Skip to content

Changelog

Shared change log for the just-dna module format/compiler ecosystem. Because just-dna-format + just-dna-compiler are consumed by just-dna-pipelines, just-dna-marketplace, and just-dna-agents, cross-repo integration changes are recorded here so parallel work in the other repos isn't surprised. Newest first.

A version heading names the release a change will ship in, not a development batch. Entries dated 2026-08-03 carried a 0.5.1: label for a while; at the time nothing 0.5.x had been published, all three packages sat at 0.5.0, and that work therefore shipped as 0.5.0 — the label was relabelled to match. Keep it that way: a number here should answer "which published version introduced this", and a batch inside an unpublished release is not a version.

0.5.0 is published — tagged v0.5.0 and released to PyPI on 2026-08-07 (just-dna-format, just-dna-compiler, and just-dna-enricher, the last being that package's first release). Everything below dated 2026-08-07 or earlier is in it. The next heading is therefore a real new number: additive work — including a new optional column or table — is 0.6.0, while removing a column, promoting one to required, retyping one, or changing what an identity key means is 1.0. That is Principle 3 as amended on 2026-08-11; the earlier "anything that moves artifact.digest is 1.0" rested on a premise that expired when content_signature took over content identity in 0.4.1. See ROADMAP § 0.6.

That rule is about the schema surface, and the three packages version independently — so just-dna-enricher can take a patch. "Additive work is 0.6.0" sorts changes by what they do to a compiled module's identity, which is a format/compiler question. Work confined to the network tier touches no parquet, no model and no manifest field, so it can ship as an enricher patch release while format and compiler stay where they are. The first of those is just-dna-enricher 0.5.1, whose content is RM38 — a cache for the licence-gated sources, so a hosted enricher stops fetching them live per request. This does not reopen the paragraph above: that one is about labelling a batch inside an unpublished release, which 0.5.1 is not — 0.5.0 shipped, so 0.5.1 is a real next number rather than a name for work in progress. 0.5.2 follows the same rule and stretches it by one package: its ClinVar-drafting, query-shape and cache-location work is enricher-only, and the one compiler change (a warning when resolve_with_ensembl=False discards an injected resolution.csv) writes no parquet and moves no signature, so just-dna-compiler took the patch alongside while just-dna-format stayed at 0.5.0.

main is the patch line; minor-class work lives on the 0.8 branch (since 2026-09-25). RM256, RM257, RM259, RM260 and RM262 were committed to main on 2026-09-24/25 although each sizes as a minor, and were moved to the 0.8 branch by a history rewrite the same evening. Their entries below stay here because this log is shared across the ecosystem, but their code is not on main and ships with 0.8. main is v0.7.2 plus patch-scope work only. The rule is in RELEASE_CYCLE.md. The 0.8 branch merged into main at the 0.8.0 cut (2026-11-05).

2026-11-18 — RM455: the pre-0.9 instances of four classes 0.9 fixed in its own code (patch, enricher only, on main)

0.9's review round 1 fixed four error classes in its own code; these are the older instances on main, each fixed in the form 0.9 gave its own. Nothing a module resolves moves.

  • A pull record that is not UTF-8 reads as unreadable, never as a traceback. download._pull_record, locations.pull_record_repo (and pulled_from through it) and locations.snapshot_record_missing caught JSONDecodeError but not UnicodeDecodeError, so such a record escaped them. Each now takes its own unreadable arm: no record, no repo named, and an unrecorded cache with a warning.
  • A batch size below 1 is refused where it is read. GnomadSettings, EutilsSettings, EuropePmcClient and PmcIdConverterClient never checked batch_size, and net.batched yields no batch for 0, so a client set to 0 asked nothing and answered {}, which every caller reads as no record for any id. Each now refuses it with a ValueError naming the field (net.check_batch_size), when it is built and again by each batched read.
  • A lane's root-file payload is a field, never its name. cache rebuild --publish and its --dry-run chose the payload a snapshot is checked and uploaded by with lane.name == "strchive". CacheLane.payload_filename now states it for strchive and acmg, a test holds it equal to what each lane's resolver finds, and an AST walk refuses a lane.name compared with a literal anywhere in the enricher.
  • Not here: mitomap_miss calling itself present over moved parents. main has no stale state for any lane; that one is fixed with RM431 on 0.9.

2026-11-18 — RM453: a mirrored site where no spelling was witnessed keeps both spellings (patch, a declared corrected derivation, on main)

A correction to 0.8.9, declared (P3). 0.8.9's RM449 left an rsID row unpositioned where a second source recorded unplaced on both spellings of a mirrored insertion and deletion ("witnessed neither"). But unplaced on every spelling means only that the witness held no counterpart under the rsID. ClinVar files only some variants, and it files pathogenic_clinvar's rs80356821 deletion under rs281864900, which dbSNP merged into it. So an offline enrich refused that row under --strict, where the online table placed it. That is no evidence for either spelling, the same knowledge as undecided or an empty cell. Both spellings now stay filled, under a warning (borrowed code rsid_expanded_to_multiple_loci until 0.9) saying the source asked "placed neither as the same event". A site where one spelling was witnessed still fills only that one. What moves: a table with unplaced on every spelling of a mirrored site compiles that rsID row onto both spellings again, as before 0.8.9; nothing the enricher writes changes. Compiler only.

2026-11-18 — v0.8.9: format, compiler and enricher 0.8.9 (a patch, cut from main)

The compiler moves for RM447, RM449 and RM451, the enricher for RM446, RM448, RM449's ClinVar drafter and RM452. Format 0.8.9 changes only its release record (RELEASE_RECORDS["0.8.9"], which declares RM449's corrected derivation): no vocabulary, model or field moves, so a reader on just-dna-format 0.8.6 still reads every 0.8.9 manifest (measured: all 22 reference modules compiled by 0.8.9 read under format 0.8.6). Asking needs_recompile(stamp, "0.8.9") needs format 0.8.9, since 0.8.6 holds no record for it. Each intra-workspace floor is >=0.8.9: the compiler's just-dna-format>=0.8.9, the enricher's just-dna-format>=0.8.9 and just-dna-compiler>=0.8.9. No third-party floor moved, and uv lock moved only the three workspace versions. Sized by the pm, 2026-11-18: RM449 first, the silent one, and RM448 beside it; both block three of the registry's production modules (S158).

  • What it carries: RM449 (an rsID row's indel genotype is filled on the spelling a second source witnessed, not on both events at a mirrored insertion and deletion; a declared corrected derivation), RM448 (online, a witness holding one ALT of a multi-allelic indel site no longer withholds the whole site), RM451 (reverse no longer invents a resolution.csv row for a coordinate row the table never filed), RM447 (an allele refusal on a row Ensembl REST answered names REST's older dbSNP build and the remedy that reaches it), RM446 (a reused client's earlier 410 no longer tells a later run's outage to upgrade) and RM452 (a NULL start in the Ensembl snapshot places nothing instead of raising TypeError). The pathogenic_clinvar example gets back its coordinate row 11:5226762 C>CAAAG (RM449's remedy), so its content_signature moves; nothing publishes it.
  • No new warning code. RM449's sentences borrow the existing code of what the row became until 0.9 adds rsid_genotype_reads_as_two_events: rsid_no_hosting_locus where no position is written, and rsid_expanded_to_multiple_loci (carried) where one or more loci are filled. Until 0.9 a site narrowed to one locus counts as an expansion in warnings_summary, and the filled arms' sentences sit in compilation.carried although an authored coordinate row clears them.
  • A new refusal under --strict: an rsID row at a mirrored indel site whose recorded placement_verdict places neither spelling as same_event (an unplaced answer from a witness) is no longer filled, and --strict refuses it as unresolved. A recorded undecided on both spellings reads as no verdict and keeps both. RM447 rewords the two allele refusals for an ensembl-rest row; their pinned phrases stay.
  • New public names, additive (RM446): Withholding.retired_met, net.retired_mark(client) and an optional since= on net.withheld_reason. A patch should not add a public name; these are recorded as 0.8.4's size_refusals was.
  • Known issue, RM453 (for 0.8.10): an offline re-derivation of pathogenic_clinvar leaves rs80356821 unplaced, and --strict refuses it; an online one clears it. Offline, ClinVar files that rsID's deletion under the retired rs281864900 (dbSNP merged it into rs80356821), so the witness holds no record for it and answers unplaced on both spellings, which RM449 reads as witnessed neither. The same can happen to any mirrored indel site whose ClinVar record carries a merged rsID. It is loud and names its rsID.
  • Release record (RELEASE_RECORDS["0.8.9"], measured from 0.8.6, the last published compiler, both sides on polars 2.1.0 and Python 3.14.8): of 22 reference modules, 21 measured on both sides. cyp2c9_warfarin_grch37 moves parquet_bytes (RM451: 12 pharm_variants rows' resolution_loci null → [], declared an addition) and hboc_palb2 gains two RM449 sentences under rsid_expanded_to_multiple_loci; content_signature moves on none. pathogenic_clinvar is unmeasured: 0.8.6 refuses its restored row as Duplicate (variant, genotype), the loud case RM449 fixes. A second sweep, the same two compilers over v0.8.8's spec of that module, measures the correction: two weights.parquet rows fewer (rs80356821 and rs193922563 lose the mirrored insertion), and stats.weights_rows and compilation.resolution_subjects 338 → 336, compilation.expanded_keys 11 → 9, compilation.expanded_rows 22 → 18 (and warnings_summary's rsid_expanded_to_multiple_loci 11 → 20). The record declares those four fields and weights.parquet as RM449 corrections, scoped on compilation.expanded_keys (a module that applied a resolution table with rows); RM451's addition states no reach, since a header-only table writes its [] too.
  • After upgrading:
  • Raise all three floors to >=0.8.9.
  • Recompile the modules RM449 reaches (needs_recompile(stamp, "0.8.9").declared_for(manifest), on format 0.8.9, keeps the correction for every module whose manifest carries compilation.expanded_keys; the reach is narrower, an rsID row whose genotype one position hosts as both an insertion and a deletion). On recompile weights.parquet, artifact.digest, stats.weights_rows, compilation.resolution_subjects, compilation.expanded_keys and compilation.expanded_rows move; content_signature does not. The fill reads resolution.csv's placement_verdict at such a site, so a table enriched online with the reference read is corrected by the recompile alone. A table with no verdict there, or undecided on both spellings (an offline run, a ClinVar fill, a table written before 0.8), keeps both events and says so; re-derive it online (just-dna-enricher enrich <spec-dir> --rederive) and recompile. The registry's two published cases are superhuman@2.4.1 rs766212849 and thrombophilia@2.0.1 rs1572088853.
  • Where an RM449 sentence fires (each one carries "(RM449)"), the rsID row claims both events, or none. If it is about one of them, replace it with a row authoring that event's coordinate and re-run the enricher so the table files it; at a mirrored site an authored coordinate row now wins over the expansion onto its exact key.
  • A table re-derived online before this release may lack rows RM448 withheld: a plain enrich on 0.8.9 asks again for every subject with no row.
  • Re-derive mirrored indel sites online, not offline, until RM453 ships (the known issue above).
  • Nothing to do for RM446, RM447, RM451 or RM452. RM451 moves artifact.digest on recompile for a module that read a table and holds a coordinate row the table never filed; the table reverse gives back is the same.

2026-11-18 — RM452: a NULL start in the Ensembl snapshot places nothing instead of raising (patch, cut as v0.8.9)

A snapshot row naming an rsID with no start raised TypeError in lookup_loci, and so did a back-fill probe with a chrom and no start. The scan now leaves such a row out of the answer. An rsID named only on such rows is said in one warning, and the next link is asked. The probe gets no candidates. No published release on this machine carries a NULL start; the RM431 builder met it with a synthetic row. Enricher only; nothing public moves.

2026-11-18 — RM451: reverse no longer invents a table row for a coordinate row the table never filed (patch, cut as v0.8.9)

resolution_loci was null both where a row's own cells answer (RM364's saving) and where the module's table files nothing under the row's key. So a coordinate row authored after the last enrich came back from reverse as a table row rebuilt from its cells, and compile → reverse → compile moved resolution_signature. Restoring pathogenic_clinvar's 11:5226762 C>CAAAG (RM449's remedy) found it, and a two-row module with no mirrored site shows it too. A module that read a table now writes [] there, for a key with no positioned locus on its build, and reverse writes nothing for it. What moves: that cell, null → [], so artifact.digest moves on recompile for a module that read a table and holds such a row. Among the reference examples that is cyp2c9_warfarin_grch37 (12 pharm_variants rows). No value a row states changes, and reverse gives back the same table. A module that read no table is unchanged. Compiler only; no column, code or vocabulary moves.

2026-11-18 — RM448: a witness holding one ALT of a multi-allelic indel site no longer withholds the site (patch, cut as v0.8.9)

Online with the reference read (verify_ref), the placement witness (RM267) compared ClinVar's one ALT of a multi-allelic indel site with Ensembl's other ALT of the same kind and length. Ensembl places rs1452575772 as A>AC,AT and ClinVar as A>AC, so A>AT was judged "a different event". The whole locus was withheld, A>AC with it, and the rsID was left with no resolution.csv row. The registry's re-derived pathogenic module lost 12 rsIDs this way and cardio one (S158). RM267's rule says a different allele is unplaced, never a disagreement. A witness allele that is already the same event as one of the rsID's placed alleles is now accounted for, and is no counterpart of the others. Those read unplaced, and the locus is written. A witness allele that matches none of the rsID's alleles still withholds, as S117's rs8176719 does. Offline runs never hit this, since nothing judges an indel offline. Enricher only; no column, code or vocabulary moves. A table re-derived online before this release may lack those rows: a plain enrich on this release asks again only for subjects with no row.

2026-11-18 — RM449: an rsID row's indel genotype is filled on the spelling a second source witnessed, not on both (patch, a declared corrected derivation, cut as v0.8.9)

A correction, declared (P3). At one position an insertion and a deletion hold the same allele strings: C/CT is one inserted T on C>CT and one deleted T on CT>C. The string hosting rule accepted both, so the fill put an rsID row's claim on both events. Alone, the row compiled under --strict, and the claim sat silently on the event it is not about. Beside an authored coordinate row for the other event, --strict refused with Duplicate (variant, genotype). S158 (the registry) hit 21 such refusals in cardio, 25 or more in cancer and 14 or more in pathogenic. Its lead found 2 of its 69 published modules carrying the silent claim (superhuman@2.4.1 rs766212849, thrombophilia@2.0.1 rs1572088853). The fill now reads ResolutionRow.placement_verdict at such a site. It fills the spellings a second source placed as same_event and not the others. Both witnessed: both stay, and a warning says so once per rsID, in both modes. Neither witnessed: neither is filled, and the row is left unpositioned (--strict refuses it as unresolved). No second source asked at all (an offline run, a ClinVar fill, a table written before 0.8): both stay as before, and a warning says so. Whatever the witness, an expansion member re-keyed onto an authored coordinate row's exact (variant_key, genotype) is not filled, and a warning names it: the coordinate row picks its event, which is RM449's own remedy, and never a strict duplicate. None is escalated in 0.8. What moves: an artifact compiled before this may hold the claim on both spellings. The reference examples move two rows (pathogenic_clinvar: rs80356821 and rs193922563 now fill only the witnessed deletion). The ClinVar drafter now writes an indel record with its coordinate, never its rsID alone, so a new draft cannot pose the question. A re-draft names an older rsID-only indel row as one no current draft would write. No new warning code in 0.8.9; the sentences borrow two (the pm, before the cut). warnings_summary validates its keys against the closed vocabulary, so a code added in a patch makes a just-dna-format 0.8.6 reader refuse the manifest of every module that triggers it, and an uncoded sentence cannot ship (findings.classify refuses a part-classified channel). Each RM449 sentence takes the existing code of what the row became: no position written is rsid_no_hosting_locus; one or more loci filled (both spellings kept, or one dropped as unwitnessed or under an authored coordinate row) is rsid_expanded_to_multiple_loci, which is carried. The text and its pinned phrases are RM449's own. Until 0.9, a site narrowed to one locus is counted in warnings_summary as an expansion, and the filled arms' sentences sit in compilation.carried although an authored coordinate row clears them. 0.9 adds rsid_genotype_reads_as_two_events and moves these sentences to it. just-dna-format stays 0.8.6. A recorded undecided reads as no verdict (before the cut, the pm). Where a mirrored site's only recorded placement_verdict is undecided (the witness was asked and could not judge, as a run with no reference sequence writes), both spellings stay, as for a site nobody was asked about, under their own sentence (pinned: "could not decide whether either is the same event"). Could not judge is not witnessed-neither, so a recorded undecided no longer turns a module that compiled into a strict refusal. undecided beside unplaced reads as an empty cell beside unplaced: neither is filled. An unplaced answer is a witness's, offline too, and still fills neither.

2026-11-17 — RM446: a reused client's earlier 410 no longer tells a later run's outage to upgrade (patch, cut as v0.8.9)

A withholding client (EnsemblResolver, CrossrefClient, the PMC clients, Grch37Client, ClingenAlleleClient) keeps the first 410 it met as retired for its life, and enrich and literature read that as the reason a run went unanswered. A caller that reuses one client across runs, which RM404's LookupClients makes the designed path, had run 1's 410 tell run 2's timeout that Ensembl retired the route and that a re-run would not help. Every 410 is now also counted in Withholding.retired_met. A run reads net.retired_mark(client) before it asks, and net.withheld_reason(client, since=mark) names retired only for a 410 met after that mark. enrich's live-Ensembl warning and LiteratureResult.crossref_unanswered use it. The client's retired message is unchanged and never reset. A host's own client without the counter reads as before. Enricher only and additive: a new attribute, a new function and an optional argument.

2026-11-17 — RM447: a resolved allele refusal names Ensembl REST's older dbSNP build when REST answered the row (patch, cut as v0.8.9)

At Ensembl 116, REST serves dbSNP 156, while the same release's VCF dump, which the snapshot is built from, carries dbSNP 157. Measured over every rsID the reference examples name on 2026-11-17: of the 454 rsIDs both answered, 54 carry alleles in the snapshot that REST lacks, and none go the other way. rs646776 is one: REST answers C/G/T, while the snapshot and dbSNP 157 carry C>A,G,T. REST answers a subject the snapshot holds only when a run has no snapshot, so such a run records the older set, and --strict refuses a correct genotype on it. genotype_allele_not_at_locus and effect_allele_not_at_locus on a row whose source is ensembl-rest now say the row came from a live link that can serve an older dbSNP build than the snapshot, and give the remedy that reaches it: enrich --rederive on a host where the Ensembl snapshot is provisioned. The phrases earlier snapshot release, release not recorded and --rederive stay in the text. Rows answered by any other link read as before. Compiler only; no column, code or vocabulary moves.

2026-11-16 — v0.8.8: enricher 0.8.8 (a patch, cut from main)

Only the enricher changed source since v0.8.7 (download.py), so it alone moves; format and compiler stay 0.8.6 and are not rebuilt. The enricher's floors stay just-dna-format>=0.8.6 and just-dna-compiler>=0.8.6, and no third-party floor moved. Sized by the pm to cut at once and alone: a silent wrong answer under a current-looking label, and sticky, so every day uncut is another cache that never repairs.

  • What it carries: RM445. A pull no longer places the hub's release.json over cached files it kept and nothing attests, so old bytes no longer carry the new release's label. cache pull and cache prepare repair such a cache, including one whose label already equals the hub's: a pull record without release_sha256 (written before 0.8.3), and any pulled cache whose files the hub's listing contradicts by size or, up to 256 MiB, by LFS sha256. Neither enrich nor cache pull fetches beside a local build any more; a missing part is reported as could-not-run.
  • A new refusal for a direct caller: a release.json that could not be fetched (a timeout, a 429, a refused token) stops a fresh or unfinished pull before anything in the cache moves; before, it read as "the repo publishes none". A cache already there is served as it is.
  • Release record: none, as for 0.8.1 and 0.8.7. The compiler did not change.
  • After upgrading:
  • Raise the enricher's floor to >=0.8.8; format and compiler stay >=0.8.6.
  • Before the first cache pull, read each lane's pull-record.json: one with no release_sha256 key marks a cache that may have served an older release under a newer label, and so may any pulled cache whose parquet files are much older than its record. Modules enriched against such a lane since that record's date may carry old answers; re-enrich them after the pull.
  • Run just-dna-enricher cache pull once. It repairs every affected pulled cache; a local build is left alone.

2026-11-16 — RM445: a pull never places the hub's release.json over files it kept and nothing attests (patch, cut as v0.8.8)

A pull of a cache with no pull record and no manifest.json kept every whole parquet and then fetched the hub's release.json, so old bytes stated the new release. The cache was then recorded and whole, and every later cache pull read it as current. This machine's ClinVar cache served August data as clinvar_2026-11-01 that way (the pm's remote-read trial, finding 8). Now the hub's release.json is placed only over files the pull fetched, or kept because the manifest attests them, or because this repo's own pull recorded that they are the release the repo still publishes; AlphaGenome's avi_knots.parquet, which reconstructs PHRED for one release, follows the same rule. cache pull and cache prepare fetch an unattested kept file again, as RM420 does for a release that moved, so they repair. enrich's own pull keeps it (no release is downloaded mid-run) and places no label: the cache states no release, so its label is withheld rather than false. An unrecorded cache stating a release of its own is a builder's and is not refetched on either path: left alone with another label (RM420), kept as it is with the hub's. A pull never fetches beside a local build (a release.json no pull of this repo placed), on either path: a build lacking a part the repo publishes, citations/ say, is left untouched, nothing removed or fetched and its release.json kept, and every run warns that the part could not be reached, naming the remedies (rebuild it with the part, or move it aside and cache pull). Before, enrich completed such a build from the hub and labelled the mixture with the hub's release. A pull's own release.json beside files it can no longer vouch for is withdrawn before the first byte, and the next cache pull re-pulls the cache. The record states the digest of the release.json the pull placed and null when it placed none, so an enrich run over a build no longer makes the next cache pull take the build for a pull. A pull fetches release.json first, staged beside the cache, and names it in the record it writes before the first download, so a resumed pull keeps what it fetched while the hub still publishes that release and fetches it again past a release boundary. A repo that publishes no release.json pulls as before. STRchive's provisioner keeps no file, so the mechanism does not reach it; it now withdraws an old release.json before every catalogue it moves in, not only on a refresh.

A new refusal for a direct caller. ensure_*_snapshot (the parquet lanes) now raises the lane's error when the repo's release.json could not be fetched (a timeout, a 429, a refused token: anything but the Hub answering that nothing is there) and the cache is absent or unfinished; before, that failure was read as "the repo publishes none" and the pull went on unlabelled. A cache already there that nothing calls unfinished is served as it is, with a warning, and nothing in it moves, since the question is asked before the manifest is placed. Ensembl's lane, which publishes no release.json, now asks for one on every pull, so an outage of that one request stops its pull too.

Caches already wrong are repaired by the next cache pull. A finished pull record without release_sha256 was written before RM420 (0.8.3), by RM339's pull, which placed the hub's label over whatever it kept. cache pull now re-pulls such a cache even when its label equals the hub's, and keeps re-pulling one whose label differs, as RM420 did. enrich still serves a whole recorded cache without the network, this one included, so run cache pull once.

Did an enrich run record the new label over old answers? Possibly, on any machine whose cache carried such a record, between its pull and this fix; nothing here knows which modules were enriched on one. The marker is the lane's pull-record.json with no release_sha256 key. Read it before running cache pull, which rewrites it; where the key is absent, re-enrich the modules enriched against that lane since the record's date. A cache that 0.8.3 to 0.8.7 completed from no record at all carries the key, so nothing on disk marks it. That cache, too, is repaired by cache pull, which now compares a pulled cache whose label equals the hub's with the published listing, as below. Its modules enriched since the pull are not identifiable from here either; re-enriching against the lane after cache pull is the safe move where the timing fits.

cache pull checks a current label against the listing. For a pulled cache (a pull record of this repo; never a local build, which differs from the hub by design) whose label equals the hub's, cache pull and cache prepare list the repo with detail and compare every kept file: its size, and, for a file of at most 256 MiB, the LFS sha256 the listing states. A file that differs, or is absent, is fetched again, with the label withdrawn first and placed again once the file is the release's. The outcome is refreshed, and the note names the files. A listing that cannot be read leaves the cache as it is, outcome unknown. enrich reads no hub for a whole recorded cache and does not run this check. Above 256 MiB only sizes are compared: hashing an AlphaGenome AVI shard on every pull would be minutes of disk reads, and a release that moved changes a parquet's size in practice. A file stored in git rather than LFS carries no digest in the listing and is compared by size.

No public name, field, flag or vocabulary moves. Tests: test_download.py (the reproduction on both paths, the pre-RM420 record with an equal and with a differing label and on the enrich path, a build an enrich run passes over, a build stating the hub's own label on either path, the label a fetch withdraws and the cache pull after it, a pull's own label an enrich run withdrew, a build lacking a part on both paths, a release.json that could not be fetched, a pull killed after placing its label, a resume of the same release on both paths and one across a release boundary, each repair's second lap) and test_cache_pull_follows_the_hub.py (the knots, with and without a release.json, Ensembl's manifest left in place when release.json could not be asked, STRchive's old label never kept over a new catalogue, and the listing check: a size, a digest, the hashing bound, agreement, enrich and builds untouched, an unreadable listing, a refetch cut short). Each arm of the fix reverted on purpose turns at least one red, and RM420's tests pass unchanged.

2026-11-14 — v0.8.7: enricher 0.8.7 (a patch, cut from main)

Only the enricher changed source since v0.8.6, so it alone moves; format and compiler stay 0.8.6 and are not rebuilt. The enricher's floors stay just-dna-format>=0.8.6 and just-dna-compiler>=0.8.6, and no third-party floor moved. Both items are text only: nothing adds a public name, field, table, flag or warning code, and no outcome, return value or raise changes for a direct caller.

  • What it carries: RM442 (the two AlphaGenome licence rows' notice and license say what the saved terms say: the Atlas row states the training bar of the Output Terms effective 2026-10-09 beside June 2025's, and the AVI row cites the Output Terms' restriction 2 instead of the Additional Terms' prohibition 4) and RM444 (grch37.recover_rsid's text says no more than it matched: a none beside overlapping records counts them and says to retry at a neighbouring anchor, and a position-only recovered makes its advice conditional).
  • Release record: none, as for 0.8.1. The compiler did not change, so a recompile moves nothing; the release record axes measure compiled output only.
  • After upgrading:
  • Raise the enricher's floor to >=0.8.7; format and compiler stay >=0.8.6.
  • A module whose AlphaGenome rows were enriched before this carries the old licence text in sources.csv. Re-enrich it to take the corrected text; its digest moves, which is the intended effect. No licence axis moves, and no gate reads notice.
  • A caller that matched recover_rsid's old sentences ("records nothing here" beside overlapping records, the --ref/--alts flag names) reads the new ones; the outcome field is the stable key.

2026-11-14 — RM444: recover_rsid says no more than it matched (patch, cut as v0.8.7)

grch37.recover_rsid's result text over-claimed in four places, reported on CFTR F508del (S155), whose GRCh37 record is padded at Ensembl's own anchor, one base right of VCF's spelling. A none beside overlapping records said "dbSNP records nothing here"; it now says no record starts here, counts the overlaps, and says to retry at a neighbouring anchor before reading it as absence. "Records nothing here" stays for an empty answer. The note no longer says "with the alleles given" when none were. A recovered with no ref says it matched on position alone and makes its advice to author the rs-number conditional, since at a padded indel's VCF position the record starting there is the anchor base's SNV. The ambiguous text names the arguments ref/alts instead of the CLI flags. The docstring and ENRICHER.md say the anchor may be neither VCF's nor HGVS's. No outcome or field changes; the tests replay Ensembl's real answers at the three F508del positions.

2026-11-14 — RM442: the two AlphaGenome licence rows say what the terms they quote say (patch, cut as v0.8.7)

Three sentences in licensing.py misstated the AlphaGenome terms, found when the liaison saved the live texts (26db64cb). The Atlas row's notice gave only the June 2025 training bar; it now also states the bar in the Output Terms effective 2026-10-09, which forbid training any model except an AlphaGenome-like one under the Model Parameters Terms, for Output generated from that date. Its license dated the Output Terms with the Additional Terms' 2026-09-08; it now names 2025-06-25 and 2026-10-09. The AVI row cited the Additional Terms' prohibition 4, which exempts Permissive artifacts such as AVI; it now cites the Output Terms' restriction 2 (June 2025 text), which binds a holder who has not agreed to the Additional Terms and has no such carve-out. No licence axis moves and no gate reads notice. A module re-enriched after this carries the new text in sources.csv, so its digest moves; that is the intended effect.

2026-11-13 — v0.8.6: format, compiler and enricher 0.8.6 (a patch, cut from main)

Format and the enricher move because each changed source since v0.8.5: format for RM439, the enricher for RM440. The compiler's source did not change. It moves because its floor must name the format's current version (test_workspace_versions), so its pyproject.toml changes, and a published 0.8.5 cannot take a second set of metadata. The enricher's floors are just-dna-format>=0.8.6 and just-dna-compiler>=0.8.6, the compiler's just-dna-format>=0.8.6. No third-party floor moved. Nothing adds a public name, field, table, flag or warning code, and no return value or raise changes for a direct caller: RM440 adds one log line.

  • What it carries: RM439 (VariantRow.stat_significance's description says the level is the module author's verdict across the variant's evidence, with no threshold of its own), RM440 (enrich --rederive logs one warning per run naming the rows that lost a back-filled rsID on RM438's no-host arm), and RM437's doc half (INTEGRATION_0_8 § 3 says the ensembl lane carries the core only, to keep the current source for consequence, clinical flags and frequencies until 0.9, and why a hand-fetched groups/ sidecar is unsafe). RM437 itself, the route that pulls those sidecars, stays open for 0.9.
  • Release record (RELEASE_RECORDS["0.8.6"], measured from 0.8.5, the last published compiler): 22 reference modules, nothing moved on any axis. RM439's description reaches one line of describe variants.csv and no compiled output. RM440 writes nothing to any table. No recompile is needed.
  • After upgrading:
  • Raise floors to >=0.8.6.
  • Nothing else is required. RM440's warning appears on enrich --rederive only, and only where a recorded row loses its back-filled rsID; its phrase is "lost its back-filled rsID on this re-derivation".

2026-11-13 — RM440: a re-derivation that drops a back-filled rsID on RM438's no-host arm says so (patch, cut as v0.8.6)

Since RM438 (0.8.5), enrich --rederive writes a coordinate-only row with no rsID where every rsID at its position is refuted by the row's alleles, which replaces a recorded row that carried one. That is right, and it was silent: only a diff of the table showed which rows lost theirs. The run now logs one warning, aggregated over the rows: "N row(s) lost its back-filled rsID on this re-derivation: no rsID at its position hosts an allele the row states", with the first five variant keys. A log line, no warning code; nothing written to the table changes. The registry's re-derivation of its three ClinVar panels is where it will be read. RM438's residual.

2026-11-13 — RM439: a variant's stat_significance says it is the module author's verdict across the evidence (patch, cut as v0.8.6)

VariantRow.stat_significance listed its four levels and said nothing more, while RM434 had just explained the same word on studies.csv. The row-level description now says the level is the module author's verdict across the variant's evidence, with no threshold of its own, and points at each study's own grade in studies.csv and its p_value_num for a common scale. A description only: no digest, member or check moves. RM434's residual.

2026-11-12 — v0.8.5: format, compiler and enricher 0.8.5 (a patch, cut from main)

Format and the enricher move because each changed source since v0.8.4: format for RM434, the enricher for RM432, RM436 and RM438. The compiler's source did not change. It moves because its floor must name the format's current version (test_workspace_versions), so its pyproject.toml changes, and a published 0.8.4 cannot take a second set of metadata. The enricher's floors are just-dna-format>=0.8.5 and just-dna-compiler>=0.8.5, the compiler's just-dna-format>=0.8.5. One third-party floor moved down: the enricher requires tenacity>=9.0.0 where 0.8.1 to 0.8.4 required >=9.2.1 (RM436). Nothing adds a public name, field, table, flag or warning code. Two behaviours change for a direct caller: two snapshot readers raise where they returned an empty value (RM432), and resolve_variants may now withhold an rsID it used to attach (RM438).

  • What it carries: RM432 (gene_metrics.lookup_snapshot and civic_refutation.civic_snapshot_rows raise FileNotFoundError on a reference they cannot read; the CIViC refutation skip's log line is reworded, and the citations and gene-metrics passes now say when they could not read a snapshot), RM434 (StudyRow.stat_significance's description says whose threshold a level is), RM436 (the enricher accepts tenacity 9.0 again, so it resolves beside google-genai), and RM438 (a coordinate-only row's back-filled rsID is narrowed by the row's alleles on both paths, a corrected derivation of resolution.csv). Beside them, the GWAS pass's docstrings quote the Catalog v2's published budget of 15 queries a second (no behaviour moves), and RELEASE_CYCLE states that a moved floor moves its package.
  • Release record (RELEASE_RECORDS["0.8.5"], measured from 0.8.4, the last published compiler): 22 reference modules, nothing moved on any axis. RM434's description reaches describe studies.csv and no compiled output. RM438's correction is to resolution.csv, an enricher output the record does not measure; its entry below declares it, with the registry's exposure measured on production (nothing published moves on either arm). No recompile is needed.
  • After upgrading:
  • Raise floors to >=0.8.5.
  • A project that carried a tenacity>=9.2.1 override to resolve the enricher beside google-genai can drop it.
  • A direct caller of gene_metrics.lookup_snapshot or civic_refutation.civic_snapshot_rows catches FileNotFoundError, raised for a reference with no parquet behind it; {} and [] now mean only "read, and nothing matched".
  • A direct caller of resolve_variants expects a coordinate-only row at a position several rsIDs share to come back with no rsID and an rsid_ambiguous warning naming every candidate, where it used to get the first in string order; and none, under rsid_without_resolution_label, where the row's alleles refute every rsID there.
  • Re-derive (just-dna-enricher enrich <spec-dir> --rederive) any module whose coordinate-only rows were back-filled from a snapshot, to apply RM438's correction; a recorded row is otherwise kept as it is. The three ClinVar panels still on pre-0.8 tables are where it applies on the registry.

2026-11-12 — RM438: a coordinate-only row's back-filled rsID is narrowed by the row's alleles, on both paths (patch, cut as v0.8.5)

resolve_variants gave a coordinate-only row the first rsID at its (chrom, start, ref) in string order and never read the genotype. dbSNP 157 added a 16 bp insertion (rs2517910443) at CYP2D6*4 (22:42128945, rs3892097), so an A/G row there got the insertion, with no warning. The enrich path flagged its pick ambiguous but chose it the same blind way. One private rule, resolver._hosting_rsids, now narrows a position's rsIDs on both paths. It reads the row's alts when stated (with its ref), else every genotype the variant's rows author, and judges them with the forward direction's hosting predicate (site_verdicts). An undecided verdict keeps a candidate. An allele it cannot read (*, symbolic, unparseable, no genotype) narrows nothing. On the enrich path a row stating one ALT was already allele-exact and is unchanged; resolve_variants never read alts before, so there it now matches the ALT too. A key whose rows name different rsIDs (A/C and C/<insertion> at one coordinate) gets none, since each label is refuted by a row. No public name and no warning code is added.

  • resolve_variants (deprecated, removed at 1.0; called directly by consumers). One rsID left is attached. More than one is withheld with rsid_ambiguous naming every one, since this path writes nothing that could mark a pick. None left, where rsIDs sit at the position, attaches nothing and says, under rsid_without_resolution_label, that none there can host the row. A behaviour change for a direct caller: a row that got an rsID may now get none and a warning, and the ref-less multi-allelic pick with its "Specify ref to disambiguate." warning is gone.
  • enrich writes a corrected derivation of resolution.csv (P3). It keeps pick-and-flag, applied to what is left: an ambiguous row becomes resolved where exactly one rsID hosts the genotype, and keeps ambiguous with a shorter rsid_alternates where several do. Where none hosts, the row is written with no rsID (source=authored), as an allele-exact miss already was. That reaches rows that were ambiguous, and rows that were resolved on a single back-filled rsID the genotype refutes. A link whose candidates the row refutes is a miss, so ClinVar is read next, and a row's source can move from cache to clinvar. Rows move only when a table is next written (a fresh enrich, or --rederive). Merge-not-clobber keeps a recorded row otherwise.
  • Exposure, measured. The registry ran the exposure query on production, read-only, over the latest version of all 24 published modules: 439,752 resolution.csv rows, 0 with status=ambiguous, 0 with rsid_alternates. So no published ambiguous row moves. The query, kept for the re-derivation ahead: status=ambiguous, a coordinate variant_key, and alts empty or multi-valued. The registry then ran the second query on production, read-only: status=resolved, source cache or clinvar, a coordinate variant_key with rsid set, and alts empty or multi-valued. 7 rows carry a snapshot back-fill (source=cache), all in one module (smell 0.1.1), and each row's authored genotypes sit within its own ref/alts, so none is refuted. Nothing published moves on either arm. 51,249 more rows carry a ClinVar back-fill, all in the three ClinVar panels still on pre-0.8 tables (cancer, cardio, pathogenic 2.0.0); their re-derivation on 0.8.5 is where this change applies. The test registry was not measured.

From S151 and S152.

2026-11-12 — RM436: the enricher accepts tenacity 9.0 again, so it resolves beside google-genai (patch, cut as v0.8.5)

RM371 floored tenacity>=9.2.1 so every retried client could spell wait_exponential_jitter(multiplier=…), the keyword 9.2.1 renamed from initial. Every google-genai release caps tenacity<9.2.0, so a consumer depending on both (just-dna-lite) could not resolve the 0.8 trio without an override, and a relock that did not raise the enricher's own floor kept a mixed trio silently. The floor is now tenacity>=9.0.0. One private helper in net.py builds every backoff and picks multiplier= or initial= by the installed constructor's signature, so RM371's deprecation warning stays silent on 9.2.1 and newer, and 9.0 and 9.1 get the keyword they take. A test holds that no client calls wait_exponential_jitter itself, and the floor is the release a run proved: the enricher suite passed on tenacity 9.0.0 and on the locked 9.2.1. Backoff timing is unchanged. No public name moves. From S149.

2026-11-12 — RM434: a study's stat_significance says whose threshold it is (patch, cut as v0.8.5)

StudyRow.stat_significance listed its four levels and stated no threshold behind any of them, so a reporter graded a look-up replication at P 0.002 against the genome-wide scale and wrote not_significant, beside a failed look-up at P 0.6. The description now says the level is the authors' verdict at the threshold of the study's own test, that a pre-specified look-up which passes is significant, and that p_value_num is the column to threshold on one common scale. A description only: no member, digest or check moves. A fifth member (nominal) was declined, since p_value_num already lets a reader compute it. From S147.

2026-11-11 — RM432: two snapshot readers raise on a reference they cannot read, instead of answering "nothing found" (patch, cut as v0.8.5)

gene_metrics.lookup_snapshot returned {} and civic_refutation.civic_snapshot_rows returned [] for a reference with no parquet behind it, the value of a lookup that read the snapshot and matched nothing. Both now raise the built-in FileNotFoundError, naming the path; {} and [] keep meaning "read, and nothing matched". A behaviour change for a direct caller of either function, which must now catch it. The three callers in the enricher report could-not-read: the gene-metrics pass treats the snapshot as not consulted (offline, it no longer writes a not_found row naming a release it never opened, and names the genes as never asked), the CIViC refutation check is skipped with a warning naming the path, and the citations pass warns that it mapped variants from the curated name identities alone (a mistyped --civic-snapshot used to fall back silently). No new public name or warning code. From S146.

2026-11-11 — v0.8.4: format, compiler and enricher 0.8.4 (a patch, cut from main)

Format and the enricher move because each changed source since v0.8.3: format for RM425, the enricher for RM426, RM427 and RM428. The compiler's source did not change. It moves because its floor must name the format's current version (test_workspace_versions), so its pyproject.toml changes, and a published 0.8.3 cannot take a second set of metadata. The enricher's floors are just-dna-format>=0.8.4 and just-dna-compiler>=0.8.4, the compiler's just-dna-format>=0.8.4. One public attribute is new, GwasCatalogClient.size_refusals (RM428's count of refused first pages; a patch should not have added it, and RM429 puts it on the result in the next minor). Nothing else adds a public name, field, table, flag or warning code.

  • What it carries: RM425 (effect_size's description names the single strongest track), RM426 (GeneLocusConflict names both readings of a gene paired with a variant on another chromosome), RM427 (the ClinVar-pass skip names cache pull --only clinvar and clinvar build, commands that exist), and RM428's asks 1 to 3 (the GWAS pass asks for a page of 500, and asks a refused first page once more with no size). RM428's ask 4, a v0.7.4 for the 0.7 line, was declined by the maintainer, so RM428 closes here. Beside them, two changes no installed package carries: a test fix that keeps huggingface-hub 2.4's agent-registry fetch inside the network fence, and the weekly dependency bump of the lock (huggingface-hub 2.4.0 among it; no published requirement moved).
  • Release record (RELEASE_RECORDS["0.8.4"], measured from 0.8.3, the last published compiler): 22 reference modules, nothing moved on any axis. No recompile is needed.
  • After upgrading:
  • Raise floors to >=0.8.4.
  • Re-run just-dna-enricher gwas <spec-dir> on any module whose GWAS pass ran on enricher 0.8.0 to 0.8.3 after the GWAS Catalog's change (some time after 2026-11-04). Those runs asked for a page the Catalog now refuses, so they filled no gwas_effects.csv row and reported every rsID as could-not-ask; nothing was written in their place, so the re-run fills the table. Enricher 0.7.3 fails for another reason (its v1 API answers 410), and no 0.7.4 will be cut: upgrading to 0.8.4 is the only fix.

2026-11-11 — RM428: the GWAS pass asks for a page the Catalog accepts, and survives the next drop (patch, cut as v0.8.4)

The GWAS Catalog's REST API v2 began refusing any page above 500 associations with a bare 400 some time after 2026-11-04, and the pass asked for 1000, so its first request failed for every rsID and it filled no gwas_effects.csv row on any module (reported as could-not-ask, never as empty). It now asks for 500 and still follows _links.next, so a variant with more associations comes back whole. A 400 to an rsID's first request is asked once more with no size, at the Catalog's own default page; the run warns once naming the refused size, and the extra request is counted in requests_made. A second 400, or one on a later page, stays could-not-ask. An opt-in live test (JUST_DNA_NETWORK_TESTS=1) asks with the page size the pass uses and expects a 200. No public name, field or check result moves. Every installed 0.8 enricher fails this way until it upgrades. 0.7.3 fails for another reason (its v1 API answers 410) and stays broken: the maintainer declined a v0.7.4 (desk #252), so upgrading to 0.8.4 is the only fix. From S143.

2026-11-10 — RM427: the ClinVar-pass skip names commands that exist (patch, cut as v0.8.4)

With no ClinVar snapshot reachable, the clinical-assertion pass told the reader to run clinvar pull, in its log line and in the assertions command's own line. No app registers that command; the fetch is cache pull. Both lines now read "provision one with cache pull --only clinvar, or build one with clinvar build" (the log line with the just-dna-enricher prefix). A new test resolves every command quoted in the three packages' sources against the two apps' command trees, so a quote left behind by a moved command fails the suite. Wording and a test only: no public name, field or check result moves. From S142.

2026-11-09 — RM426: the gene-versus-chromosome finding stops guessing which half is wrong (patch, cut as v0.8.4)

check-identifiers' GeneLocusConflict closed on "a real symbol beside an invented rsID is what a machine-written summary produces". The check cannot see which half of the pairing is false, and S141's was a real rsID (rs1801394, in MTRR) beside the wrong gene. The closing sentence now reads "Either half may be the wrong one: a real symbol beside an invented variant, or a real variant paired with the wrong gene, and a machine-written summary produces both. Look up which gene sits in, then check the row against the source you took it from." The first sentence, the fact a reader greps, is unchanged byte for byte. No check result moves. From S141.

2026-11-09 — RM425: effect_size's description says it is the single strongest track, not an aggregate (patch, cut as v0.8.4)

ExpressionEffectRow.effect_size's description read "aggregated over the scorer's tissue tracks". The enricher writes the single strongest track with its sign kept, so a size whose sign contradicts effect_direction (the majority sign) is correct, and the description now says so; ranking by |effect_size| ranks each variant's most extreme single tissue. Wording only: no field, type or artifact byte moves. Reaches describe and the authoring reference. From S140.

2026-11-09 — v0.8.3: format, compiler and enricher 0.8.3 (a patch, cut from main)

All three packages move, because each changed source since v0.8.2: format for RM423, the compiler for RM422, the enricher for RM423's other half and the rest. The enricher's floors are just-dna-format>=0.8.3 and just-dna-compiler>=0.8.3, the compiler's just-dna-format>=0.8.3. Nothing adds a public name, field, table, flag or warning code; RM422 is one more arm of the existing derived_row_orphan, and RM423 corrects what merge_records returns over moved bytes, declared in INTEGRATION_0_8.md § 4.

  • What it carries: RM418 (one NCBI pacing gate per process), RM420 (cache pull and cache prepare follow the hub's release over a pulled cache), RM409 (Ctrl-C stops the Open Targets mirror and the AlphaGenome build within a chunk), RM421 (a 120 s default timeout on every bulk download, so the stall RM409's entry leaves open now raises and is retried), RM422 (a resolution.csv row no authored table names is reported), RM423 (a later command no longer re-signs records of checks it did not run over moved bytes).
  • Release record (RELEASE_RECORDS["0.8.3"], measured from 0.8.2, the last published compiler): 22 reference modules, nothing moved on any axis. No recompile is needed.
  • After upgrading:
  • Raise floors to >=0.8.3.
  • cache pull or cache prepare once: a lane pulled earlier moves to the hub's current release (gnomAD constraint to gnomad_v4.1.2_constraint, STRchive to strchive_v2.28.0). A STRchive catalogue pulled before this release carries no pull record and is left alone; move it aside and pull.
  • A verification.json written by 0.8.2 or earlier over authored bytes edited since may hold records its last writer carried across the edit; the fix does not repair it. Re-run the checks it names, or delete it and re-run every pass that writes it.
  • Delete any resolution.csv rows a new derived_row_orphan warning names: enrich never removes them.

2026-11-09 — RM423: a later command no longer re-signs the records of checks it did not run over rows they never read (patch, cut as v0.8.3)

  • merge_records drops the earlier records once the authored bytes have moved (RM423, from S139). With existing_still_binds=False the result is this run's records alone, newest-wins among them. Before, a record for a check the run did not put was kept whatever the bytes did, and record_verification signed it under the binding of the module as it now stands. So after an author edited a table, any later command that wrote verification.json (enrich, literature, a check-*) published the old answer as this module's: a finding on a row the author had removed, or a findings=0 pass over rows the check never read. The compiler then saw a bound document and kept the block.
  • The run names what it drops. record_verification logs one warning listing each check whose record it dropped because the bytes moved; re-run those checks to record them again.
  • Unchanged while the bytes stand still: the default stays True, and a check a run did not put keeps its earlier record (RM45), as does an answer a fresh skip would otherwise replace (RM72). The closure was already dropped on moved bytes (RM73). No public name moves. A corrected derivation of a public function's output, so declared rather than silent: see INTEGRATION_0_8 § 4 for documents written by 0.8.2 or earlier.

2026-11-09 — RM422: a resolution.csv row for a variant the module no longer names is reported (patch, cut as v0.8.3)

  • validate and compile name a resolution row no authored table asks about (RM422, from S138). A module that dropped a variant after enrich kept that variant's resolution.csv row, and every check over the table asked only about rows an authored table names, so --strict said nothing and the compile hashed the file into manifest.derived. Now one more arm of derived_row_orphan reads: "resolution.csv places {n} identity(ies) no table in this module names: {orphans}. Delete those rows: enrich keeps a row it is not asked about, so no run removes them."
  • A row is the module's own when its variant_key or rsid is named by variants.csv or any locus table (haplotypes.csv, pharm_variants.csv, heteroplasmy.csv), or is a study's proxy_rsid: the subjects enrich asks about. No reference example gains the warning.
  • Unchanged: a warning in both modes, never a refusal (a hand-kept row is plausible mid-edit), and no new code or public name. The row never reached a parquet, so artifact.digest and content_signature do not move.

2026-11-08 — RM421: a bulk download from a server that goes silent is given up on, not waited on forever (patch, cut as v0.8.3)

  • Every bulk download now has a timeout (RM421). net.stream_to_file defaults to 120 s when its caller names none; before, the default was no timeout at all, and eleven of the fourteen builder downloads named none (clinvar twice, constraint, drug_labels, mitomap, strchive, clinpgx, pubmind, and the Open Targets listing and parts). A server that accepted the connection and then sent nothing blocked the build forever, with no retry line and, on the Open Targets pool, no Ctrl-C. Now the stall raises a read timeout, which is retried as a transport error and then reported as the lane's own error, so a cache rebuild moves on to the next lane.
  • A slow mirror is not cut off: the timeout bounds the gap between two chunks, not the whole download.
  • Unchanged: no public name moves, and this is not a setting. A caller's own timeout= still wins, including an explicit None, which no caller in the package passes; civic, clinvar_gkm and mane no longer name the 120 s they passed, since it is the default now.

2026-11-08 — RM409: Ctrl-C stops the Open Targets mirror and the AlphaGenome build mid-download (patch, cut as v0.8.3)

  • An interrupt stops the in-flight work within a chunk (RM409). opentargets build and cache rebuild --only opentargets used to wait, after Ctrl-C, for every part already downloading to stream to its end, so an interrupt about 1 GB into the ~22 GB mirror was ignored twice and needed SIGTERM. The workers now stop at their next chunk, and a part cut short is removed rather than left under its name; the parts that finished stay, and the next run resumes past them. A failed part stops the others the same way, so the failure is reported when it happens.
  • The AlphaGenome AVI build had the same shape over whole contigs and stops at the next block now, with no contig written and no tabix left running.
  • Unchanged: no public name moves. A part whose connection stalls without sending anything still stops only when its next chunk arrives, since bulk downloads carry no read timeout.

2026-11-08 — RM420: cache pull and cache prepare bring a newer published release (patch, cut as v0.8.3)

  • A pulled lane follows the hub's release (RM420). cache pull and cache prepare compare the release label a cache holds with the one its repo publishes, read by the lane's own label reader, and re-pull a cache that came from a pull of that repo when the two differ, keeping a file only where the new manifest attests it. On 0.8.2 a machine that already held a lane stayed on it: gnomAD constraint at gnomad_v4.1_constraint though gnomad_v4.1.2_constraint is published, STRchive at strchive_v2.26.0 against strchive_v2.28.0.
  • A lane built on this machine is left alone and printed with both labels, since it may not be uploaded yet. A builder writes no pull record, and the record now states the release.json its pull placed, so a build over an old pull is not taken for one. cache pull also stops writing a record over an unrecorded build and replacing its release.json with the hub's.
  • Unchanged: enrich never reads the hub for a whole cache, and a hub that cannot be read serves the cache as it is, with a warning. No public name moves.
  • Still needed once: a STRchive catalogue pulled before this patch carries no pull record, so it is left alone; move it aside and pull. On 0.8.2 itself the workaround in INTEGRATION_0_8 § 4 stays.

2026-11-08 — RM418: two NCBI clients in one process share one pacing gate (patch, cut as v0.8.3)

  • An EutilsClient built without a gate takes the process's NCBI gate (RM418), one per interval, so the rsID check and the literature pass, or two enrichments a host runs on two threads, pace against NCBI's one budget per key or address instead of each to the full rate. A gate handed in is still used as handed, and no public name moves. A shared gate's spent now counts every default-built client's attempts against NCBI.
  • Not in it: separate processes still pace apart (RM404, 0.9), and the other services' clients still build a gate each, the two Ensembl REST clients first (RM419).

2026-11-07 — v0.8.2: format, compiler and enricher 0.8.2 (a patch, cut from main)

All three packages move, because each changed source since v0.8.1: format and compiler for RM414 and RM415, the enricher for the rest. The enricher's floors are just-dna-format>=0.8.2 and just-dna-compiler>=0.8.2, the compiler's just-dna-format>=0.8.2. Nothing adds a public name, field, table, flag or warning code; the new names are private in the 0.8 line (RM412 makes them public in 0.9).

  • What it carries: RM408 (the Open Targets snapshot has a repository and a pull route), RM410 (gnomAD constraint 4.1.2), RM411 (cache rebuild --dry-run takes nothing), RM413 (the cache-lane prerelease gates), RM414 and RM415 (two repeat-bin warnings reworded), RM416 (--pin taken or refused), RM417 (gnomAD API constraint rows stop claiming v2.1.1).
  • Release record (RELEASE_RECORDS["0.8.2"], measured from 0.8.0, the last published compiler): 22 reference modules, nothing moved but warnings in 2 (RM414's text). No recompile is needed.
  • After upgrading:
  • Raise floors to >=0.8.2.
  • cache pull --only opentargets fetches Open Targets instead of building it (~5.5 GB against ~22 GB).
  • Rows labelled gnomad_v2.1.1_constraint in a module's gene_metrics.csv hold v4.x numbers (RM417, declared in INTEGRATION_0_8.md § 4); delete the file and enrich again to relabel them.
  • The constraint snapshot moves to 4.1.2 once it is published after this release; existing 4.1 rows are kept on a re-run.
  • A --pin for a lane that cannot take one now stops the run; drop it.
  • cache rebuild --dry-run no longer builds: build without --publish, then dry-run to rehearse the upload.

2026-11-07 — RM414 and RM415: two repeat-bin warnings stop misstating a module (patch, cut as v0.8.2)

  • The repeat- and copy-number warnings name the field the module reads (RM414, S136). When every bin declares a source_field and none is VCF 4.4's own (RUC, CN), measure_field_fractional and measurement_spans_bins name it (FORMAT/REPCN) and speak conditionally, instead of telling the author their field is a Float and sending them to CIRUC. Codes, firing conditions and the published fragments are unchanged; reference_examples/htt_repeat_expansion is one such module.
  • bins_ungrounded no longer says "no bin pmid" when a bin carries one (RM415, S137): with one bin cited, the clause reads "carry no grounding evidence of their own (…; N other bin(s) do cite one)".

2026-11-07 — RM417: gnomAD API constraint rows stop claiming v2.1.1 (patch, cut as v0.8.2; a corrected label)

  • Gene-metrics rows taken from the live gnomAD API are labelled gnomad_api_constraint (RM417), not gnomad_v2.1.1_constraint. The API serves gnomAD 4.1.1 or later: its answer for chrX's DMD, a gene no release before 4.1.1 has constraint for, is 4.1.2's to every digit. The answer names no release, so the label names the route.
  • Declared, per the charter's staleness clause: a row already labelled gnomad_v2.1.1_constraint, in any module enriched since 0.5.0, holds whatever the API served on its fetched_at date, a v4.x number and never v2.1.1. Nothing rewrites it: a re-run counts it as done and adds nothing beside it. To relabel a module, delete its gene_metrics.csv and enrich again, which moves its fact hash. INTEGRATION_0_8.md § 4 carries it as a known issue.

2026-11-07 — RM410, RM416: gnomAD constraint 4.1.2, and --pin taken or refused (patch, cut as v0.8.2)

  • The constraint snapshot is gnomAD 4.1.2 (RM410), was 4.1: 19,511 genes where 4.1 had 18,111, chrX and chrY for the first time, and gnomAD's recomputed pLI, LOEUF, o/e and z-scores (LOEUF moved for 17,214 of 4.1's genes). The builder reads the release's .tsv.bgz and labels the snapshot from the file's own name, gnomad_v4.1.2_constraint; it was the constant gnomad_v4.1_constraint.
  • After upgrading: a module's gene_metrics.csv rows from the 4.1 snapshot keep their label and values on a re-run (a row from any release of the snapshot counts as done); to move a module to 4.1.2, delete the file and enrich again, which moves its fact hash. cache pull --only constraint brings the new snapshot once it is published.
  • cache rebuild --pin is taken or refused (RM416). Nine lanes ignored it and built their default, so --pin constraint=4.1.2 built 4.1 without a word. Seven lanes take one (acmg, civic, clinvar_gkm, constraint, mane, opentargets, strchive); a pin for another lane is refused before the run, by cache rebuild and cache prepare.
  • Known, not fixed here: gene-metrics rows taken from the live gnomAD API are labelled gnomad_v2.1.1_constraint, and the API serves 4.1.1 or later (RM417, open).

2026-11-07 — RM413: no cut while a lane with a repository has nothing published in it (patch, cut as v0.8.2)

  • Every cut now passes three cache-lane gates (RM413, enricher/tests/test_prerelease_gates.py): stage parity (a published lane can be pulled, every publish command has a --dry-run, and a lane built here and published nowhere names what bars it, so "no repository yet" fails), idempotency (cache status, cache rebuild --dry-run and every publish --dry-run write nothing and answer the same twice), and, under JUST_DNA_NETWORK_TESTS=1, a published release in every lane's repository. 0.8.0 passed none of the three for Open Targets. Tests and the release rule only; nothing a consumer imports moves.

2026-11-07 — RM411: cache rebuild --dry-run takes nothing (patch, cut as v0.8.2)

  • cache rebuild --dry-run no longer downloads or builds (RM411, a bug: a dry run that writes is not one). It ran the whole rebuild and skipped only the upload, so rehearsing --only opentargets --publish began a 22 GB download into <out>/opentargets/source/. It now prints, per lane, what a run would build, from where and into which directory, the release this machine holds and, with --publish, the release the lane's repo states (nothing published and could not ask kept apart). A snapshot already in <out>/<lane> has its upload rehearsed.
  • After upgrading: a script that used --publish --dry-run to build and then rehearse now gets the plan only. Build with cache rebuild (no --publish), then run --publish --dry-run to rehearse that build. No public name, flag or warning code; 0.9 adds an opt-in download rehearsal (RM412).

2026-11-07 — RM408: the Open Targets snapshot has a repository and a pull route (patch, cut as v0.8.2)

  • cache rebuild --only opentargets --publish publishes the Open Targets snapshot to just-dna-seq/opentargets, and cache pull --only opentargets pulls it (RM408). The lane shipped in 0.8.0 with a builder and no repository, so every deployment had to download ~22 GB of Platform parquet to cut a ~5.5 GB snapshot, and a registry asking for the lane was told it was unpublished. The release is CC0-marked, and no decision had kept it unpublished: the code named the missing repository as the only gap.
  • After upgrading: cache pull with no --only now fetches this lane too (~5.5 GB) once the repository holds a snapshot; until then it reports not published yet, as it does for clinvar_gkm, and cache prepare no longer builds the lane when it is absent (a pullable lane pulls). Build it locally with opentargets build meanwhile. A cache you already built is left as it is. No public name, field or warning code: the pull route is private in the 0.8 line.

2026-11-07 — v0.8.1: enricher 0.8.1 (a patch, cut from main; format and compiler stay 0.8.0)

Only just-dna-enricher moves. Neither just-dna-format nor just-dna-compiler changed a line of source since v0.8.0, so both keep 0.8.0, are not built, and the enricher's floors stay >=0.8.0. Their uv-build<0.13 bound (below) therefore ships with their next release, not this one. Nothing adds a field, table, parameter, public function or warning code: RM401's and RM403's new names are private in the 0.8 patch line, and 0.9 makes the batch knobs settings with RM404. Recut the same day, before the publish, to take RM398 and RM407; nothing was published under 0.8.1 before them. The compiler is unchanged, so there is no release record: a sweep would compare 0.8.0 with itself, and a compiled module needs no recompile.

  • RM398 — the third placement witness is dbSNP's own record, read from the merge check's esummary, where it was a live Ensembl REST request: independent of the snapshot, and no new request.
  • RM407 — an indel's VRS mint reads the reference once per locus, not once per normalize step.
  • RM401 — the live Ensembl link and the REST placement witness (until RM398) ask 20 rsIDs per request.
  • RM400 — a corrected derivation: live REST no longer writes loci on GRCh38 patch and alt contigs, and a re-derivation leaves recorded ones out instead of reporting lost_allele.
  • RM403 — every retry writes a WARNING line, and every give-up names its attempts and seconds.
  • RM397 — the wheel no longer ships the Atlas .proto sources a second time.
  • Tests and tooling — the cross-tier floor check reads just_dna_compiler.compiler as a package (it read every import of it as absent from 0.8.0), and scripts/build-dists.sh.

After upgrading: just-dna-enricher enrich <spec> --rederive online rewrites a resolution.csv that live REST filled with patch or alt-contig loci (INTEGRATION_0_8 § 4 names the case); new resolution.csv rows name dbsnp where they named ensembl as placement_witness, and sources.csv gains a dbsnp row. No HuggingFace snapshot changed.

2026-11-07 — RM398: the third placement witness is dbSNP's own record, not a live REST request (patch, cut as v0.8.1)

  • The placement witness reads dbSNP's GRCh38 placement from the esummary db=snp record the rsID merge check already fetches (RM398, S132), for the same subjects live Ensembl REST was asked about: snapshot indels neither ClinVar witness holds. S117's misanchored ABO insertion is still withheld. The merge check is served the same records, so an online run asks dbSNP nothing new and REST nothing as a witness. New rows name dbsnp as placement_witness where they named ensembl, and sources.csv gains a dbsnp row at the resolution layer (public domain, release withheld). A run with the merge check off counts those indels as not asked. No public name, field or warning code.

2026-11-07 — RM407: an indel's VRS mint reads the reference once per locus (patch, cut as v0.8.1)

  • Minting an indel's VRS id reads one seqrepo window per locus instead of one HTTP read per step of vrs-python's normalize (RM407): 1,024 bases each side, read once and served from memory; a slice past it is read as before, so no id changes. The registry's replay made 2,472 such reads, about ten per indel. The run's minter also keeps its ids, so the placement comparison and the minting pass normalize an allele once. No public name, field or warning code.

2026-11-07 — RM401: live Ensembl asks many rsIDs per request (patch, cut as v0.8.1)

  • The live Ensembl link and the REST placement witness ask 20 rsIDs per POST /variation/:species instead of one GET each (RM401, S133). Each rsID gets the answer its GET would give: an rsID the batch leaves out is asked by GET, a merged one is found by its synonyms, and a failed batch is could-not-ask for every rsID in it, never not_found. No public name is added: the batch size (20) and the POST's read timeout (120 s) are fixed private constants in the 0.8 patch line, and 0.9 makes them settings with RM404. Only EnsemblResolver batches; an injected resolver is asked one rsID at a time, as before.

2026-11-07 — RM400: live REST no longer writes loci on GRCh38 patch and alt contigs, and a re-derivation leaves recorded ones out (patch, cut as v0.8.1)

  • Live Ensembl REST drops a mapping off the GRCh38 primary assembly (HSCHR17_1_CTG5, HG76_PATCH) instead of writing it as a locus (RM400, S132), as the snapshot and the tier's other readers already do. An rsID REST maps only there is written not_found, as the snapshot answers it.
  • enrich --rederive leaves a recorded off-primary locus out of the comparison and logs one warning counting them. A table REST filled on 0.7 or 0.8.0 no longer reads such a subject as lost_allele; the committed table drops the patch rows. No drift arm, carry reason, result field or warning code is added.

2026-11-07 — RM403: every retry writes a line, and every give-up says how many attempts and how long (patch, cut as v0.8.1)

  • An enricher request that is retried now says so, at WARNING on the just_dna_enricher.net logger: EnsemblResolver._rest_rsid(rs1801133): attempt 1 of 3 failed (ReadTimeout: …); retrying in 0.6 s, and on giving up …: gave up after 3 attempt(s) over 46.2 s (ReadTimeout: …) (RM403, S133). Every client's policy carries it. Until now a timed-out attempt wrote nothing. No exception, return value or warning code moves: the give-up raises what it raised before. A host that filters the enricher's warnings will see more of them during an outage, one per retried attempt.

2026-11-05 — the enricher wheel shipped its Atlas build input beside the generated tree (patch, cut as v0.8.1)

  • The enricher wheel no longer carries _atlas_protos/ (RM397), the fetched AlphaGenome .proto sources the build generates the bindings from. Everything the runtime reads is under generated/, as before; the sdist still carries the sources, so a build from it needs no network. 0.8.0's wheel shipped five files a second time (about 50 KB). Nothing a consumer imports moves.
  • scripts/build-dists.sh builds every workspace package into a fresh dist/ and exits 3 when a wheel's non-.py files pass 3% of its bytes or a wheel holds the same bytes under two names.
  • just-dna-format and just-dna-compiler bound uv-build<0.13 in [build-system].

2026-11-05 — 0.8.0: the cut

Cut as 0.8.0 on 2026-11-05, all three packages at one number: just-dna-format, just-dna-compiler and just-dna-enricher each changed since v0.7.3 and each takes 0.8.0 (PyPI's latest at the cut: format 0.7.1, compiler 0.7.2, enricher 0.7.3). Every intra-workspace floor is >=0.8.0; polars>=1.42.0 stays, because just-dna-lite stays on polars 1.x. Publishing is the maintainer's, and the tag v0.8.0 follows the publish. Every entry below headed (0.8, uncut) ships in it. The surface delta and the per-consumer lists are INTEGRATION_0_8.md.

  • Consumer pre-integration was SKIPPED, by the maintainer's call: an intentional, expedited shipment (RELEASE_CYCLE.md § Minors step 6 lets only the maintainer skip it). The consumers integrate against the published 0.8.0, from INTEGRATION_0_8.
  • The charter wording was vetted. The changed sentences of Principle 4, Principle 7, the Goals bullet on the reference-snapshot layout and the Goals toolchain clause (and the recorded toolchain) were read old → new by the maintainer in chat on 2026-11-04 and accepted ("Good, I accept"). Each 2026-11-04 entry of CONSTITUTION_AMENDMENTS_HISTORY.md records it.
  • The release record, RELEASE_RECORDS["0.8.0"], measured by two sweeps. The release sweep (published compiler 0.7.2 under polars 1.44.2 against 0.8.0 under the locked polars 2.1.0, one spec root) measured 6 of the 22 reference examples: content_signature moved on 1 (hla_alleles, whose meta.csv 0.7.2 skips as unknown), every other axis on all 6. The other 16 are unmeasured: each commits a resolution.csv written by 0.8, which 0.7.2 refuses. With polars held at 1.44.2 the same six move the same way, so every byte move in the interval is the compiler's. The fixed-compiler sweep (0.8.0 under polars 1.44.2 and 2.1.0) moved parquet_bytes on 22 of 22 and nothing else. Because it explains no measured module on its own, the gate refuses a toolchain declaration, and none is made; the record's evidence states the move. toolchain_differs is the normal case between 0.8 and just-dna-lite (polars 1.x against 2.x): one compiler, two digests, one content_signature. Declared: RM363's contig spelling and RM361's :None key as corrections; the manifest additions of RM300, RM304, RM256, RM28, RM324, RM331, RM333 and RM292; the parquet columns of RM270, RM331, RM364, RM380, RM387 and RM325–RM336; RM380's DuckDB-path parity, which holds for facts and not bytes. The example data that moved (the split-snapshot regeneration, pathogenic_clinvar's duplicate row, cpic_gene_pairs' NUDT15 anchor) is in the record's evidence, not declared, since no reach grammar confines a declaration to the examples. sweep --toolchain-before … --release 0.8.0: release record for 0.8.0 covers the measurement. RM342's tests pass under polars 1.44.2 and 2.1.0, and a manifest polars 1.44.2 wrote reads under both.
  • Known issues (INTEGRATION_0_8 § 4): RM385 (offline enrich cannot left-align an indel, for 0.9), RM376 (an offline and an online fill of one snapshot can differ), RM388 (placement readers that bypass usable_loci, parked), RM390 (the enricher does not say what compile will refuse, parked) and RM386 (publish pinning with no live run yet).
  • The release-gate sweep added literature.quotes_checked (RM256) to INTEGRATION_0_8 § 1.2 and a § 1.44 of the new public names the reference docs document, and filed every closed item of RM_TOC_300_399.md under Shipped. The P3 scan: the 11 published modules in the registry mirror cloned here get the same validate verdict and errors under 0.8 as under 0.7.2 (three are invalid under both, for reasons older than 0.8), and no module there or in reference_examples/ holds a DOI in pmid or an HLA name 0.8 refuses.
  • The roadmap files. ROADMAP_0_8.md closed into history/, and its one waiting item, RM28's predicate half, moved to ROADMAP_0_9.md, the open minor file. PROPOSAL_0_8_PT2 is concluded.

2026-11-05 — RM394: the re-derivation's report says what happened, and the remedy each case needs (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • The drift label has seven arms. An allele the source added (A → A,T) reads "gained an allele" where it read "moved: the source names a different event" (94 subjects of the examples' regeneration); an rsID newly named reads "gained an rsID". SubjectDrift.case carries the arm (respelled, gained_allele, lost_allele, gained_rsid, renamed, moved, undecided); respelled reads as before. The drift warning counts by arm.
  • The derived-layer currency line names the release that is behind, one line per source. When the local snapshot is the release the table recorded, it says cache pull --only <lane> first: a re-derivation reads the same snapshot and would repeat the line. When the snapshot is current, it says enrich --rederive.
  • The carry-forward groups the subjects it kept by why, each with its remedy; "no source could be asked" is said only of subjects nobody could ask.
  • One module, enrich/rederive_words.py, maps each case to its sentence; a test walks every case.

2026-11-05 — RM393: a re-derivation with no snapshot keeps a positional row's back-filled rsID (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • enrich --rederive no longer erases what it could not ask. A coordinate-only row's rsID comes from a position→rsID map, which only the Ensembl and ClinVar snapshots supply. With neither read (a bare machine under --offline, or a failed pull), the re-derivation wrote the row rsid empty and source=authored, and the recompiled weights.parquet/annotations.parquet carried rsid=null. A recorded back-fill whose map was not read this run (per link: a failed Ensembl pull beside a ClinVar snapshot counts) is now carried forward unchanged, so the module recompiles to the same artifact.digest. A map that was read and has no rsID at the allele still replaces it. Measured on hfe_hemochromatosis (audit round 3).

2026-11-05 — RM395: the commands the user guides run do what the guides say (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • just-dna-compiler keygen --out <path> creates a missing parent directory, as compile and reverse do, instead of a traceback.
  • just-dna-enricher draft-repeats without --catalogue reads the provisioned STRchive lane, as its --help says and check-repeat-bands does.
  • cache pull ends with what the run left on disk and which lanes had nothing published (nothing pulled; nothing published for: clinvar_gkm), never caches available: and the registry's list, and a pull that found nothing published leaves no empty directory for cache status to call occupied. cache status names both routes for an absent lane that can be pulled or built. ENRICHING.md quotes the re-run output.

2026-11-05 — Docs pass: the user guides re-run against the 0.8 candidate, and 0.8's surfaces added (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

RELEASE_CYCLE § Minors step 8, run at 64c835b9. Every command in index.md, GETTING_STARTED.md, COMPILING.md, CONSUMING.md and ENRICHING.md was re-run on a scratch copy of the module it uses. Quoted outputs now match the run (apoe_epsilon's digest is c86809… with three parquets; its content_signature did not move). Three commands had been wrong as written and are corrected: hint in COMPILING takes a kind and --file/--row, hint variant/hint citation take --rsid/--pmid, and draft/draft-clinpgx draft nothing from a no-sale source without --use. scaffold stubs only the kinds named with --kind.

Added, each with a command that ran: validate's strict switches and the refusals new in 0.8 (COMPILING), meta.csv and model_organisms.csv with their table pages, the studies.csv columns and the pmid NONE + doi pair (COMPILING, CONSUMING), what 0.8 added to the artifact including source_release inside resolution_loci (CONSUMING), and in ENRICHING enrich --rederive after the snapshot switch, cache pull --only ensembl, SNAPSHOT_REPO_IDS, the ClinVar-GKM placement witness, the Open Targets lane and the publisher's PublishRefused family. CLAUDE.md's doc map gains the INTEGRATION_0_8.md row.

Two guide commands still fail as written because the code is wrong, and are left for the fix rather than worked around: keygen --out into a directory that does not exist raises a traceback, and draft-repeats without --catalogue does not find a provisioned STRchive lane although its help says it does. Both are fixed in code by RM395 (above).

2026-11-05 — hboc_palb2 and pathogenic_clinvar regenerated against the split snapshot (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • The two examples the first regeneration withheld are re-derived online now that RM381, RM382 and RM387 are on 0.8, so every example that commits a resolution.csv answers from ensembl-variation-2026_07_13-vep116.0. Both pass validate --strict and compile → reverse → compile is a fixed point; neither raises rsid_placements_disagree, positional_identity_contradicted or derived_row_orphan. pathogenic_clinvar's rs613985 resolves again (ClinVar, 11:5226724).
  • pathogenic_clinvar drops one authored row, the coordinate-only 11:5226762 C>CAAAG, which dbSNP 157 now files under rs80356821 with the same genotype and annotation as the module's own rs80356821 row. Its content_signature moves and the module was closed again. The first regeneration's entry below, its table and INTEGRATION_0_8 § 1.29 now count all 16 examples.

2026-11-05 — INTEGRATION_0_8 names the known issues 0.8 ships with (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • INTEGRATION_0_8 § 4 lists what 0.8 ships with unfixed, each line with its item: RM385 (offline enrich cannot left-align an indel, for 0.9), RM376 (an offline and an online fill differ), RM388 and RM390 (parked, each with its return condition), and RM386's first live run being the next real snapshot publish. The status section says the release candidate holds every item listed as able to land, but the two examples' regeneration and audit round 3.
  • RM362's and RM374's **Residuals** name RM388 and RM390, their parked remainders.

2026-11-05 — RM392: a count beside a registry or table is dropped or compared, in the docs and in code comments (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A total stated in the comment above a registry must be its size. test_source_counted_prose.py walks every module-level registry in the three packages; a bold headline number (**Nine, and …) or a number before the registry's own noun ("the three states" over LANE_STATES) must equal it. Its first run found OVERRIDABLE_TABLES saying nine over twelve, and VALID_CAID_OUTCOME and VALID_RECOVERY_OUTCOME saying three and four over five.
  • SCHEMAS.md's lists are compared to their registries as sets. The authored roster gains meta.csv and drops gwas_effects.csv (a derived sidecar its own paragraph leaves out); the overlay key table gains locus_to_gene.csv, colocalisations.csv and consequences.csv; the exclusions sentence names ld_pairs.csv, the third table outside the overlay.
  • COMPILER.md describes reverse as it is since RM364 and RM380 (the table's own loci, every column but RESOLUTION_LOCUS_EXCLUDED), and pipeline.py no longer calls two resolution ladders compile-only after RM367.
  • Stale counts dropped in INTEGRATION_0_8.md (the excluded locus columns, the 0.8 parquets), ENRICHER.md, tables.py, binning_checks.py, fact_checks.py, civic_vcf.py and identifiers.py. No code path or published surface changed.

2026-11-05 — RM391: no test reaches the network without JUST_DNA_NETWORK_TESTS=1, whichever client it uses (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • The root conftest.py fences the transport: socket.getaddrinfo, socket.socket.connect/connect_ex and gRPC's channel constructors refuse every host but this machine unless the opt-in is set. The error, NetworkReachedInTests, names the test and the host, and a test whose body swallowed it fails at teardown. JUST_DNA_NETWORK_GUARD_LOG=<file> lists every refusal.
  • The 31 tests in 7 files that asked ClinGen, Ensembl REST, eutils or the seqrepo proxy, and passed on whatever came back, now script those clients or turn the incidental pass off; none lost its point.
  • enricher/tests/test_the_network_is_fenced.py drives httpx, requests, urllib, huggingface_hub and gRPC into the fence and holds the enricher's imported transports to that set. Nothing for a consumer to do: the packages are unchanged.

2026-11-05 — RM361, audit round 2: a 410 from a bulk download or a function-style fetch is its lane's *Retired (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A source that retired a file now says so in every lane, not only in the clients. net.stream_to_file takes the lane's *Retired as a required retired= keyword. A 410 raises it with the body quoted and "upgrade just-dna-enricher", followed by the file and the lane's own remedy (a release the operator picked can be the thing that went), where it raised the lane's outage type ("could not download"). That covers every bulk-download lane. clingen.fetch_curation_list, gene_validity.fetch_validity_export, mane_build.discover_current_release and acmg.fetch_acmg_page raise theirs too. The new types are AcmgRetired, CivicRetired, ClinGenRetired, ClinPgxRetired, ClinVarRetired, ConstraintRetired, DrugLabelRetired, GeneValidityRetired, ManeRetired, MitomapRetired, OpenTargetsRetired, PubMindRetired and StrchiveRetired, each a subclass of its lane's *Unavailable, so every existing handler still catches it.
  • clinvar_gkm.lane_currency records retired (…) for a 410, where it wrote unreachable (…). The placement witness's not-asked sentence says Ensembl REST answered HTTP 410 (a fixed sentence; the source's body stays in the log, out of the attested record) instead of "unreachable after retries".
  • Why RM361's guards missed these: the client roster walks classes, and these were functions; the handler walk compared constants, and missed an f-string. Now a walk finds every module-level class and function that calls an httpx transport, drives each function with a 410, a 503 and a refused connection, and asserts the 410's type. A second walk checks each stream_to_file call's retired= against its error_cls. Record: ROADMAP_HISTORY § RM361, INTEGRATION_0_8 § 1.26.

2026-11-05 — RM386: a publish refuses a concurrent writer, and a snapshot's description describes the bytes that were sent (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

Every write the enricher makes to Hugging Face now goes through one helper that commits on top of the revision its plan read (parent_commit), and every remote read after that names the revision. A second publisher or a prune landing in between makes the hub refuse the write, raised as the new upload.RemoteMovedError (a PublishRefused), where before it was overwritten unseen; this covers upload_module (the versioned-path check and both commits), publish_reference_snapshot (the orphan check, the retirement listing, the payload and the description) and cache prune. A snapshot publish hashes every file once (SnapshotPlan.sha256, with SnapshotPlan.revision), checks the tree the hub holds at the payload commit against those hashes, and only then sends release.json as the bytes it read, so a rebuild during a publish can no longer put the next release's description over this one's rows; a mismatch withholds the description (PlanChangedError). prune_repo deletes exactly the listed paths at PrunePlan.revision, and refuses a plan with none; cache prune --yes prints PRUNE REFUSED instead of a traceback. Because the library skips an unchanged write without checking the precondition and pins only the first part of a multi-commit upload, each write also walks the history back to its parent and refuses when a commit there is not its own. An AST walk holds every hub write to the one helper and every read to a revision; the tests run on a new fake_hub fixture, a repository with a head. → ROADMAP_HISTORY § RM386

2026-11-05 — RM389: a locus with a vrs_spec and no id survives the round trip (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • resolution_loci is withheld only when the row's cells rebuild the locus on every column the cell rebuild writes. The comparison was a hand-kept tuple of six that left out vrs_id/vrs_spec, so a locus stating vrs_spec with no id (legal) or an id on a row with no ref was withheld and reverse wrote the value back empty. It now reads parquets.CELL_FACT_FIELDS, derived from _CELL_REBUILT. Reachable only from a table without provenance (hand-filled, reversed from an older artifact, or the deprecated path's); an enriched table's loci were already kept. Where it bites, resolution_loci is filled where it was null, so artifact.digest moves; content_signature does not.
  • Guarded by an equality with _CELL_REBUILT, a walk over every carried column, and a round trip over every reference example with a resolution.csv (test_reproduced_locus_compares_every_rebuilt_column.py). The 0.8 audit, round 2, class 4.

2026-11-05 — RM384: every pulled lane's repository has one public name, the Ensembl lane's included (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

locations.SNAPSHOT_REPO_IDS maps each lane this tier pulls to its HuggingFace repository (org/repo), so a consumer pinning manifest.compilation.ensembl_reference takes just-dna-seq/ensembl_variations_split from SNAPSHOT_REPO_IDS["ensembl"] instead of spelling it or stripping a private prefix (S130). It is the only place a repo id is written: SNAPSHOT_REPO_PREFIXES is derived from it as datasets/<repo id>/data, the upload.DEFAULT_*_REPO_ID names stay as aliases of its entries, every CacheLane.publish_repo reads one, and four --repo help texts that wrote their default by hand now interpolate it. The private download._*_HF_PREFIX constants are gone; every prefix and pull record is byte-identical to before. A test asserts the equalities over the walked map and an AST walk over the three source trees refuses a repo id written anywhere else. → ROADMAP_HISTORY § RM384

2026-11-05 — RM387: a positional row at a sibling locus of a one-to-many rsID no longer refuses a valid module (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • RM374's rsid_placements_disagree compares against the source's whole answer. The rows under an rsID's own key are the loci its genotype can sit on (the hosting filter runs before they are written), while a positional row's rsID is back-filled from the unfiltered answer, so a positional row at the sibling locus read as a second placement, compile --strict refused, and --rederive wrote the same table. The rsID-keyed rows now record every place the links offered in a new derived resolution.csv column, source_loci (chrom:start, comma-joined; written only where it is wider than what the rows place, outside resolution_signature), and the check reads the union. A table recorded before the column still fires; enrich --rederive writes it.
  • The parquets' resolution_loci struct carries source_loci, and reverse writes it back, so the second lap reads the first lap's answer. artifact.digest moves on a recompile of a module carrying the struct; content_signature does not.
  • New just_dna_format.resolution.source_loci_cell / split_source_loci, the one spelling both tiers use.

2026-11-05 — RM383: a re-derivation retires the licence row for a source no resolution row names (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • enrich --rederive removes a resolution-layer licence row whose source no committed row names (mt_common_deletion's clinvar,resolution, left behind when its loci moved to the Ensembl snapshot), inside the table's commit, with an info line naming it. The compile's source_row_unused for it goes away. A plain enrich never removes a row.
  • The concordance record, rewritten whole on each run that writes it, retires an authority's clin_sig_authority_call row once no call row names it.
  • licensing.record_table_sources takes answered_every_row (default False); every call in the package states it.
  • reference_examples/mt_common_deletion is re-derived with this in place (enrich --offline --rederive, as its regeneration ran): the regeneration had run before RM383 merged and kept the row. The row is gone, so is the compile's source_row_unused, and resolution.csv gains RM387's source_loci header. content_signature is unchanged; compile → reverse → compile gives one artifact.digest. The RM383 tests plant the old row on their copy.

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A rejected answer no longer ends the chain. Each link's answer passes the allele-aware filter as it arrives; one that every subject of the rsID rejects (one of them rsID-only) is set aside and the next link is asked (ClinVar cache, staged answers, live Ensembl, gnomAD). pathogenic_clinvar's HBB rs613985, whose only split-snapshot locus is 3:159434329 T>G, now resolves from ClinVar at 11:5226724 CATAACAGCATCAG>C.
  • An rsID-only subject every link rejected gets no resolution.csv row — this reverses S85 (0.7), which kept the status=not_found row beside the "the source HAS these variants" warning; the maintainer accepted the reversal on 2026-11-05. INTEGRATION_0_8 § 1.36 and § 2 step 23 say what a reader of the table sees now. It stays unresolved; strict names the rejection in its own arm. --rederive drops a recorded not_found for it and keeps a recorded placement.
  • AlleleMismatch gains links and placed_by; EnrichmentResult.superseded_answers names each set-aside answer and the link that placed the rsID after it.

2026-11-05 — RM381: a split-snapshot site is filled ALT by ALT in the caller's spelling, offline too (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • Every reader of a snapshot's alt spells each ALT alone (alleles.trim_allele, regrouped by sequences.per_alt_loci): the Ensembl and ClinVar rsID lookups, REST's allele_string, gnomAD's variant ids and the three position→rsID readers. The split repo pads every ALT of a site to one ref, and the readers grouped by it, so rs1555461597 was filled as 16:23635706 GTT>G,GT,GTTT; it is now G>GT and GT>G (GTT>G is the 2 bp deletion its genotype cannot host), and SHOX's rs1569493663 regains C>CAG beside CAG>C. The trim reads no reference; left-alignment past it is still online only, so an offline indel piece records left_aligned unknown.
  • A caller-spelled coordinate back-fills its rsID from the split rows (16:23635706 GT>G → rs1555461597).
  • A genotype authored in a site's padded frame is read in it, by trimming its alleles with the same helper (site_verdicts, both tiers): GTT/GTTT names the insertion G>GT alone, the compound het G/GTTT names GTT>G and G>GT, and the old repo's AGA/AGAGA names SHOX's duplication alone. A caller-spelled genotype reads as before. No new column or code.

2026-11-05 — INTEGRATION_0_8 is finished for downstream to integrate against the branch (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • INTEGRATION_0_8.md now covers every § 1 item from § 2 and § 3: an upgrade order (format, compiler, enricher pins; uv lock; the ensembl_variations cache moving to the split release; recompile; enrich --rederive after the snapshot switch), § 2 steps 12–21 for RM267 through RM379, and per-consumer lists for just-dna-registry (checked against its M1–M11 map), just-dna-lite, just-dna-pipelines, just-dna-agents and the new longevity-loci. The status line is the honest pre-cut one (0.8 at 1007a26d, not on PyPI, git dependency for testing only), and the file records that consumer pre-integration (RELEASE_CYCLE § Minors step 6) was skipped for 0.8 by the maintainer's call. Two wrong pointers fixed (RM368 is § 1.27, RM267's undecided § 1.18), and RM_TOC's RM364 row now points at § 1.28. Docs only.
  • Merged beside RM380 and the examples' regeneration, it gains RM380 as § 1.34, the regeneration's paragraph under § 1.29, and a § 2 step 21 for a tool that parses a reversed resolution.csv (the levers are step 22 now). Steps 12, 14 and 15 and the § 1.19, § 1.28 and § 1.32 rows say what RM380 carries, and the status names what has landed since 1007a26d and what still can: RM381, RM382, RM383, hboc_palb2 and pathogenic_clinvar regenerated, and the second audit round.

2026-11-05 — The reference examples' resolution.csv regenerated against the split snapshot (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • All 16 examples that commit a resolution.csv were re-derived (enrich --rederive) against just-dna-seq/ensembl_variations_split, release ensembl-variation-2026_07_13-vep116.0 (the maintainer's decision of 2026-11-04): 14 in the first pass, and hboc_palb2 and pathogenic_clinvar once RM381, RM382 and RM387 had landed on 0.8. Every rewritten row now records source_release, and RM267's, RM270's and RM375's columns. Ten ran online (the two later ones among them), because their committed tables held what only an online run produces (an indel vrs_id, an rsid_status, a REST-answered row, or online checks in verification.json). Six ran offline (marked off below), because their tables held nothing an offline run would lose.
  • content_signature did not move on any example but cpic_gene_pairs and pathogenic_clinvar, where an authored row was corrected or removed (below). compile → reverse → compile is a fixed point for artifact.digest and every signature on all 16 (verification is outside the round trip by REVERSE_MANIFEST_EXCLUDED). Every example passes validate --strict, no example raises rsid_placements_disagree or positional_identity_contradicted, and the only authored_event verdict in the corpus is rs761191455's same_event (cpic). Every closure survived except cpic's and pathogenic_clinvar's, which were closed again.
  • cpic_gene_pairs: NUDT15 *1 rs1457579126 is authored at 13:48040977 GA, no longer at 48040982. 48040982 is the HGVS deleted-base position of a one-base deletion in G AAAAA T (977–983), inside the event and not a VCF anchor. The row now uses the snapshot's left-aligned anchor and the *1 (reference) allele. The README records the edit, since the example keeps no logs/. assets/consumer_join grew its chr13 window to cover the row.
  • hboc_palb2 and pathogenic_clinvar were withheld from the first pass and regenerated after RM381. Re-derived before it, they spelled six multi-ALT indel alleles with the snapshot's shared padded ref (16:23635706 GTT>GT, 11:5226773 TGG>TG, …), which a bcftools normed VCF never meets, and test_consumer_join refused it. With the per-ALT trim every indel keeps its caller spelling, the hosting filter leaves out the pieces the authored genotype cannot carry (16:23634953 CAA>C, 11:5226773 TGG>T), and no row of frequencies.csv or clinical_assertions.csv in hboc_palb2 is orphaned, so neither pass was re-run. pathogenic_clinvar's rs613985 resolves again: both snapshots place it where its authored C/CATAACAGCATCAG cannot sit, and since RM382 the chain asks ClinVar, which has it at 11:5226724 CATAACAGCATCAG>C; source_loci names the three placements it saw.
  • pathogenic_clinvar: the coordinate-only row 11:5226762 C>CAAAG is removed, with its studies.csv row. dbSNP 157 files that insertion under rs80356821, beside the deletion CAAAG>C the module authors by that rsID with the same genotype string C/CAAAG, so the rsID row expanded onto the positional row's key and --strict refused Duplicate (variant, genotype). The two rows state the same annotation (risk, pathogenic, HBB, clin_sig=pathogenic, PMID 29165669), so the positional one went and the rsID row covers both events. The example's README records the edit, since it keeps no logs/.
  • Seven test files read a shape off an example that the regeneration changed. Each now builds that shape on its copy, so the test still checks what it was written to check. apoe_epsilon had no ALT (test_verification_block's no-ALT line, test_opentargets' whole-release case). hfe_hemochromatosis had no verification records and no concordance record (test_verification_record, test_clin_sig_concordance). rs111033563 lacked G (test_studyrow_0_8's wrong proxy allele is now T, the lead's own). shox_par1 carried a split locus (test_rsid_placements_disagree puts the old pair back on a copy). And assets/consumer_join did not cover cpic's corrected row.
Example Subjects Respelled Allele gained rsID filled Source only Other derived files rewritten Digests that moved
apoe_epsilon (off) 2 0 2 0 0 new licensing.csv artifact, resolution, sources
apoe_klvs_meta 6 0 2 0 0 licensing.csv artifact, resolution, sources
cpic_gene_pairs 62 2 14 3 0 licensing.csv artifact, resolution, sources, content
cyp2c19_star_alleles (off) 35 0 10 0 0 licensing.csv artifact, resolution, sources
cyp2c9_warfarin_grch37 5 0 0 0 0 none none (GRCh37: nothing re-asked)
hfe_compound_het 2 0 2 0 0 new licensing.csv artifact, resolution, sources
hfe_hemochromatosis 12 0 4 0 8 sources.csv, new concordance pair artifact, resolution, sources, concordance
longevity_studies (off) 6 0 1 0 0 licensing.csv artifact, resolution, sources
model_organisms (off) 1 0 1 0 0 licensing.csv artifact, resolution, sources
mt_common_deletion (off) 3 0 0 0 2 licensing.csv, new concordance pair artifact, sources, concordance
par_boundary 3 0 0 0 3 licensing.csv artifact, sources
pgx_slco1b1_simvastatin (off) 1 0 0 0 0 licensing.csv artifact, sources
pleiotropy_tradeoff 3 0 0 0 0 licensing.csv artifact, sources
shox_par1 10 1 2 0 1 new concordance pair artifact, resolution, sources, concordance
hboc_palb2 16 0 5 0 11 licensing.csv artifact, resolution, sources
pathogenic_clinvar 327 0 51 7 249 new licensing.csv, new concordance pair artifact, resolution, sources, concordance, content

Respelled: a subject's locus moved (cpic's two NUDT15 indels; shox's split rs1569493663 is now the one locus X:634689 CAG>C, which hosts its authored C/CAG). Allele gained: the snapshot lists ALTs the old table did not, or the row had none (apoe_epsilon, hfe_compound_het). rsID filled: a coordinate-keyed subject gained its rsID. Source only: the facts are unchanged and the answering link moved (clinvar or ensembl-rest to the snapshot). The sources signature moved everywhere because each licence row is now labelled by the release its rows state (RM373). The concordance pair (clin_sig_concordance.csv, clin_sig_authority_calls.csv) is a 0.8 enrich output, first written into these four examples by this run. verification.json was rewritten on all 16. The new warnings: clin_sig_concordance_contested (three examples with one subject each, and pathogenic_clinvar with 21, 4 of them opposed calls, every one between the two authorities themselves), verification_findings_recorded on hboc_palb2 and pathogenic_clinvar (a dataset_currency finding: ClinVar has published clinvar_2026-11-01 since the snapshot these were answered from), and on mt_common_deletion source_row_unused, for the clinvar,resolution licence row that no resolution row uses any longer.

The Digests that moved column compares each example against its own 0.7-era table, before RM380, which landed beside this regeneration and widens the locus struct (the two later rows compare against the same committed table compiled at 80506f72, after RM380, so they show only what their regeneration moved): it moves artifact.digest again on 20 of the 22 examples, cyp2c9_warfarin_grch37 included, and moves no signature. The final digests are measured by the release record at the cut, not read off this table.

2026-11-05 — RM380: reverse gives back every resolution.csv column a parquet carries (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compile → reverse → compile now keeps authority, placement_witness/placement_verdict, left_aligned, source_release, caid and status on every resolution.csv row a parquet carries, so a best-effort ambiguous pick is no longer laundered into resolved by a round trip. The locus struct (resolution_loci, proxy_loci) is derived from ResolutionRow minus parquets.RESOLUTION_LOCUS_EXCLUDED, a map naming each left-out column with its reason, and a test walks the model against the two. A key whose locus holds any of those columns keeps its loci, so a cleanly resolved module carries them too. Reverse's resolution.csv header is derived from the same list; its variants.csv/studies.csv headers and the enricher's frequencies.csv/gene_metrics.csv headers are held to their models by equality tests. compile --strict's refusal for an authored rsID↔coordinate pair the table contradicts now says the table's position is carried and the next compile reaches the same finding, instead of saying it is lost. 20 of 22 reference examples get a new artifact.digest; no content_signature or resolution_signature moves. → ROADMAP_HISTORY § RM380

2026-11-05 — RM373 amendment: a row Ensembl REST witnessed contributes REST's release (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • The placement witness's release map covers every witness. It named ClinVar and ClinVar-GKM and predated the REST witness, so a REST-witnessed row stated no release for it; the REST witness now asks /info/data once (or reuses the live link's answer), and both release maps, by link and by witness, refuse to drift from their registries. The ensembl licence row of such a run is still withheld: the snapshot and REST spell one Ensembl release two ways.

2026-11-05 — RM379: every authority the concordance record names has its licence row (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • enrich records ClinVar's and PubMind's terms at the new clin_sig_authority_call layer for each authority a clin_sig_authority_calls.csv row names and that was consulted, inside that table's commit, dataset labelled by the release the rows state. An authority nobody asked writes none; CIViC, whose checks write no table row, writes none. The compiler does not corroborate the layer (its authority column admits manual), so the new rows raise no source_row_unused.

2026-11-04 — RM373: every resolution row names the release that answered it, and no label speaks for a table two releases answered (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • resolution.csv gains source_release: the release of the link that answered the row, as the lane states it (ensembl-variation-2026_07_13-vep116.0 from the split repo's manifest, clinvar_<date>, live REST's ensembl_116 from /info/data; none for gnomAD, whose gnomad_r4 names a release line). Written only by the run that answered the row, never back-filled; --rederive stamps every row. Outside resolution_signature; reverse writes none.
  • Every derived pass labels its licence row by the committed table (licensing.record_table_sources): the one release its rows state, withdrawn when they state two or any states none. enrich and four fact passes left dataset blank; five more kept their first run's label forever. A row is labelled by its source's own release, never by the day it was asked: gwas reads the Catalog's from /api/search/stats, the AlphaGenome atlas states none (alphagenome_atlas), assertions uses clinvar_dataset_label (an unstated release is clinvar, no longer clinvar_unknown), and gene-metrics reads the snapshot's own release.json rather than a constant.
  • rsid_coordinate_agreement.release is withheld over a mixed table, with the count of rows another release, or no recorded release, answered. The probe's case (an all-old table re-enriched against the new snapshot without deleting it) now reads null and says why. Rows live REST answered are not counted against the snapshot's label.
  • A derived table's source publishing past it never makes strict refuse (@a-source-recuring-is-not-a-strict-matter): on every layer but annotation the currency finding warns in both modes and names enrich --rederive; only a drafted layer refuses and says re-draft.

2026-11-04 — RM375: two spellings of one indel are one event, judged by the enricher and trusted by the compiler (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • enrich records whether an authored coordinate spells the resolved event: resolution.csv gains authored_event (same_event / different_event / undecided), written against the reference where the authored chrom/start/ref differs from the locus, and judged_coordinate, the coordinate it judged, so a verdict never outlives an edit or speaks for another table's row. NUDT15's rs761191455, authored at the old snapshot's 13:48041104 and placed by the new one at 13:48041103 T>TG, reads same_event online and undecided offline.
  • The compiler reads the record and judges nothing itself. A recorded same_event draws no positional_identity_contradicted and no rsid_coordinate_disagrees, and fills only rsid/chrom; undecided keeps the finding without accusing either side. The verdict rides resolution_loci through reverse.
  • rsid_coordinate_agreement no longer reads an insertion's one-base ref as a substitution, so a respelled indel is agreement online and undecided offline, never a finding.
  • --rederive says respelled, moved, or undecided: SubjectDrift.respelled is three-valued, and a change this run cannot settle (no vrs_id, no reference read) is warned apart from the drift.
  • One judge for every comparison of two spellings, just_dna_enricher.events, with a registry test and an AST walk over its callers.

2026-11-04 — RM374: one rsID placed two ways in one resolution.csv is reported, and refused under strict (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • rsid_placements_disagree, a new warning code. read_resolution asks, for both validate and compile, whether a resolution.csv places one rsID at one (chrom, start) under its own key and at another under a positional key whose rsID was back-filled: what merge-not-clobber leaves when a module is enriched before a snapshot moves and again after it. The 0.8 migration probe's rs761191455 (13:48041104 G>GG beside 13:48041103 T>TG) now draws it. The sentence names the rsID, each answer, the releases (ResolutionRow.source_release when rows record it, else releases not recorded) and the remedy, enrich --rederive or a deleted resolution.csv. A warning in best_effort; compile --strict and validate --strict refuse.
  • One key is one answer. A one-to-many rsID, a pseudoautosomal pair and a multi-ALT locus split into pieces at one start (shox_par1's rs1569493663) never fire; no reference example does. Where the table cannot tell one answer from two (no row keyed by the rsID itself, one key against itself, two spellings the enricher judged left_aligned true and false) the check withholds.
  • A resolved allele refusal names the stale reading. genotype_allele_not_at_locus and effect_allele_not_at_locus, on alleles an external source put in resolution.csv, add that the row may predate the snapshot, name the release it records (or release not recorded) and give --rederive as the remedy. The probe's rs646776 A/C refusal blamed only the author or ClinVar.
  • See COMPILER § One rsID, one placement set and INTEGRATION_0_8 § 1.30.

2026-11-04 — RM376: why one rsID fills three ALTs offline and one online on the same snapshot (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • No code change. The 0.8 migration probe saw rs746071566 filled with three ALTs offline and one online from one split snapshot. The snapshot, Ensembl REST and dbSNP 157 all carry the same three alleles at one site, and the resolver groups them correctly. The difference is RM270's rule: online, each ALT is left-aligned against the reference and the piece the row's genotype names is written; offline there are no reference bases, so the locus is written whole with left_aligned left unknown. enrich --rederive online respells such a row. Closed in ROADMAP_HISTORY § RM376.

2026-11-04 — RM377: the two refusals only strict adds have published phrases (patch, on main)

compiler — errors carries no code, so a consumer telling one strict refusal from another reads its words. The two refusals strict adds rather than promotes now have phrases that are API, published in COMPILER.md's Warning texts a consumer keys on: have unresolved genomic positions after resolution (resolution left rows without a coordinate) and whose coordinates the enricher diagnosed as another assembly's (the module's rows contradict its genome_build). Both open with strict compile:, which nothing else opens with. A test reads the phrases from that table and walks the source, so a third such refusal cannot ship unpublished. No message, field or identity moves. Coded errors are RM378, a minor.

2026-11-04 — RM372: the enricher reads the Ensembl snapshot from its new repo, and never serves two releases as one (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • ensure_snapshot pulls just-dna-seq/ensembl_variations_split (dbSNP 157, VEP 116, one row per ALT, left-aligned indels), prepare-annotations' release of 2026-11-04. Every lane's repo is a row of download.SNAPSHOT_REPO_PREFIXES.
  • A cache pulled from the old repo is re-pulled by cache pull, never returned as current: the file names are the same, so the pull record's repo is what tells them apart. enrich still reads such a cache as the older release it is.
  • The release's manifest.json is pulled first and attests every file: a kept or downloaded file whose size and sha256 differ is refetched or refused. A cache holding a data/ file its manifest does not attest (old chrMT beside new chr21), or lacking one it attests, is withheld whole by the resolvers, with a warning naming the files and which of the two it is.
  • cache status and rsid_coordinate_agreement.release name the release (ensembl-variation-2026_07_13-vep116.0), read from the manifest by locations.ensembl_release_label.
  • No test reaches the real HuggingFace hub unasked (@no-test-reaches-the-real-hub): an autouse fence in the enricher conftest, after a test re-pulled six chromosomes of a shared cache on this branch.

2026-11-04 — RM364: reverse gives back everything the compile carried, or names why not (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • provenance.json and the logs come back. A compile copies both into the artifact and hashes them into manifest.provenance / manifest.logs; reverse copied back only the README and logo, so the second lap's manifest had neither. All four now go through one list, reverse.shipped_files, each copied byte for byte after a hash check; a file that disagrees with its record, or a name that leaves the module directory, is not copied.
  • Every manifest field is on one side of a registry. REVERSE_MANIFEST_RECONSTRUCTED and REVERSE_MANIFEST_EXCLUDED (with a reason each) cover every ModuleManifest and Compilation field, asserted by a walk over both models. The corpus sweep and the every-field fixture compare reconstructed_projection lap against lap, which replaced two hand-kept projections that both missed logs and provenance.
  • A variant row is rebuilt from what the table said. weights.parquet and the three positional parquets gain resolution_loci, the loci resolution.csv files under each row's authored key (the proxy_loci shape), kept wherever the row's own cells would not rebuild them and null on a cleanly resolved row. Reverse used to rebuild a row the fill left alone from its own cells, losing the rows the fill did not take and writing the authored coordinate back as a resolution fact: the contradiction went silent on the second lap and resolution_signature moved. Two positional matrix cases strict accepts, and five weights-side cases it refuses, are now fixed points. Every module carrying variants.csv or a positional table gets a new artifact.digest on recompile (a new column); content_signature does not move.
  • No resolution.csv comes back for a module that read none (grch37_build, mt_heteroplasmy, cyp2d6_structural), so the second lap no longer publishes a signature and VRS coverage the first did not. resolution_signature is now compared from the first lap on every reference example, with no exemption; cyp2c9_warfarin_grch37 was the one the GRCh37 exemption hid (RM69, addendum).
  • reference_examples/cpic_gene_pairs has its NUDT15 *1 row rs1457579126 back, removed while the example was built because of this defect; re-enriched (enrich --strict, which added the ClinVar licence row for the placement's witness) and re-closed.
  • Not fixed, and named in REVERSE_MANIFEST_EXCLUDED: the second lap's warnings, which lose the attestation and the resolution provenance reverse discards. On eight examples that provenance (authority, placement_witness) is why the second lap says a licensing.csv source is unused.

2026-11-04 — RM362 (amended): a fact row at the resolved spelling of an identity the module carries is not an orphan (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A position is every spelling the module's own resolution.csv gives a carried identity. A row authored at one spelling of an indel (CPIC's right-shifted NUDT15 allele) and the table placing its rsID at another (Ensembl's left-aligned one) are one place, and the frequency pass asks gnomAD at the table's. module_subjects now joins every usable locus the injected table files under a row's variant_key or rsid, on the module's build, so frequencies.csv, clinical_assertions.csv and studies.csv rows there are no longer orphans. Only the injected table counts, only where the fill reads it (GRCh38, resolution on), and a not_found row places nothing.
  • BA1 and the study facts reach the same spelling. clin_sig_contradicts_frequency now fires for a pathogenic variant whose common frequency sits at the table's spelling, and a coordinate-only study there feeds the variant's weights.parquet study columns (a variant strict refuses for the contradiction). A row the table placed keeps its own locus only, and one the table places at several other positions keeps its own only, so no row is handed a locus it may not mean. BA1 reports a variant once per ALT when both spellings carry a frequency.
  • New in just_dna_compiler.compiler.subjects: identity_placements, row_positions, position_key, ModuleSubjects.resolved_positions; resolution.usable_loci is public. No reference example gains or loses a warning.

2026-11-04 — RM368: a module publish sends exactly the bytes it checked, and every dry run runs the real run's checks (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — what the publisher plans, checks and sends is now one reading of the disk.

  • The plan's bytes are what is sent. UploadPlan.sha256 records every planned file's hash, manifest.json from the bytes it was parsed from. After the versioned-path check each file is read once more, compared, and those bytes are committed through create_commit; a file that changed in between (a recompile still writing) raises PlanChangedError before either commit. It used to be re-read by upload_folder after the check.
  • A malformed declared field refuses. artifact.files that is not a list, a sha256 that is not a string, a logs/logo/readme/provenance of the wrong shape used to read as absent and drop the check it feeds; the plan now refuses, quoting the field. Absent or null still withholds. A whitespace-only name names no file.
  • The manifest itself is compared at the versioned path. The same data under a new title replaced the published manifest of a version without --force; the refusal now names the dotted paths that differ.
  • Every dry run rehearses. upload --dry-run reads the published manifest of the version (read-only, anonymous without a token) and refuses as the real run would; every snapshot publish --dry-run runs the orphan check that two of them ran and the rest skipped.
  • One error family, upload.PublishRefused: ModuleNotPublishable and SnapshotNotPublishable (still FileNotFoundErrors within 0.8), PlanChangedError, and the collision and orphan types re-based under it. cpic publish and clinpgx publish print a refusal as one line instead of a traceback. INTEGRATION_0_8 § 1.27. No format or compiler change, and no identity moves.

2026-11-04 — RM267: Ensembl REST is the placement's third witness online, so S117's rs8176719 is withheld against the real witnesses (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher, format (a description) — For an rsID the Ensembl snapshot placed, with an allele that is not a substitution, and that neither the ClinVar snapshot nor ClinVar-GKM holds, an online enrich now asks Ensembl REST where it sits (through the run's resolver and RM268's anchored path) and compares the answer by RM267's ladder. REST places rs8176719 (ABO O1) at 9:133257521 T>TC, the snapshot at 9:133257520 G>GC, so the rsID is withheld with both placements named; before, neither loaded witness held it and it was counted as one witness. The witness is named ensembl on the row (placement_witness), the licensed source sources.csv already carries. REST is Ensembl, so the comparison is self-consistent, not independent: the placement_agreement record names its checked side ensembl once REST answered, so independence reads self_consistent (REST alone) or is withheld (REST beside ClinVar), and an rsID only REST agreed with stays in the one-witness count. A REST request unanswered after its retries is not asked for that rsID and counted in detail; --offline asks nothing. The REST leg of EnsemblResolver.resolve_rsid now retries a 5xx and a 429, and a 429 left after the retries is unreachable rather than "no record". New: verification.ran(reported_by=) and skipped(reported_by=), EnsemblResolver.resolve_rsid_rest (REST's half of resolve_rsid), PlacementCheck.self_consistent/not_asked/reported_by/placed_by, placement.live_witness_subjects and ask_live_witness. No new column, no warning code. INTEGRATION_0_8 § 1.18.

  • An injected resolver without resolve_rsid_rest no longer kills the run. enrich(resolver=) asks for resolve_rsid only (ensembl.RsidResolver); REST's mapping is an optional capability (ensembl.RestWitness), asked where the resolver has it. A host resolver written before 0.8 (the shape just-dna-marketplace's own test fake has) raised AttributeError at the first eligible indel; now REST is not asked for that run, the rsIDs are counted (PlacementCheck.not_asked, with not_asked_reason naming the missing method in detail), and one warning says why. The class was hunted across enrich's other injected collaborators: gnomad_client.resolve_rsids, grch37_client.reference_bases/variants_at and the release_probes callables are each their parameter's original contract, and non_nucleotide_rsids was already read by getattr (RM272).
  • resolve_rsid no longer asks the retired beta GraphQL surface before REST. Probed 2026-11-04: beta.ensembl.org/api/graphql/variation answers 301 to www.ensembl.org, where the query no longer validates and a bare rsID is VARIANT_ID_NOT_FOUND; it never answered one, so it cost a request per rsID. The leg is removed with EnsemblSettings.graphql_endpoint and human_genome_id; resolve_rsid is REST alone and every row it writes says ensembl-rest. ensembl-graphql stays a readable source value (rows written before 0.8) and keeps its ensembl authority. The other clients' endpoints were probed the same day, redirects unfollowed: GenCC's bulk download answered 301 (search.thegencc.org → thegencc.org) and gene_validity.DEFAULT_GENCC_URL now names the new host; every other one answers where it is named (ClinPGx's 303 and ClinVar-GKM's directory 302 are hops to files that answer 200).
  • The REST rung is paced and honours Retry-After. EnsemblResolver holds a PacingGate (EnsemblSettings.interval, 0.1 s against Ensembl's published 15 req/s; injectable as gate=) and waits on it once per attempt, so a host's shared resolver is really paced (just-dna-marketplace's pacing table already said it was). A retried 5xx or 429 waits at least the answer's Retry-After (net.wait_retry_after, net.retry_after_seconds: delta-seconds or an HTTP-date); one past net.RETRY_AFTER_CEILING (10 s) is not retried (net.retry_if_transient_within), and the resolver asks nothing, answering could-not-ask, until the time the server named.

2026-11-04 — RM361: a failed, retired, unread or undecided answer is never recorded as a definite one (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A 410 is retired, never unreachable. Every *Retired error carries net.EndpointRetired, and every handler that records why a question went unanswered reads it with net.could_not_ask_reason. A client whose contract is a withheld value (Crossref, the PMC id converter, BioC, GRCh37 REST, Ensembl, the ClinGen Allele Registry) mixes in net.Withholding and says it through retired without raising. retired is a new member of VALID_VERIFICATION_SKIPS and of VALID_FULLTEXT_STATUS (asked again on the next run); LitVar's coverage reasons gain index_retired.
  • "No copy" and "no record" are one named status per client. A fulltext route's 404 and BioC's recorded "no result" body are absent; a 429, any other 4xx, a 200 that is not JSON or not a JATS <article> are unreachable. Ensembl REST's and GRCh37 REST's 400 are their "nothing there"; every other status is could-not-ask. The Hub's own not-found errors are SnapshotNotPublished; a timeout or a refused token is the lane's failure, and cache pull counts it. Those routes retry a 5xx and a 429.
  • A body of another shape is could-not-ask: an esummary with no summary block (it made every uid "no such document"), an OLS4 or HGNC body without its answer block, a Registry body that is not an allele.
  • Open Targets answers only from datasets it read whole (opentargets.read_is_complete, READS): a partial or unstated snapshot replaces no row, names unmatched keys not_read rather than not_in_release, and leaves the posterior as it was; strict refuses either.
  • gnomAD's per-alias errors are unanswered, not not_found (GnomadClient.unanswered, FrequencyResult.unanswered); a CIViC read short of its stated total withholds "rejected or withdrawn"; an unparsable studies.csv is no longer nothing_to_check; an Atlas UNIMPLEMENTED on an SNV is AtlasRetired; a CPIC table with no stated count refuses; a reference read that fails is unread, and all failing is unreachable; a dbSNP outage is a skip of rsid_currency, not ran(findings=0).
  • An undecided provenance_regex withholds quotes_found instead of counting a quote checked and missing; a Europe PMC flag the record does not state is None, not closed.
  • An absent cell is never formatted into an identifier. just_dna_format.base.coordinate_key writes an absent ref as an empty segment (13:48037749, 13:48037749::A), and a release a ClinGen or GenCC file did not state is withheld from the dataset label (net.release_dataset) instead of spelled unknown. The nine :None keys in reference_examples/cpic_gene_pairs/resolution.csv are corrected, which moves that example's haplotypes.parquet and its derived hash (content_signature unchanged): a corrected derivation the 0.8 cut's release record declares.

2026-11-04 — RM362: the orphan checks, stats.genes and the enricher agree on which genes and loci a module is about (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • One answer, just_dna_compiler.compiler.subjects.module_subjects, over two rosters derived from the table registry: MODULE_GENE_TABLES (variants.csv and every gene-bearing table kind) and MODULE_LOCUS_TABLES (variants.csv and every positional table kind). model_organisms.csv is on neither.
  • The five derived-row orphan checks (gene_metrics.csv, gene_validity.csv, frequencies.csv, clinical_assertions.csv, gwas_effects.csv) read it instead of variants.csv alone. A diplotype-scored APOE module no longer warns that its own gene_metrics.csv row names "a gene this module never mentions", and a module without variants.csv is now checked rather than skipped. Each keeps its join and its sentence.
  • study_variant_orphan reads the locus roster, so a study on a heteroplasmy.csv locus is carried; its tail names the roster's tables. The meta arm is gone: a member reaches rows already on the roster.
  • The overlay's resolution.csv reachability counts a positioned row of a positional table, which reverse writes back: a correction aimed at one is no longer "no artifact of this module can carry" on lap 1 only.
  • manifest.stats.genes and the enricher's module_genes call the same function; neither answer moves.
  • Measured on every reference example: no warning gained or lost, stats.genes unchanged.

2026-11-04 — RM371: the enricher no longer warns at import under tenacity 9.2.1, and floors it (patch, on main)

enricher — Importing any retried enricher client under tenacity 9.2.1 printed DeprecationWarning: The 'initial' parameter is deprecated, use 'multiplier' instead, once per @retry site. All 18 sites now pass multiplier=, with the same backoff timing. The dependency floor moves to tenacity>=9.2.1, the first release that has multiplier. An install that also caps tenacity below 9.2.1 now fails to resolve rather than importing. just-dna-lite, just-module-creator and the marketplace all accept 9.2.1, and dagster does not depend on tenacity. No format or compiler change, and no identity moves.

2026-11-04 — RM370: a pass that answered nothing leaves no table behind (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • An enrichment pass with no subject to ask about writes no sidecar, and so no licence row. Seven writers left a header-only table instead: enrich's clin_sig_concordance.csv and clin_sig_authority_calls.csv on a module with no clin_sig cell; frequencies.csv and clinical_assertions.csv when every resolved row is off-build; gene_metrics.csv (both the constraint and dosage passes) and gene_validity.csv on a module with no gene; and literature.csv when every citation is DOI-only. One predicate, layout.nothing_to_write, now sits in front of every enricher sidecar writer.
  • A table already on disk is still the pass's to empty. The concordance record still clears a subject the authorities stopped contesting, and a merge pass still prunes rows whose variant was deleted; only the creation of a file that says nothing is withheld. Two empty tables are still created on purpose: resolution.csv when enrich placed nothing (every later pass needs it to run), and a concordance record the module's overrides.csv answers (so the answer still compiles to overlay_answer_vindicated).
  • cpic_gene_pairs, longevity_studies and model_organisms lose their header-only concordance pair, so their compiled artifacts carry no empty clin_sig_concordance.parquet / clin_sig_authority_calls.parquet. The manifest already reported no concordance block for them.
  • Guarded by a test running every DERIVED_PRODUCERS pass on a module that gives it nothing to answer (runner set equal to the registry), and an AST walk asserting every enricher CSV writer calls the predicate.

2026-11-04 — RM369: a snapshot build that dies before its last write leaves no release.json, never the previous build's (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • Every builder withdraws release.json before its payload moves and writes the new one last. just_dna_format.snapshot.begin_snapshot_write(dir) removes the record and returns a SnapshotWrite whose commit(release) writes it; all sixteen builders and opentargets subset go through it, opened immediately before the first payload byte moves, so a fetch or parse that fails earlier leaves the old snapshot whole and described. Before, a crash between the payload swap and the record left new bytes described by the previous build, and every reader (the cache status label, SourceRow.dataset, a child's parent pin) believed it. Now the worst case reads as an unstated release and is withheld.
  • ClinVar's citations half withdraws its own block only (begin_block_write(dir, "citations")), so the records' provenance is never touched; clinvar_build._merge_release_block is gone.
  • AlphaGenome AVI and CIViC write LICENSE.txt before release.json, the order SNAPSHOT_ROOT_FILENAMES states; both wrote the record first.
  • SnapshotRelease types the four keys readers act on: parts_read and parts_total (dict[str, int], Open Targets), source_release and vrs_version (str, ClinVar-GKM). A value of the wrong kind in an old release.json now costs that key (read as unstated, with the existing warning) instead of raising in the Open Targets pass or being stringified into a VRS comparison.
  • Guards: an AST walk asserting the builders opening a write equal the builders and none names write_release; a reader walk over all three tiers refusing an untyped release read; and every builder run offline with its describe step failing.

2026-11-04 — RM367: validate and compile reach one verdict, and a refused compile writes nothing (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • One resolution step for both commands. validate_spec now runs the compile's own resolution step (compiler/resolution_step.py), so validate --strict refuses what compile --strict refuses on a GRCh37 rsID-only module, under --no-resolve, for locus_cannot_host_genotype, rsid_coordinate_disagrees and a post-resolution duplicate, and the pre-flight emits every warning code the compile does (expansion, labels, ploidy on a resolved MT/Y row, the derived-table orphans, BA1, source_row_unused, log_file_not_shipped). validate_spec takes ba1_threshold; the validate command takes --resolve/--no-resolve and --ba1-threshold.
  • --no-resolve reads nothing it says it skipped. The table is no longer overlaid, VRS-verified (a tampered id refused), counted into vrs_alleles_identified, or used for allele membership; the artifact equals the one the spec compiles with the table deleted.
  • No refusal after the first write. A malformed provenance.json, logo or readme, and every derived-table refusal, now refuse before output_dir is created (a non-UTF-8 provenance.json raised). compiler/check_paths.py rosters the checks one path asks; test_check_paths.py holds the roster, the refusal placement and the flag forwarding as equalities over the code, and test_one_verdict.py holds verdicts and codes over the reference examples and a bad fixture per instance. No new warning code. Consumer note in INTEGRATION_0_8.md § 1.23; the instance list is in ROADMAP_HISTORY.md under RM367.

2026-11-04 — RM363: a gene, drug, DOI, PMID and contig compare by one rule in every tier (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

The 0.8 audit's class B2 (@one-normalizer-two-spellings). What now works:

  • An Apoe diplotype is reached by its APOE allele in meta_resolved.parquet, and diplotypes[gene=apoe] matches APOE rows; the phase checks, the derived-row orphan checks and manifest.stats.genes treat a gene written in two cases as one gene.
  • chr12 on studies.csv, haplotypes.csv, pharm_variants.csv and heteroplasmy.csv (and the four derived tables) is stored as 12, the way variants.csv has stored it since 0.5, so a coordinate study no longer false-orphans against its coordinate variant. Nothing that validated stops validating.
  • Azathioprine and azathioprine are one (sex, drug) scope in meta_resolved.parquet, and a new warning, meta_member_drug_disagrees, says when a drug-scoped meta row's member reaches rows about another drug.
  • duplicate_study_citation treats [PMID: N] and N, and a DOI with or without trailing punctuation, as one citation (spec.citation_key); the literature pass keys a DOI-only row by the same token as its PubMed path.
  • The enricher compares an authored gene or drug with PharmVar, ClinPGx, the drug labels, ClinGen, GenCC and HGNC by the same keys, and the drafters' --gene/--drug filters do too.

The mechanism is just_dna_format.spelling (COMPARISON_KEYS, comparison_key, ChromCell), with three guards: a model walk asserting the set of gene/drug/DOI/contig fields equals the registry and every chrom field folds; an AST walk over the compiler and the enricher asserting no site compares one of them raw (six sites comparing one source's own records are listed with their reason); and an AST walk refusing a hand-typed .lower()/.casefold() on one, and DOI_PATTERN outside the module owning the DOI rule. Gene and drug cells stay as written. The duplicate-row refusal folds only on a table new in 0.8 (meta.csv, model_organisms.csv), where no published module holds a row it would refuse. Consumer view: INTEGRATION_0_8 § 1.22.

2026-11-04 — RM366: a column that means nothing without its companion is refused half-stated (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • Refused at load, on columns new in 0.8: gwas_effects.csv posterior_probability and fine_mapping_method (both or neither); colocalisations.csv h4 or clpp without method; resolution.csv placement_witness and placement_verdict (both or neither).
  • One mechanism. Each model declares _PAIRED_FIELDS (base.Companion: a field and the columns one of which it needs), and one validator on the base.PairedFields mixin enforces every declaration. paired_fields(model) reads them, the element-rule pair included. The eight pair validators written by hand (quality_from/min_quality, confidence→confidence_unit on two tables, sample_size→sample_unit, stratum/stratum_level, the copy-number modifier, source_element, trait_role) are now declarations: the same rules, one message shape — <field>=<value> is set and <companion> is empty — <why>, each rule keeping its own explanation, so a consumer that matched an old full sentence should match the explanation.
  • A new warning code, paired_field_half_stated, for a pair a published module may already hold: effect_size with no effect_measure on variants.csv or studies.csv. Reported once per table and column, in both modes, never refused.
  • The opentargets pass withholds a posterior or a colocalisation statistic whose method the release left empty, rather than stopping on the refusal, and counts the posterior as its own reason (PosteriorCounts.method_missing); 26.09 has none.

2026-11-04 — RM365: the registry guards that were floors or subsets are equalities, and the registries they guard name what they describe (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • DERIVED_PRODUCERS names every check its passes write. resolution.csv's enrich producer gains placement_agreement, published_refutation and evidence_status_currency; literature's gains provenance_quote. Each command's claimed checks now equal what its modules pass to ran/skipped, read by one AST walker the vocabulary-wide emitter guard shares.
  • The first-seen guards measure. The model registry is held equal to every pydantic model the format package defines (less the documents the tooling writes and two named bases) instead of >= 31 models; _RELEASES equals the release tags plus the uncut version (it lacked 0.7.1–0.7.3); and every field's since() must equal the first tag whose model carries it, measured from the tags' own sources. All 538 declarations hold, ResolutionRow.vrs_spec's 0.5.0 included.
  • COMPILER.md's ARTIFACT_PARQUETS diagram drew 26 of the 30 names, partly out of order; it is now the tuple, compared in order by test_counted_prose.py, and the _INPUT_FILES bullet (which lacked meta.csv and model_organisms.csv) is compared as a set.
  • Every *_FACT_FIELDS tuple partitions its model: a column is in the fact set or named, with its reason, in the comment above the tuple. Fifteen columns sat outside their signatures with no word about them; each now has its documented reason, and literature.csv's doi_exists/doi_checked say that none was ever recorded. No tuple changed and no signature moves.
  • Three more guards narrower than their claim: the sidecar-path roster (seven derived tables behind) is derived from the producers, the validate/compile parity cases equal the injected tables, and the vocabulary-closedness walk measures the shared vocabularies it skipped.

2026-11-04 — RM28: a meta.csv row can be scoped to a drug, so one phenotype pair states each thiopurine's recommendation (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • meta.csv gains an optional drug (MetaRow.drug, free text as on diplotypes.csv, first in 0.8.0), and the key is now (members_key, trait_efo_id, sex, drug). An azathioprine row and a mercaptopurine row for one TPMT × NUDT15 phenotype pair are two statements; the same drug twice is still a duplicate key.
  • A drug-scoped row outranks the same key without a drug, for that drug, as a sex-scoped row does for its sex. A row scoped on both outranks either; a male row and an azathioprine row are incomparable where both apply, so a column they disagree on is withheld.
  • meta_resolved.parquet gains a drug column, and its grain is (row reached, sex, drug). The scopes of one row are every combination of the sexes and drugs the rows reaching it state, none included on each axis, so a consumer taking, per axis, its own value where the target has rows for it and null otherwise finds at most one row, and none means nothing applies. A module stating no drug writes the same rows as before, with drug null. A combination's replaces holds within its own sex and drug scope.
  • meta.parquet carries the authored cell, reverse writes it back, and the round trip is a fixed point.
  • reference_examples/cpic_gene_pairs/ carries all three thiopurines: 27 rows, CPIC's nine TPMT × NUDT15 phenotype pairs for azathioprine, mercaptopurine and thioguanine, from the CPIC API call its README records (re-run 2026-11-04T17:15:31Z; the azathioprine texts were unchanged). Re-enriched and closed.

2026-11-04 — RM270: an indel's spelling-independent vrs_id reaches every variant parquet, and a filled coordinate is written left-aligned (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • weights.parquet gains vrs_id (List(Utf8), parallel to alts) and vrs_spec; the three positional parquets (pharm_variants, haplotypes, heteroplasmy) gain the same pair as the comma-joined cell resolution.csv uses, and studies.parquet's proxy_loci struct gains both fields. Both spellings of rs72613567 (4:87310240 T>TA, 4:87310241 A>AA) now carry one id on the parquet and rs8176719's two placements carry two, so a consumer that mints VRS from its VCF has a join key that does not depend on the spelling. Each row's ids are copied from its own resolution.csv row by allele, never minted by the compiler (compiler/vrs_columns.py); an allele its key records with two ids is withheld.
  • reverse writes vrs_id/vrs_spec back into resolution.csv, so compile → reverse → compile keeps every id and the second lap's vrs_alleles_identified (RM241's residual).
  • Every module carrying variants.csv or a positional table gets a new artifact.digest on a 0.8 recompile; content_signature does not move. INTEGRATION_0_8 § 1.19, COMPILER.md § The VRS ids on the variant parquets. Digest parity with the deprecated DuckDB path now holds when the table records no id; with ids, the two weights.parquet differ in those two columns only.
  • enrich writes a coordinate it fills from an rsID left-aligned, ALT by ALT (sequences.normalize_locus, through the same _left_align RM267 compares with), so a caller's normalized VCF meets it on an exact join; a multi-allelic locus whose ALTs land on two references becomes two loci. An authored coordinate is judged and never rewritten: a warning names its leftmost spelling (EnrichmentResult.spelling). Every written row carries the verdict in a new resolution.csv column, left_aligned (tri-state, outside resolution_signature). A recorded row is respelled only under --rederive.
  • reference_examples/shox_par1 re-derived: rs1569493663 is now X:634689 CAG>C and X:634689 C>CAG, and compiler/tests/test_consumer_join.py passes with its RM270 pin removed — every row ClinVar carries joins.

2026-11-04 — PT2 §6: three meta-table reference examples compile, round-trip and state what they resolve (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • reference_examples/apoe_klvs_meta/: the legacy tag selector beaten by a locus row, APOE ε2 and ε4 as allele_function.csv items whose facts differ, the ε2/ε4 diplotype settling the confidence column and leaving the unstated flags withheld, KL-VS × ε4 as one combination that replaces the four rows it names, and APOE's study links on rs429358/rs7412 drawing no orphan warning.
  • reference_examples/pleiotropy_tradeoff/: per-trait rows on bare locus members; tradeoff true on SH2B3, null on protective-only LPA, null with meta_member_allele_ambiguous on CETP rs5882 (on purpose).
  • reference_examples/cpic_gene_pairs/: TPMT and NUDT15 drafted from the CPIC snapshot, and CPIC's nine azathioprine phenotype-pair recommendations as two-selector rows. Its README records three findings (the snapshot cannot reassemble a pair, meta.csv has no drug scope so one module holds one thiopurine, a pair row has no home for CPIC's classification) and one compiler defect it removed a *1 row to avoid.
  • compiler/tests/test_reference_examples_meta.py asserts each README's stated values off the compiled parquets, computed from the example's own CSVs.

2026-11-04 — PT2 §6: three reference examples carry the 0.8 study, HLA and model-organism shapes (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • reference_examples/longevity_studies/: five longevity loci and thirteen studies.csv rows from Timmers 2019 and 2020, Deelen 2019, Pilling 2017, a Research Square preprint (pmid=NONE + doi) and CHAMP. They use every 0.8 StudyRow column: phenotype_class, sample_size/sample_unit/n_cases, the age band, the stratum, ancestry, effect_unit and proxy_rsid. artifact.digest and content_signature survive the round trip; resolution_signature does not until RM331's fix lands, because reverse drops the proxy's resolution.csv row.
  • reference_examples/hla_alleles/: three HLA alleles as allele_function.csv items, with meta.csv findings scoped by sex and one cited by DOI alone.
  • reference_examples/model_organisms/: RM335's table beside a one-variant lead table, with one experiment in two sexes as two keys.
  • Each README names what the build found in the sources and the shape. docs/REFERENCE_EXAMPLES.md §12–§14.

2026-11-04 — RM331: a study measured on an LD proxy gets r² with its proxy, in the study's own superpopulation (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • just-dna-enricher ld-pairs <spec> asks Ensembl REST's pairwise LD for r² and |D′| between each studies.csv row's proxy_rsid and its lead, in the 1000 Genomes phase 3 superpopulation the row's one ancestry code names, and writes a new derived table, ld_pairs.csv, keyed (proxy_rsid, variant_key, ld_population). Online only; --offline asks and writes nothing.
  • Withheld is a status, never a zero. Ensembl's 200 [] is not_reported (below its floor or not in LD), its 400 for an unknown id is rsid_unknown, a study naming several superpopulations or none gets its row as ancestry_mixed / ancestry_unstated with no population asked, and a coordinate-authored lead with no rsID is lead_has_no_rsid. Only resolved carries a number. A merged proxy's current id is kept in answered_rsid; the release (ensembl_116) is read in band.
  • Compiled and reversed: ld_pairs.parquet joins ARTIFACT_PARQUETS after consequences.parquet (a module without the table keeps its digest), manifest.ld_pairs summarises it, reverse writes the CSV back. VALID_SOURCE_LAYERS gains ld_pair, and the (ensembl, ld_pair) licence row commits inside the table's write. No new warning code. INTEGRATION_0_8 § 1.2.4; the hand-run is probes/LANES_0_8.md § 6.
  • A proxy's resolved row now survives compile → reverse → compile. It reached no parquet, so the reversed spec lost it, resolution_signature moved on lap 2 and the proxy allele check went silent. studies.parquet gains the compiler-filled proxy_loci (so every module with studies.csv gets a new artifact.digest on a 0.8 recompile; content_signature does not move), and reverse re-emits the rows. A test holds the tables the enricher resolves equal to the tables reverse rebuilds from; the proxy was the only gap.

2026-11-04 — RM267: an rsID two authorities place as two different events is withheld, and every written placement says who else was asked (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher, format — enrich now compares every placement its links give an rsID with the two offline witnesses this tier holds, the ClinVar snapshot and ClinVar-GKM (new clinvar_gkm_cache= / --clinvar-gkm-cache, read locally, never downloaded by enrich). Two spellings are one event when they are identical, when two substitutions trim to one spelling, when they mint one VRS id (exact, no tolerance), or when they left-align to one spelling against the reference; otherwise the comparison is undecided and counted. A placement two authorities give as two events is withheld: that locus gets no row (an undisputed locus of the same rsID still does), an rsID left with none stays unresolved (so strict refuses, naming the disagreement), a pair row keeps its authored coordinate, an overrides.csv answer on the subject clears the unresolved state, and one warning names both placements. Nothing is rewritten (PROPOSAL_0_8_PT2 §4 question 3). S117's rs8176719 (ABO O1) is the motivating case: the Ensembl snapshot's 9:133257520 G>GC against the 9:133257521 T>TC dbSNP and Ensembl REST give, two haplotypes that mint two ids. Probed 2026-11-04 on real data: neither loaded witness holds rs8176719 (ClinVar's 2026-06-27 VCF has no record at 9:133257510–529 and no RS=8176719; ClinVar-GKM lacks it, RM305), so against today's witnesses S117's own case is counted in the one-witness class and named in detail, not withheld. The spelling 9:133257521 T>TC is dbSNP's and Ensembl REST's; the tests supply it as a ClinVar row to exercise the withhold. rs72613567 and rs77944059 are respellings that mint one id and are written as before. A different allele of the rsID is not a disagreement (unplaced), and a ClinVar fill has no independent witness here. resolution.csv gains two optional provenance columns outside resolution_signature, placement_witness (who was asked) and placement_verdict (VALID_PLACEMENT_VERDICTS: same_event, undecided, unplaced; same_event only when every ALT was placed the same), and each witness a written row names gets its SourceRow at the resolution layer, which the compiler's orphan check now counts as used. The comparison attests as the new check placement_agreement (witness clinvar+clinvar_gkm, with an undecided count and the one-witness count in detail); EnrichmentResult.placement carries it. A 0.7 compiler refuses a 0.8-enriched resolution.csv (extra="forbid"): upgrade first. A table written before 0.8 is compared only under --rederive or after deleting it, and under --rederive a disputed subject's recorded placement is not carried forward. No parquet and no warning code. INTEGRATION_0_8 § 1.18.

2026-11-04 — RM28: a study on a pharm_variants.csv variant is not an orphan, and a module with no variants.csv is checked (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • pharm_variants.csv is an arm of the study-orphan check of its own (the lead's decision, reversible until the cut). A study on a pharm_variants.csv row's variant is carried whether or not a meta.csv member reaches it; before, only a member's reach counted.
  • The check no longer waits for a variants.csv. A diplotype-only or PGx-only module's studies.csv is checked against the arms it has, so a study on a variant in no table now draws study_variant_orphan, and duplicate_study_citation, computed in the same pass, runs there too (@roster-is-as-wide-as-the-tables-it-reads). Measured on the 16 reference_examples/: none gains or loses a warning.
  • The message keeps its pinned prefix and its tail now names the new arm: — nor among haplotypes.csv's defining variants, pharm_variants.csv's rows or the rows a meta.csv member reaches. (INTEGRATION_0_8 § 1.15).

2026-11-04 — RM359: under --strict an escalated finding is an error only, and the docs state no registry totals (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • One finding, one channel. A compile --strict used to publish every mode-ladder finding twice: once in errors, and once in warnings as the best-effort sentence its pre-flight had run into. It is now in errors only. compile_module runs its pre-flight (validate_spec) in its own mode, so a strict compile refuses with validate --strict's accumulated errors where it used to stop at the first ladder that refused, and a resolution ladder publishes its longer refusal instead of its warning (ResolutionOutcome.channels(strict=)). best_effort output is unchanged. No new warning code. Asserted over every ladder code by compiler/tests/test_one_channel_under_strict.py; consumer note in INTEGRATION_0_8.md § 1.17.
  • No counts in prose beside a generated roster. SCHEMAS.md said authoring_reference()["vocabularies"] had "21 entries today" (47 now). That number and the other registry totals the maintained references stated with nothing re-deriving them are gone; the paragraph states the rule that generates the list. The list is in ROADMAP_HISTORY.md under RM359.

2026-11-04 — RM354: gwas reads the GWAS Catalog's REST API v2, and every client names a retired endpoint (patch, on main)

enricher — The GWAS Catalog retired the legacy REST API the gwas pass read; every legacy path now answers HTTP 410, so gwas_effects.csv could not be filled for any module. The pass now reads /gwas/rest/api/v2/associations?rs_id=, follows its paging, and takes pmid, study accession, trait and EFO id from the association itself; ancestry is one request per distinct study (/studies/<accession>/ancestries). Measured on hfe_hemochromatosis: 190 requests where the legacy API took 382. Same GwasEffectRow columns, and the association ids carried over, so a re-run merges into a table the legacy pass wrote. What changed in the values: standard_error is always empty (v2 publishes none); p_value is the Catalog's mantissa/exponent pair (3e-493 where the legacy pass wrote 0.0); an all-NR ancestry is empty rather than NR; an effect string outside v2's <number> [<unit>] <increase|decrease> is withheld and counted in the new GwasResult.unparsed_effects, which --strict refuses on and the CLI summary prints. The effect allele is taken from the snp_effect_allele entry naming the queried rsID, never the first of a multi-SNP list. The client now retries a 5xx and a 429 as well as a transport failure. --no-study-facts now skips only the ancestry request; pmid, accession and trait are filled either way. A 404 on the association search is now could-not-ask (v2 answers a variant it holds nothing for with an empty 200), so gwas left the 404-is-an-answer set.

Across the tier: a 410 Gone raises each client's own *Retired, a subclass of its could-not-ask type, so existing handlers still catch it: GwasRetired, GnomadRetired, EutilsRetired, CpicRetired, PharmVarRetired, ReleaseRetired, LiteratureRetired, IdentifierRetired, PgsCatalogRetired, LitvarRetired, CivicApiRetired. The message carries the URL and the source's own text. gnomad, eutils, cpic, pharmvar and the ClinVar release probe no longer retry a 400, 404 or 410; they retry transport failures, 5xx and 429 (net.retry_if_transient). What a 404 means for each client is unchanged and is tabled in ENRICHER.md § Exception contract. No format or compiler change and no identity moves. One text-only change in format: the GwasEffectRow column descriptions name v2's fields.

2026-11-04 — RM355: upload --dry-run reports a refusal as the real run does, and the publisher reads its manifest once (patch, on main)

enricher — just-dna-enricher upload --dry-run on a module the publisher refuses (an attested file missing, a declared file edited after the compile, no annotation table) now prints the same UPLOAD FAILED: … line as the real run and exits 1, where it used to end in a traceback. The dry run still contacts nothing, so a versioned-path collision is reported by the real run only. The plan and the publish read manifest.json once (the version used to come from a second read), and an artifact.files entry whose name is "" is refused with the entry quoted, where the refusal used to read attests [''] and suggest the file was missing. No format or compiler change, and no identity moves.

2026-11-04 — RM357: the authored tables that are not table kinds have one registry and one loader, and a refused meta.csv no longer reports its citations as uncited (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • One registry. just_dna_compiler.compiler.AUTHORED_NON_KIND_TABLES registers overrides.csv, meta.csv and model_organisms.csv (CSV name, parquets, row model, its own table checks). Every list that named them is derived from it, among them _INPUT_FILES, ARTIFACT_PARQUETS, _TABLE_DUPE_KEYS, _CITING_TABLE_KINDS, draft.DRAFTABLE, reverse_module and the enricher's gene scope; COMPILER.md § The authored tables that are not table kinds has the full list.
  • One loader, load_authored_table(spec_dir, csv_name), for all three. load_overlay stays and is that loader on overrides.csv; compiler.meta.load_meta and compiler.model_organisms (both new in 0.8) are gone.
  • A refused table keeps its rows for the citing view. A meta.csv with a duplicate key used to drop every row, so a literature.csv row only it cited was also reported as uncited, pointing the author at the wrong file. Now the duplicate is the one finding. A meta.csv or model_organisms.csv with a row that will not parse keeps the rows that did, so only the malformed row's own citation can still read as uncited; a malformed overrides.csv is still applied not at all.
  • On a module with none of those refusals nothing moves: the 16 reference examples and the meta, model-organism and overlay fixtures give the same warnings, errors, digests and artifact.files listing as before.

2026-11-04 — RM356: two genes' *2 defined by one variant are two haplotypes.csv rows (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • haplotypes.csv's duplicate rule reads gene, when both rows state it. The key stays (haplotype_name, variant_key, allele); HaplotypeRow._KEY_DISTINCT_WHEN_STATED = ("gene",) declares that two rows under it are two rows when both state a gene and the genes differ (case- and space-insensitive). The UGT1A genes share exons 2-5, so UGT1A1's and UGT1A6's *2 can share a defining variant and allele, and the second was refused. An unstated gene is unknown, so a gene-less row still collides with any row under its key (@absent-is-not-different), and every refusal keeps its text.
  • One predicate at every "same row" site: compiler.tables.same_row, asked by the compile/validate refusal, draft.append_rows (so a second gene's drafted row is added, never refused at compile) and the hints lint. RM351's _gene_key moved beside it so the two compare a gene one way.
  • hints.TableKey.distinct_when_stated publishes the column beside columns (describe_table too).
  • haplotypes.parquet may now carry two rows sharing (haplotype_name, variant_key, allele), differing by gene; a reader indexing the table on those three adds gene (INTEGRATION_0_8 § 1.16).
  • Additive: it only narrows what is refused. The reference examples carrying haplotypes.csv compile to the same artifact.digest and content_signature with and without the declaration. Tests: compiler/tests/test_haplotype_key_gene.py; one scenario in features/compiler/licensing_closure_pgx.feature.
  • The class survey found no second instance: AlleleFunctionRow and DiplotypeRow key on gene, the CPIC drafter matches (gene, allele), and RM351 had already re-keyed the cross-table checks.

2026-11-04 — RM28: a study on a variant haplotypes.csv defines, or a meta.csv member reaches, is not an orphan (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • study_variant_orphan counts what the module carries, not variants.csv alone (PROPOSAL_0_8_PT2 §5 B1 step 4). A study's variant is carried when it is a variants.csv row's, a defining variant in haplotypes.csv, or the variant of a variants.csv / pharm_variants.csv row a meta.csv member reached. The meta arm reads the rows build_meta already expanded each member to (compiler/meta.py member_variant_rows), so a selector covers exactly the rows it matched.
  • The case it was for: APOE scored only as diplotypes (longevitymap v2's R-APOE-1). Its 14 study links on rs429358 and rs7412, the SNPs that define the ε alleles, drew the warning because neither is a variants.csv row. They no longer do, and a study on a variant in no table still warns.
  • The warning fires less often, never more, and keeps its pinned prefix; the sentence gains a tail naming the wider search (INTEGRATION_0_8 § 1.15). The check still runs only in a module with a variants.csv.
  • That closes RM28's 0.8 part. Its predicate half (open-world negation, OR, trans phase) stays parked in ROADMAP_0_8, out of bounds by the maintainer's 2026-11-04 decision.

2026-11-04 — RM353: the module publisher sends and guards exactly what the manifest declares (patch, on main)

enricher — upload_module (and just-dna-enricher upload) now sends manifest.json plus every file the manifest declares in the module directory: the parquets artifact.files attests, the logo, the readme, provenance.json (named in manifest.provenance and never uploaded until now) and every log. A file on disk that the manifest does not declare is no longer sent, so a parquet left by an earlier compile, or a logo the manifest does not name, stays behind. A declared name that leaves the module directory is skipped with a warning, as RM347 did for logs. A declared file whose bytes no longer match the sha256 the manifest states (edited after the compile) refuses the publish and asks for a recompile. When the manifest is absent, unparseable or has no artifact.files list, the publisher falls back to its old allowlist, as before. The guard on data/<name>/v<version>/ now compares every declared file's sha256 as well as artifact.digest, so republishing a version whose logs, provenance, logo or readme changed is refused without --force; the PublishCollisionError message names the files that differ. Every name passed to the hub is now glob-escaped by one helper: the snapshot publisher's file list (a data/chr[1].parquet selected data/chr1.parquet instead), a declared retirement's deletion, and cache prune's deletion. No format or compiler change, and no identity moves.

2026-11-04 — RM333, RM239, RM236: the Open Targets lane fills a locus-to-gene table, its colocalisations, per-transcript consequences and a fine-mapping posterior (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

  • A new snapshot lane, opentargets. just-dna-enricher opentargets build mirrors Open Targets Platform release 26.09 from EBI's FTP over https (retried, translated to OpenTargetsUnavailable) and flattens five datasets into the lane snapshot; --source builds from parts already on disk, --parts N a test-sized one, and release.json records parts read against parts total. cache status shows it with its release label. The hand-run build is probes/LANES_0_8.md § 5.
  • opentargets enrich <spec> reads the snapshot only (it never fetches) and the module's resolution.csv, and writes three new derived tables, each with its (opentargets, <layer>) licence row inside its own commit: locus_to_gene.csv (one row per study locus, gene and module variant, for every set the variant is a member of, led or not: the L2G score), colocalisations.csv (that study locus against a molecular-QTL study locus: h4, clpp, the right side's gene and tissue) and consequences.csv (one row per variant and transcript: SO accessions and impact). Several study loci per variant are kept, never collapsed. Members, not lead-only, so HFE C282Y, which sits in 82 GWAS sets and leads none, gets their rows. In overrides.csv a correction to either table keys on the pair and reaches every member's row; an update of variant_key is refused as re-keying, an insert matches on the variant_key it supplies, and a suppress removes every member's row. Two module keys naming one release allele have these rows withheld and are named in the pass's shared_allele_keys.
  • A resolution.csv row that states an rsID and no ALT now gets rows (added 2026-11-04, build/RM333-noalt). It was skipped silently, so apoe_epsilon's rs429358 and rs7412 had no L2G, coloc or consequence row. The pass reads every allele the release holds at the row's position: one allele carrying the rsID (or its rsid_current) on the row's REF, beside no other candidate, supplies the ALT for all four tables, and the key is named in allele_supplied; several candidates are withheld and named in allele_unstated_withheld, never one picked. An allele with an empty rsIds counts as a candidate, since the release leaves some alleles unlabelled: at 6:26091590 G it names rs773443949 on G>T only, while ClinVar puts it on G>A too. A row with neither ALT nor rsID is withheld (allele_unstated_no_rsid). The supplied ALT is never written back to resolution.csv, and no derived row says which ALT it used. The snapshot's variants.parquet gains rs_ids for this, so a snapshot built earlier must be rebuilt: against one, those rows are allele_unstated_unasked and --strict refuses.
  • gwas_effects.csv gains posterior_probability and fine_mapping_method (RM239), filled by the same pass when exactly one fine-mapping value answers for a row's variant and study, and left empty and counted when the study has no credible set, the variant is not in it, or several sets disagree. No credible-set level: the release publishes the 95% set only.
  • Each table compiles to its own parquet with a manifest block (locus_to_gene, colocalisations, consequences), is overlay-correctable through overrides.csv, and survives reverse. No causal or effector confidence label is computed (RM323's cut). No new warning code. INTEGRATION_0_8 § 1.2.3.

2026-11-04 — RM305: ClinVar-GKM's indels are a lane, an offline witness for where an indel sits (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — A new cache lane, clinvar_gkm: ClinGen's ClinVar-GKM Parquet set cut to its GRCh38 net insertions and net deletions (309,405 alleles on the 2026-10-24 set), each with its VRS id, SPDI, gnomAD-style VCF spelling, ClinVar variation and rsID. It is a witness, not a check: it runs nothing of its own, and RM292's records and RM267's comparison name it as what they checked against. clinvar-gkm build downloads allele.parquet and variation.parquet whole (the source asks for whole downloads, not range reads) and refuses any file whose size or sha256 differs from the set's manifest.json, the new ClinVarGkmMismatch: the source overwrites its one Parquet set weekly, so the pin is (source_release, sha256), and a release landing mid-build is named as such. release.json records dataset clinvar_gkm_<source_release>, both digests, the line's newest weekly release, and vrs_version 2.1.1 with its basis (declared by the source's changelog; the bytes state no version). ClinVarGkmWitness.lookup takes a VRS id, an rsID or a one-ALT chrom:start:ref:alt key and returns every placement, never a pick; ids_comparable says a VRS-id comparison is sound only at the same VRS minor, so against today's 2.0 minter the VCF spelling is the key. clinvar-gkm currency and lane_currency answer three questions apart: a newer Parquet set exists, the set was overwritten under the same release, or the set itself lags the newest weekly release (which a rebuild cannot fix). clinvar_gkm joins currency.default_probes, so --verify-datasets asks it for any module whose sources.csv names it, and CLINVAR_GKM_TERMS (CC0 over ClinGen's transformation, ClinVar's attribution and disclaimer verbatim) joins TERMS_BY_SOURCE under its own name. CC0 permits publishing, so the lane has ensure_clinvar_gkm_snapshot and a just-dna-seq/clinvar_gkm repo (not created yet), reached through cache rebuild --only clinvar_gkm --publish. No warning code, no model or parquet change.

2026-11-04 — RM326, RM328: weights.parquet says whether a locus's evidence is European-only, and which outcomes were studied (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — weights.parquet gains four derived columns, repeated on each genotype row of a locus and read off its supporting study rows (same rsid or same chrom:start:ref). eur_only_evidence is true when every row states only EUR, false when one states another code, and null when any states none or there are none. The outcome strip outcome_healthspan, outcome_lifespan, outcome_exceptional is true when a row's phenotype_class maps to the cell (vocab.OUTCOME_STRIP_CELLS: HEALTHSPAN; PARENTAL_LIFESPAN and OWN_SURVIVAL; EXTREME_SURVIVAL_CC and FAMILIAL) and null otherwise, never false. None is authored and reverse writes none back. Closes RM326 and RM328.

2026-11-04 — RM326, RM329, RM330: one paper's several outcomes, bands and strata are not duplicate citations (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler — duplicate_study_citation reads the 0.8 study columns by RM140's rule: two rows for one (variant_key, pmid) are two claims when phenotype_class, an age-band cell (age_min, age_max, age_of, each on its own) or the stratum pair is stated on both and differs, as statistical_test already did. An unstated cell is unknown and the pair still warns; proxy_rsid is not read. Every case that still warns keeps the byte-identical message, and a module writing none of the new columns warns exactly as before. Closes RM329 and RM330 (both halves of RM330 are now on 0.8).

2026-11-04 — RM332: an effect states its unit on the study row, and weights.parquet carries it when the studies agree (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains effect_unit, free text and verbatim, the column gwas_effects.csv has carried since 0.6. weights.parquet gains a derived effect_unit per variant: the unit every supporting study row that states one agrees on, null when none states one or two disagree. It is not authored on variants.csv (the maintainer cut that half) and reverse never writes it back. Every module carrying variants.csv gets a new artifact.digest on a 0.8 recompile, since weights.parquet gains a column; content_signature does not move.

2026-11-04 — RM331 (the author's half): a study measured on an LD proxy says so, and its allele is checked against the proxy (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler, enricher — studies.csv gains proxy_rsid, the variant a study measured when it stood in for the row's lead. A proxy on a row naming no variant, or equal to its own rsid, is refused. study_effect_allele_not_at_locus checks effect_allele against the proxy's resolved alleles when one is stated, never the lead's, and names both in its text; a proxy with no resolution.csv row withholds. enrich resolves each stated proxy as a subject (after every table naming a module variant), so the row exists. A proxy is not one of the module's loci, so the GWAS, frequency and clinical-assertion passes leave proxy-only rows out (naming them runs the subject collector, so the frequency and clinical-assertion passes now refuse an invalid PGx or heteroplasmy table, as enrich already did). An invalid study row does not stop enrich: the valid rows' proxies are asked for and one warning names the rest. The r² between proxy and lead, computed in the study's ancestry, is still to come on 0.8.

2026-11-04 — RM330: a study row states the stratum its estimate was measured in (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains stratum and stratum_level, free text and written both or neither (sex/male, parent_sex/father); one without the other is refused. They reach studies.parquet and reverse writes them back. This is RM330's study half; its conclusion half, meta.csv's sex, shipped with the meta-table.

2026-11-04 — RM329: a study row states the age band its estimate covers, and whose age it is (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains age_min and age_max (years, finite and non-negative; either may be empty for an open band) and age_of (closed: participant, parent). No refusal ties the three together. They reach studies.parquet and reverse writes them back. age_event was cut by the maintainer before the build.

2026-11-04 — RM328: a study row names the ancestry it was measured in (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains ancestry, optional, multi-valued and closed over the 1000 Genomes superpopulation codes (vocab.VALID_SUPERPOPULATIONS: AFR, AMR, EAS, EUR, SAS). A trans-ethnic study lists its codes; population stays free text. pgs.VALID_TRAINING_ANCESTRY is now derived from the same constant plus multi, so the five codes are one vocabulary; its members are unchanged. The column reaches studies.parquet as a list and reverse writes it back |-joined.

2026-11-04 — RM326: a study row states which outcome it measured (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains phenotype_class, optional and open, with longevitymap v2's twelve classes as the recommended members (RECOMMENDED_PHENOTYPE_CLASSES: EXTREME_SURVIVAL_CC, PARENTAL_LIFESPAN, OWN_SURVIVAL, VIABILITY_AF, HEALTHSPAN, MULTIVARIATE_AGEING, FAMILIAL, DISEASE_DEPLETION, RARE_BURDEN, EVOLUTIONARY, MR, BIOMARKER). A token outside them is kept and draws the new warning phenotype_class_unrecognized, once per module, naming each token with its row count, in both modes. It reaches studies.parquet and reverse writes it back. New warning code: a reader pinned to 0.7 refuses a manifest carrying it (INTEGRATION_0_8 § 1).

2026-11-04 — RM325: a study row states how many were analysed, in what unit, and how many were cases (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — studies.csv gains sample_size (the total analysed, cases included), sample_unit (closed: individuals, lives) and n_cases, optional and declared after every 0.7 column. A sample_size without its unit is refused, since a parental-lifespan study counts two lives per participant, and so is an n_cases larger than the sample_size beside it. The three reach studies.parquet (the counts as Int64) and reverse writes them back. A module that leaves them out keeps its content_signature; every module with studies.csv gets a new artifact.digest on a 0.8 recompile, because the parquet gains columns.

2026-11-04 — RM349: a pmid=NONE row's DOI is checked to exist, and its quote matched where the text is reachable (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — The literature pass now asks about a work cited by DOI alone. Each distinct DOI on a pmid=NONE row (spec.doi_key) is put to Crossref, and the answer names the title Crossref holds for it, in the citation_existence detail and on the CLI (NONE doi <doi>: names '…'). A DOI no registration agency knows is a finding, joins doi_missing, and is refused under --strict. Its quotes are matched against the text Europe PMC holds for that DOI: a record with a PMC id goes through the Europe PMC and BioC rungs, and a bioRxiv/medRxiv preprint goes through Europe PMC's preprint repository, which serves the JATS the documented fullTextXML route answers 500 for. The abstract is the fallback (a hit settles a quote, a miss does not), and fulltext_status keeps its meaning: retrieved, absent or unreachable. Settled quotes join the provenance_quote denominator. A module citing only by DOI records citation_existence with source="crossref", or skips it as not_requested (--no-doi) or unreachable; offline both records say offline (the quote record used to say unsupported). Nothing is written to literature.csv, whose rows stay keyed on a PubMed id, so every online run asks again. LiteratureResult.doi_citations holds one DoiCitation per DOI.

A Crossref 404 is no longer read as "this DOI does not exist" by itself. /works answers 404 for every DOI another agency registered, such as a DataCite dataset (10.5281/zenodo.3360392), which is real. After a 404 CrossrefClient.work asks /works/{doi}/agency: another agency answers exists=True with CrossrefWork.agency set and no title, a 404 there is False, and anything else withholds. This also changes an authored DOI's check on a PubMed row and hint citation --doi. A literature.csv row pinned before this with doi_exists=false for such a DOI keeps its verdict, since a pinned row is not re-fetched: delete the sidecar to re-derive it.

format — literature.settled_quotes, the rule LiteratureRow.quotes_checked applies, as a function for a citation that has no row.

2026-11-04 — RM292: every verification check names its witness, and a check that reads its own source says so (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

A record in verification.json (and manifest.verification.checks[]) now states what the check read and what it read it against. VerificationRecord gains checked_source, witness_source, undecided and independence (all optional, since 0.8.0); independence is derived from the two sources over the new closed VALID_VERIFICATION_INDEPENDENCE — self_consistent when they are one source, independent when they differ, null when either is unknown. The enricher's roster just_dna_enricher.verification.CHECK_WITNESSES declares both sides for every emitting check, and ran()/skipped() fill every record from it, so a check with no entry cannot be recorded; an equality test over the roster and the vocabulary keeps it whole, and ENRICHER.md's check table gains Reads and Against columns asserted equal to it. record_verification settles authored against the module's licence table: a ClinVar-drafted, unedited clin_sig column now reads clinvar against clinvar, and a licence row naming the witness with no digest withholds the side. rsid_coordinate_agreement publishes the indel pairs its RM31 tolerance abstains on as undecided (@tautology-zero). An old-reader break: a 0.7 reader refuses a 0.8-written record (extra="forbid", 0.7's producer shape again) — INTEGRATION_0_8 § 1.13. No published verification.signature moved.

2026-11-04 — RM335: model-organism evidence has a table of its own, model_organisms.csv (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format — just_dna_format.model_organisms.ModelOrganismRow: one row per model-organism lifespan finding about a human gene. gene (required, the human symbol), orthologue, species, model_type, sex (the animals', free text; not meta.csv's karyotype condition), diet, lifespan_effect, concordance (closed: concordant, discordant, nonmonotone; empty is unstated), pmid, doi. It inherits CitationPair; the citation is optional and an uncited row draws no warning. The key is every column but lifespan_effect and concordance.

compiler — The table is authored input (in manifest.inputs and content_signature), keyed (a duplicate is refused), a citation site (_CITING_TABLE_KINDS), and compiles verbatim to model_organisms.parquet, appended last to ARTIFACT_PARQUETS; reverse writes it back and the round trip is a fixed point. It is registered beside the meta-table rather than as a _TABLE_KINDS entry, so it is not a lead table: a module carrying only this file is refused as having no recognized table, and its genes stay out of manifest.stats.genes. Annotation only: it reaches no weight, effect, tier or meta member. describe/template/stub answer for it (DRAFTABLE), which also puts its gene in front of check-identifiers. A module carrying no such file keeps its digest. INTEGRATION_0_8 § 1.2.2.

enricher — gene_metrics.module_genes, the scope of the constraint, gene-validity and ClinGen dosage passes, leaves model_organisms.csv out (a gene named only there is not one of the module's genes, as manifest.stats.genes already says); authored_identifiers/authored_rows gain a keyword-only exclude for it. check-identifiers reads the table as it reads every authored one.

2026-11-04 — RM351: the PGx cross-table checks know an allele by its gene, and an HLA allele is no longer "undefined" (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — The two cross-table PGx checks key a named allele on (gene, allele), not on the bare name. star_allele_undefined used to let one gene's definition answer for another's: a module defining CYP2C19 *2 and using CYP2C9 *2 with no definition got no warning, and now gets one. The message keeps its wording and appends the gaps grouped per gene ("By gene: CYP2C9 *2; CYP2D6 *36, *37."); genes compare case-blind, so Cyp2c19 and CYP2C19 are one gene. An allele of an HLA gene (pgx.is_hla_gene, the HLA- prefix RM334's grammar reads) that haplotypes.csv defines nothing for is typed by name and defined by no variant, so it is no longer reported as one that "can never match"; an HLA gene the module does define by variants is checked like any other. When the warning fires for another gene it names the HLA genes it did not check, and when the only undefined alleles are HLA alleles, nothing fires. In diplotype_definitions_identical and diplotype_phase_ambiguous, one gene's *2 no longer merges its defining variants into another gene's *2, which had silenced a real finding. A haplotypes.csv row with no gene still defines its name for every gene, as before, so a module that leaves gene out of haplotypes.csv is warned about the same alleles as before; only the appended text is new. The six PGx reference examples report the same gaps as before. No column, code or identity changes.

2026-11-04 — RM327: meta_resolved.parquet says which rows carry a trade-off, computed from the per-trait meta rows (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler — meta_resolved.parquet gains tradeoff, a nullable boolean between the fact columns and withheld. It is computed per allele-bearing member: true when one member has a protective per-trait row on one trait and a risk per-trait row on another (each a one-member meta.csv row with trait_efo_id and direction), and it lands on every row that member reaches, a diplotype reached through an allele_function allele included. Two members are never pooled, so two alleles meeting at a heterozygous diplotype are not a trade-off. A mediator row counts; neutral, unknown, contested and a direction the compile withheld do not. On one trait at one row the fact columns' precedence runs first, so a narrower judgement on that trait (a sex-scoped row, a diplotype over the allele it contains, a genotype over its locus) drops the broader one there. It is true or null, never false: RM327's false arm needs every trait's judgement stated, and a module states no closed list of traits.

format — tradeoff leaves vocab.RESERVED_NAMES_0_4 (three names now) and joins vocab.MISPLACED_COLUMN_REASONS, since it is real on a generated table and still never authored. A tradeoff column on any authored table is refused, as before, with a reason that now says the compiler computes it. authoring_reference() gains misplaced_columns, that map, beside reserved_names.

What changes for you. A module with per-trait meta rows recompiles with one more column on meta_resolved.parquet, so its artifact.digest moves and its content_signature does not. A module without meta.csv is untouched. A tool listing unwritable names from RESERVED_NAMES_0_4 alone no longer sees tradeoff; read MISPLACED_COLUMN_REASONS too.

2026-11-04 — RM352: a release record can declare a toolchain-only byte move, backed by a fixed-compiler sweep (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — VALID_RELEASE_CHANGE_KINDS gains toolchain: the bytes moved, the content did not, and the recorded toolchain (compilation.toolchain, RM342) says why. It is legal on parquet_bytes only, and RecompileAnswer.toolchain_moves reads it beside corrections and additions. The release gate accepts it only when a second sweep backs it, with this release's compiler on both sides and only the toolchain varied (sweep --toolchain-before DIR). Both sides must record different toolchains, the sweep must share the release's AFTER tree, and content_signature must not move. A module's move counts as the toolchain's only where the previous compiler and this one wrote the same bytes under the previous toolchain, so the cut builds its BEFORE tree under the previous polars too. When toolchain is the axis's only kind every moved module must pass that test. Each condition has its own TOOLCHAIN_* finding phrase. SweepMeasurement records each side's toolchain and digests (before_toolchain / after_toolchain, before_digests / after_digests). Measured on this branch: the sixteen reference examples compiled by one compiler under polars 1.44.2 and 2.1.0 move parquet_bytes on 16 of 16 and content_signature on none. The Constitution's Goals sentence on integrity now names the recorded toolchain beside the pinned compiler version. A 0.7 just-dna-format refuses a serialized record carrying kind: toolchain; see INTEGRATION_0_8.md § 1.11.

2026-11-04 — RM347: the HuggingFace publisher uploads the logs a module's manifest names (patch, on main)

enricher — upload_module (and just-dna-enricher upload) now sends every path manifest.logs names, so a published module's manifest no longer names logs the repository lacks. The list is read from the manifest the directory carries, never from a glob, and UploadPlan.files lists the logs as paths relative to the module (logs/authoring.log), so --dry-run shows them. A file under logs/ that the manifest does not name is not uploaded. A named path that is absolute, uses .., resolves outside the module directory, is hidden and not a *.log (the compiler's own discovery rule), or is one upload_folder's default ignore list drops (.git, .cache/huggingface) is skipped with a warning, and the publish goes on. UploadPlan.files entries were basenames until now; a caller that joins or compares them as basenames should expect a / in a log's entry. An absent or unreadable manifest behaves as before. The upload's allowlist is now the plan itself, glob-escaped, so the dry run is exactly what is sent. No format or compiler change, and no identity moves.

2026-11-04 — RM348: logs/ discovery stops trusting the filesystem (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler — Every path log discovery considers, the top-level *.log included, must resolve to exactly where it lies inside the spec directory. A symlink at any level stays behind with log_file_not_shipped's symlink reason, now "it is a symlink or lies under one, and a link can reach outside the module", and a logs/ that is itself a link is not walked: one finding names logs. Every discovered file is held to one ceiling, LOG_MAX_BYTES (5 MiB, read from stat before any byte; LOG_IMAGE_MAX_BYTES, RM338's name on main, stays as an alias of it), and the oversize reason now reads "it is larger than the ceiling for a file discovery ships". A file the compile cannot stat or open stays behind with a new reason, "the compile could not read it", where a PermissionError used to fail the compile; a directory under logs/, or logs/ itself, that the compile cannot list and search draws the same reason in place of the files it hides, which the walk used to skip without a word. This narrows what a *.log did before RM338: a symlinked *.log, one beside the spec or under logs/, and a *.log over 5 MiB used to ship and now stay behind with a warning. A *.log is still not checked for content. An explicit log_files list is unaffected, and neither identity moves.

2026-11-04 — RM28, RM323, RM324 (and RM327's rows, RM330's sex): the meta-table, meta.csv (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler, enricher — A new optional authored table, meta.csv (just_dna_format.meta.MetaRow), maps a member set to a finding and/or the per-locus facts. A member is table[column=value, …] over variants, diplotypes, pharm_variants or allele_function, members joined by &; conditions are AND-ed and = on a list column means contains. Columns: members (required), trait_efo_id, trait_role (mediator), direction, conclusion, negatives, tier, age_of_action, age_of_action_confidence, flags, pmid, doi, sex (male/female). The key is the sorted members, the trait and sex, so A & B and B & A are one key. module_spec.yaml gains tier_scale, required by any tier cell, copied into the manifest and written back by reverse. The compiler expands every member and writes meta.parquet (the authored rows, verbatim selectors, their matches, replaces, withheld) and meta_resolved.parquet (each row a one-member meta row reaches, with each fact after precedence and the columns withheld where two incomparable rows disagree). New manifest field meta_selectors (each member's match count); new warning codes meta_selector_matches_nothing and meta_member_allele_ambiguous. meta.csv alone is not a module. The enricher's rsID currency pass asks dbSNP about every rsID a member names. See INTEGRATION_0_8.md and COMPILER.md § The meta-table.

2026-11-04 — RM303: the snapshot layout and release.json's model move into just-dna-format (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, enricher — New module just_dna_format.snapshot: the reference-snapshot layout names (data/, the sidecar and root-filename tuples, LICENSE.txt, release.json, the two named payload files), SnapshotRelease, the model of release.json, and read_release / write_release / release_label / release_is_unreadable / holds_snapshot_payload. A Goals amendment records that the format package also defines this layout. just_dna_enricher.locations imports the names, so imports from there keep working, but read_release now returns a SnapshotRelease or None, not a dict. Every builder writes release.json through write_release, and an AST walk refuses a hand-built one. Keys are written in one spelling: license (CIViC wrote licence), row_count for a total and row_counts for a per-file map (three builders wrote rows for one or the other); the old spellings of published snapshots are read and never written. ClinVar's builder now writes dataset, and a ClinVar snapshot built before reads its label from clinvar_file_date through the same formula, so CacheLane.release_label's ClinVar override is gone and cache status prints the same labels. See INTEGRATION_0_8 § 1.9.

2026-11-04 — RM253: the CPIC drafter defines UGT1A1 *28, spells CPIC's DEL/INS/repeat notation, and drafts HGVS-named genes (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — draft --gene UGT1A1 now defines *28, *36 and *37 (and *80+*28, *80+*37), so validate no longer warns that they are used but not defined. CPIC's DEL<bases>, INS<bases> and repeat-count UNIT(n) values are spelled as the VCF-anchored, left-normalized indel a caller writes, reading the anchor and the run from the GRCh38 reference (*28 is 2:233760233 C>CAT, *36 CAT>C, *37 C>CATAT; the reference count is written C>C there). A repeat row is keyed by its coordinate and carries no rsID, since rs3064744 is in no caller's ID column. RYR1's DEL + INS pair at one locus is read as the delins it is. HGVS-named alleles are written with spaces dropped and & for the comma (c.1129-5923C>G&c.1236G>A(HapB3)), so DPYD, RYR1, CACNA1S and MT-RNR1 draft; --allele accepts either spelling and keeps Reference as it keeps *1. What still cannot be spelled is reported by reason: a bare DEL, alternatives (GAGTCG(3) OR GAGTCG(4)), and a reference that disagrees or could not be read (every notation row under --offline). draft_gene gains a sequences argument; the CPIC client no longer reports notation rows as skipped. See ENRICHER.md § PharmVar and CPIC gotchas and ROADMAP_HISTORY § RM253.

2026-11-04 — RM334: an HLA allele name is checked against IPD-IMGT/HLA's grammar (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, enricher — An HLA allele is an allele_function.csv row (there is no HLA table). Where gene starts with HLA-, allele must be *, then two to four colon-separated numeric fields of at least two digits, then an optional expression suffix (N, L, S, C, A, Q): *15:01, *57:01:01 and *02:01:01:02N pass, and 15:01, *1501, *15:01X and *15 are refused by validate and compile alike, with a message naming the shape. Every other gene keeps the permissive name rule. The published-module scan found no HLA row in any reference example or in any of the 46 versions in the registry's catalogue, so the refusal breaks no published module. Whether the allele exists in IPD-IMGT/HLA's list is not checked; that lane is in 0.9. The CPIC drafter skips an HLA allele the model refuses, with the same reason (pgx.hla_allele_refusal), instead of failing the draft.

2026-11-04 — RM322: a diplotype row states its trade-off and its report group (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler — diplotypes.csv gains two optional free-text columns, negatives and category, the variants.csv columns of the same names with the same descriptions. A scored diplotype such as APOE ε, which v2 scores once as a pair, can now carry the trade-off and the report group a variant row could. Both reach diplotypes.parquet and come back through reverse. category binds no vocabulary, as on VariantRow. A stated cell moves content_signature; a module that leaves both out keeps its signature, but its artifact.digest moves on a 0.8 recompile because the parquet has two more columns. They are the last two columns, so no 0.7 column changes position. The maintainer cut priority, actionability and flags from the item.

2026-11-04 — RM336: a DOI in pmid is refused, and pmid=NONE cites a preprint by its doi, on every citing table (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler, enricher — A pmid cell whose only identifier is a DOI is refused, naming the DOI; before, 10.1101/2025.03.01.640000 parsed as six PubMed ids and the literature pass looked each one up. extract_pmids skips a DOI's digits, so a PMID beside a DOI is kept. A work PubMed does not index is cited as pmid=NONE (exactly), which makes doi mandatory and DOI-shaped. One rule, spec.CitationPair, on every model with a pmid (StudyRow, the four binning kinds, PharmVariantRow, GwasEffectRow), asserted by an equality over the model registry; the last three gain an optional doi column (nullable in their parquets, so their artifact.digest moves on recompile). Two NONE study rows on one variant are told apart by their DOI in the duplicate check, and study findings name a NONE row by its DOI. A quote on a NONE row, which the literature pass cannot match (it reads text by PubMed id), is counted and reported as unsupported rather than as no quote, and a module citing only by DOI gets an unsupported existence record instead of a refusal. Every published module (the prod registry, 24 modules, all 47 versions, and reference_examples/) was scanned first: no pmid cell holds a DOI. A 0.7 reader refuses NONE and the new columns (INTEGRATION_0_8 § 1.5).

2026-11-04 — RM304: a vrs_id says which VRS minted it, and the module pins one (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler, enricher — VRS 2.1 re-ids some ambiguous insertions, so an id now carries the version that minted it and the manifest pins the version a module's ids are meant under.

  • The per-row version is resolution.csv's vrs_spec, which the enricher has written since 0.5 and nothing read. No new column: the item was filed believing nothing recorded a version, and a second column for the same axis would break P5. The enricher now reads the version off the library (ga4gh.vrs.VRS_VERSION, "2.0.1") and stamps it at the spec's minor ("2.0", as before) through vrs.vrs_spec_release. It stamps only rows it mints. A recorded id's cell is left alone, and empty means unstated.
  • The pin, manifest.compilation.vrs_version, is the VRS minor the compiler verifies against ("2.0"), or null when no allele in resolution.csv carries an id.
  • New warning code vrs_version_differs: a row whose vrs_spec names another version is reported beside each allele's own verdict. A substitution that recomputes differently is still the mismatch error within the pin's major, and is counted in this line only under another major, where the digest serialization differs. One line per stated version, a warning in both modes, actionable (re-mint). vrs mint and enrich report the same code for recorded ids under a version other than the minter's.
  • ga4gh.vrs stays capped <2.4. The insertion rs72613567 is pinned to ga4gh:VA.Jml7SNku3QQBCVIj78BGiFvR21bNkos7 behind JUST_DNA_NETWORK_TESTS=1. No parquet carries the version yet; RM270 carries the minted id and its version onto the artifact.

2026-11-04 — RM342: the manifest records the toolchain that wrote the bytes, and a digest it explains says so (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format, compiler, enricher — Every compile now stamps manifest.compilation.toolchain, the Python and polars that wrote the bytes ({"polars": "2.1.0", "python": "3.13.1"}; a library added later is a new key). It is optional and outside artifact.digest, like compiler_version, and null on older manifests. From polars 2 every parquet footer names its polars release, so one compiler beside two polars mints two digests for equal content. Where a fresh compile meets a recorded digest, that case now says so: upload raises ToolchainDiffersError (a subclass of PublishCollisionError, still a refusal unless --force) and prints TOOLCHAIN DIFFERS, and the release sweep reports bytes_reason: "toolchain_differs" per module. If either manifest records no toolchain, or both record the same one, the old outcome stands. Principle 4 is amended: reproducibility is pinned via compiler_version, the recorded toolchain and the resolved reference. verify_manifest is unchanged. See INTEGRATION_0_8.md.

2026-11-04 — RM337: a coded warning names each file under logs/ that does not ship (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format — VALID_WARNING_CODES gains log_file_not_shipped (actionable). An older reader refuses a manifest that carries it (INTEGRATION_0_8 § 1).

compiler — Every path under logs/ that discovery leaves behind now draws one warning, in both modes and never a refusal: <path> does not ship with the module: <reason>. The reason is one of the fixed sentences in LOG_NOT_SHIPPED_REASONS: a hidden path; a symlink, including one to a directory or to nothing; an image over LOG_IMAGE_MAX_BYTES, with both sizes; an image whose first bytes are not its format's signature; an SVG, which can carry script; any other extension; text that is not UTF-8, with the offset of the first bad byte; a .json that does not parse; and a .jsonl line that does not parse, with the first such line's number. _log_content_agrees became _log_left_behind_reason, which returns the reason or None and so decides and explains in one place, and _collect_logs returns its warnings beside the entries. The warnings reach CompilationResult.warnings, manifest.compilation.warnings_summary and the compile output. One line per file, with no cap. An explicit log_files list draws none, and the top level is unchanged. What ships, and both identities, are as RM338 left them. Before this, the registry's server-side compile dropped such a file without a word to the uploader (S125).

2026-11-04 — RM338: logs/ ships checked text and images, not only *.log (patch, on main)

compiler — Under a module's logs/, discovery now ships .txt, .md, .json and .jsonl when the content agrees with the name (UTF-8 text; .json parses; every non-blank .jsonl line parses), and .png/.jpg/.jpeg when the file starts with its format's magic and is at most 5 MiB (LOG_IMAGE_MAX_BYTES). SVG never ships, and neither does a hidden path or a symlink of the new kinds. *.log ships unchecked, as before, and the top level is unchanged. Each newly shipped file is listed in manifest.logs like any log and named by an info line on just_dna_compiler.compiler, and by a log ships: <name> line in just-dna-compiler compile, so an author sees it before publishing. A file that fails its check stays behind as before, with no warning: that warning is RM337, a minor. An explicit log_files list is unaffected. Neither identity moves, since logs are outside artifact.digest and content_signature. A module with a stray notes.md under logs/ ships it at its next recompile (S125).

2026-11-04 — RM340: the VRS-coverage warning points at the record that explains its no-ALT rows (patch, on main)

compiler, enricher — When an attested verification.json carries an rsid_coordinate_agreement record saying the injected Ensembl snapshot has no record for some authored rsIDs, the no-ALT line of vrs_coverage_incomplete now names that record. Those rows' empty ALT is explained there, and a newer snapshot may fill it (S127). The line's existing phrase is unchanged and leads it. The phrase the enricher writes is now one constant, compiler.SNAPSHOT_UNPLACED_PHRASE, shared by writer and reader. No new warning code.

2026-11-04 — RM339: an interrupted snapshot pull no longer reads as complete (patch, on main)

enricher — A snapshot pull now writes pull-record.json from the remote listing before its first download. A pull that stops part-way therefore names the files it never got: every parquet resolver reports the lane absent instead of present, and the next cache pull, cache prepare or ensure_* fetches only what is missing. Before, a cache holding 3 of 25 whole chromosome files was trusted, and enrich would have resolved against it (S126). A cache with no record, built locally or pulled before this release, is compared with the published listing once; with no route to the Hub it is used as it is, with a warning. A download that times out names HF_HUB_DOWNLOAD_TIMEOUT. New in locations: PULL_RECORD_FILENAME, parquet_footer_ok, snapshot_record_missing.

2026-11-04 — a new major of an upstream: the liaison maps it, and a cap needs a proven break (seats, on main)

repo — The person's rule, written into the liaison's and the lead's primings. The liaison decides what blocks adopting a new major of an upstream and hands the lead an adoption map. A dependency is capped only once the new major is shown to break something here, and the cap is lifted once support is integrated. polars 2 and duckdb 2 broke nothing that RM343 and RM344 do not cover, so neither is capped. The lead caps a dependency only as a safety, when the liaison is unavailable and work has to move. No package changes.

2026-11-04 — RM343: a query-shape test reads duckdb 2.0's Hash Join (patch, on main)

enricher (tests) — duckdb 2.0 prints the join operator as Hash Join where 1.x printed HASH_JOIN, so on a fresh install, which takes 2.0 under duckdb>=1.1.0, one test failed although the plan is the same hash join. The assertion now reads the name in either spelling. The rest of the enricher suite passed under 2.0.0 (the liaison's run), so duckdb stays uncapped. No package changes.

2026-11-03 — the triage seat's kit watchers replace the hook-armed inbox watcher (repo tooling, on main)

repo — Since the seating, the triage seat's watchers in .blueprints.toml wake it on a consumer's note. watch.inbox is the kit's file watcher on CONSUMER_SUGGESTIONS.md, with the same 150 s settle; watch.inbox-commits catches an item pushed to origin instead. The old pair would have rung a second time on every note, so it is removed: the UserPromptSubmit hook .claude/hooks/arm_triage_watcher.py, the wrapper .claude/wake-on-suggestions.sh, the watcher .claude/watch-suggestions.sh and their test schema/tests/test_triage_watcher.py. .claude/settings.json now holds no hook. Triage confirmed its kit watcher fired on the real S125–S127 edit before the removal. CONSUMER_TRIAGE_LOOP.md §1 describes the seat's watchers, and the gist's watch-inbox.sh is no longer adopted here. No package changes.

2026-11-03 — triage of S125–S127: four items filed, one doc fix (docs only, on main)

docs — The triage seat's first pass answered the three consumer items that had sat in the inbox, uncommitted, since October. None had been worked on anywhere, and all three reproduced against the code. Filed as patches: RM337 (a file under logs/ that is not *.log is dropped from a green compile with no warning), RM339 (an interrupted snapshot pull leaves a cache every caller reads as complete), and RM340 (the VRS-coverage warning does not point at the verification.json record that already explains its no-ALT alleles). RM338 records the undecided half of S125, which files under logs/ ship. TABLES.md's overrides.csv section now says that a hand-written uncited literature.csv row (RM147) is an addition the refusal there does not cover. No package changes.

2026-11-03 — the repository runs on seats: lead, pm, triage, liaison (docs only, on main)

repo — Four standing seats from the seats kit replace the two-seat description in RELEASE_CYCLE.md § Seats. The pm sizes minor and major scopes, the lead executes them, triage keeps the consumer inbox and the patches on main, and the liaison reads upstream releases. Their primings are in docs/seats/, now listed in mkdocs.yml's not_in_nav; the seat pull requests had left test_nav_and_not_in_nav_partition_the_docs_directory red. No package changes.

2026-11-03 — RM327, in part: tradeoff is a reserved column name (patch, on main)

format — tradeoff joins vocab.RESERVED_NAMES_0_4. The maintainer decided that a trade-off (an allele protective on one trait and risk on another) is computed from a per-trait benefit/harm table, never authored. That table is RM327's minor half and is not built yet. Reserving the name keeps a hand-set column out before the computed one lands. longevitymap v2 has trade-off loci to write, and this is the word an author would reach for.

What changes for you. A tradeoff column on any authored table was already refused by extra="forbid". It is now refused with a specific reason that points to negatives for describing a trade-off today, and says that direction: contested means something else. Every reserved-name error now lists four names instead of three. No module's acceptance changes, and no digest moves.

2026-10-02 — RM320: PubMind's table is CC BY-NC-ND 4.0, so a module carrying it compiles only as non-commercial (patch, on main)

enricher — WGLab answered in writing (WGLab/PubMind#5) that the PubMind table ANNOVAR and Zenodo distribute is CC BY-NC-ND 4.0. PUBMIND_TERMS now records commercial_use=False, share_alike=False and redistribution=True, with NoDerivatives in notice. Before, every axis was None.

What changes for you. A module whose sources.csv carries a freshly written pubmind row now refuses to compile under unstated or commercial and compiles under non_commercial. Before, it compiled under any declaration and manifest.sources named pubmind in unknown_terms_sources. pubmind build gains --use and skips under unstated, as clinpgx build does, and cache rebuild applies the same gate to the lane. Pass --use non-commercial (or JUST_DNA_USE=non-commercial for scripts/rebuild-caches.sh). draft-panel --source pubmind is still not gated, and its warning now names the licence. pubmind publish still refuses, now because the snapshot is adapted material that NoDerivatives bars sharing.

A module drafted before this keeps its all-null pubmind row, because a sources.csv merge never overwrites a recorded row. It goes on compiling as before. To record the terms, delete the pubmind row and re-draft.

2026-09-30 — RM316: a local-allele field is resolved through LAA before an element rule applies (patch, on main)

format / docs — annotated_alt is defined by ALT index, and VCF 4.5's local-allele fields (LAD, LPL, …) order their values by the sample's LAA instead. A consumer whose merged file carries only LAD would read another allele's depth. The rule's printed meaning and SCHEMAS.md now say to resolve the local field to its global equivalent through LAA first, as the spec recommends. Text only; no new rule.

2026-09-30 — RM315: the VCF cardinality table is all of 4.5's reserved keys, so INFO/RUC is checked (patch, on main)

format — vocab.VCF_FIELD_NUMBER held only Tables 1 and 2 of VCF 4.4, which left out the keys §3 and §4 reserve, among them RUC, the field binning names for a repeat count. A source_field=INFO/RUC with no element rule therefore drew no warning. The table now transcribes every key VCF 4.5 reserves, and CICN joins the INFO/FORMAT collision set. VCF_NUMBER_MEANINGS covers 4.5's LA, LR, LG and M. Existing modules may see vcf_pointer_unselected_element or vcf_pointer_key_collision on pointers that were silent before, with the same codes as always. None of the shipped examples does.

2026-09-30 — RM314: a reference block's extent is read per sample, from FORMAT LEN first (patch, on main)

format / docs — VariantRow.callable_from's description and SCHEMAS.md's VCF-reading contract told a consumer to find a gVCF reference block by INFO END. VCF 4.5 made END the maximum over every sample's FORMAT LEN, so on a multi-sample file it credits one sample's block to another's positions: a confident absence where nobody called. The rule now reads LEN first and END only where LEN is absent, and a . GT inside the sample's own earlier block is that block's call. Text only; single-sample and 4.4 files read exactly as before.

2026-09-30 — RM288: the CPIC and PharmVar builders and the PGx drafter export their error families (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — cpic_build.ERRORS, pharmvar_build.ERRORS and pgx_draft.ERRORS are callee-owned tuples of every type each surface raises, and every caller in the tier catches <module>.ERRORS instead of spelling its own list (RM310 was such a list missing one member). An AST test refuses a caller that spells the list. draft --use commercial against CPIC now prints DRAFT FAILED instead of a traceback.

2026-09-30 — RM283: a ClinPGx build says when its archive has gone quiet (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — A ClinPGx build that downloads now HEADs its archive and two sibling archives, records their Last-Modified in release.json as archive_last_modified, and warns (never refuses) when the lane's archive is 180 days or more behind the newest sibling. RM175's incident, a fourteen-month-old archive still answering 200, would have been named on the first build.

2026-09-30 — RM263: a DOI's PMID, through Europe PMC (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — lookup_citation(doi=…) / hint citation --doi with no PMID now asks Europe PMC which PubMed record the DOI names (EuropePmcClient.pmids_for_doi, new) and offers it as an advisory, never a fill, then asks PubMed which paper that is. Crossref's title and PubMed's are reported as two info findings; several PMIDs withhold the advisory and are named; none is reported as Europe PMC's answer.

2026-09-30 — RM258: the literature pass records how the fulltext question ended, and asks an outage again (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher + format — A Europe PMC or BioC 5xx used to read like "no copy": the row was written on the abstract and merge-not-clobber never asked again, so an outage became a permanent abstract-only pin. literature.csv gains the derived column fulltext_status (retrieved | absent | unreachable, vocabulary VALID_FULLTEXT_STATUS), and a later run re-asks only the fulltext question for a pinned row that is unreachable, or null and not already matched against a fulltext (rows written before 0.8, asked once). Both clients gain fetch_fulltext, which raises the new FulltextUnavailable for an unanswered request; fulltext() keeps its old contract.

2026-09-30 — RM272: EnrichmentResult names three answered-but-withheld states (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

enricher — unanchored_rsids (live Ensembl's one-sided indel whose anchor base could not be read, RM268; transient), non_nucleotide_rsids (every allele either Ensembl rung gave is not a base, RM271; permanent) and reference_rejected_loci (per locus, the indel the reference showed is a different event, RM274). Each rsID is in exactly one list and still in unresolved. The logs are unchanged. resolver.lookup_loci takes an optional non_nucleotide= collector, and EnsemblResolver gains non_nucleotide_rsids.

2026-09-30 — RM278: compile --ensembl-reference and --ba1-threshold (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler — The compile command gains --ensembl-reference (stamps manifest.compilation.ensembl_reference, which a CLI compile could not set before) and --ba1-threshold. test_cli_parity.py now asserts the command's options equal compile_module's parameters minus a declared set of four, each with its reason (the auto-discovered log, provenance, logo and readme paths).

2026-09-30 — RM300: the manifest counts rows and variants per ClinVar tier (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

format + compiler — stats.pathogenic_count/benign_count fold a tier pair each, so a catalog could not split pathogenic from likely_pathogenic without reading the artifact. Two new optional Stats fields, clin_sig_counts (authored rows per tier) and clin_sig_variant_counts (distinct variants per tier), keyed on every VALID_CLIN_SIG member with zeros included, from the effective clin_sig. The existing counts keep their meaning. Outside artifact.digest; a 0.7 reader ignores them.

2026-09-30 — RM313: the QUAL warning that stated the spec backwards is retired, and the true finding replaces it (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler + format — RM57's quality_floor_inverted told authors that a high QUAL on a reference record means the position is probably variant, and recommended GQ. VCF §1.6.1 item 6 says QUAL is the confidence in what ALT asserts, so a high QUAL there is a confident reference call, and GQ is conditioned on the site being variant. The same rows (requires_callable with a QUAL floor) now warn as qual_floor_on_reference_evidence: the reference block such a row is decided from usually leaves QUAL missing, so the floor cannot be read. No field is recommended. quality_floor_inverted stays in VALID_WARNING_CODES, emitted by nothing, and is declared in the new vocab.RETIRED_WARNING_CODES; QUAL_INVERSION_PHRASE is replaced by QUAL_REFERENCE_EVIDENCE_PHRASE. The new code is an old-reader break, in INTEGRATION_0_8.md.

2026-09-30 — RM308: a conclusion naming another genotype warns at validate and compile (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler + format — RM279's rule ran only on the authoring surface (hint, lint_rows), so a module nobody linted compiled green under --strict with two conclusions swapped. It now runs in the pre-flight as conclusion_names_other_genotype, a new member of VALID_WARNING_CODES (an old-reader break, in INTEGRATION_0_8.md), one sentence per reason: a swapped pair, or text copied from a neighbour. A warning in both modes.

2026-09-30 — RM275: two fact-table rows under one merge key that disagree now warn (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler + format — frequencies, gene_metrics, literature, gene_validity, clinical_assertions, gwas_effects and expression_effects had no duplicate rule, so two rows under one merge key stating different values compiled green under --strict and a consumer read whichever came first. They now warn in both modes as fact_rows_contradict, a new member of VALID_WARNING_CODES. That is an old-reader break: a 0.7 just-dna-format refuses a manifest carrying the code, recorded in the new INTEGRATION_0_8.md. Refused from 1.0.

2026-09-30 — RM241: reverse writes back every spec field the manifest carries (0.8, uncut)

Cut as 0.8.0 on 2026-11-05, with the rest of the 0.8 branch; the entry below was written while the line was uncut.

compiler — reverse_module now reads the artifact's manifest.json for the whole spec header: the module: block's display fields, version, genome_build, the defaults' curator/method, license, weighting, authorship, authority_precedence and panel. The README and logo are copied back after a hash check. Before, the display block was rebuilt from the module name and the rest was dropped, and the round-trip test passed because both identity digests leave the manifest out; it now also compares the manifest's authored projection. Surface: reverse_module(icon=, color=) and the CLI's --icon/--color default to None rather than "database"/"#6435c9" (the fallback for a bare parquet directory is unchanged), and compiler.reverse gains REVERSE_RECONSTRUCTED and REVERSE_EXCLUDED. No digest moves.

2026-09-29 — RM310: a CPIC outage during cache rebuild is a failed lane, not a traceback (patch, on main)

enricher — caches._rebuild_cpic now catches cpic.CpicError, which CpicClient raises on a transport failure or a 5xx. Before, it escaped cache rebuild and rebuild_caches, so every lane after CPIC was neither built nor reported. A new test walks CACHE_LANES and asserts every adapter returns a failed outcome for its builder's documented errors.

2026-09-29 — RM290: each overridable table states whether an unmatched correction can mean the source caught up (patch, on main)

format — overrides.OverlayTarget gains vindication (sole / withheld / none, closed as VINDICATION_READINGS) and a mandatory vindication_reason. VINDICATING_OVERLAY_TABLE (a string) is replaced by VINDICATING_OVERLAY_TABLES (a frozenset read off the registry). Only clin_sig_concordance.csv is sole, so no finding moves and no manifest changes. A caller importing the old constant should switch to the set (nothing in the ecosystem did). RM311 carries the gene_validity case, which is a value comparison rather than an absence.

2026-09-29 — RM307: the Atlas client waits on a pacing gate, at the measured interval of zero (patch, on main)

enricher — The AlphaGenome Atlas showed no throttle up to 21 calls/s (about 1,270 a minute, 3,570 calls in all), so atlas_client.ATLAS_REQUEST_INTERVAL is 0.0. AtlasClient now waits on a net.PacingGate once per attempt, so client.gate.spent counts Atlas attempts like every other client's, and AtlasClient(..., gate=) / connect(..., gate=) take a shared gate with a real interval. Behaviour for an existing caller is unchanged. The measurement is in ALPHAGENOME_ATLAS.md § 6.7.

2026-09-28 — RM279: hint names a conclusion written for another genotype (patch, on main)

compiler — hints.inspect_rows on variants.csv (so just-dna-compiler hint variants.csv and the creator's lint_rows) now warns on the conclusion column when the text names genotypes built from the locus's own alleles and never the row's own, and gives the line of the row carrying the genotype it names. A swapped C/C/A/A pair shows as two findings pointing at each other. Over 653,706 diploid rows it fired 10 times, all real. It does not run at validate or compile: that half needs a new warning code, which is a minor (RM308). The rule is just_dna_compiler.conclusion.

2026-09-28 — RM280: the Atlas client retries in the house layer, and JUST_DNA_HTTP_RETRY_ATTEMPTS reaches it (patch, on main)

enricher — The AlphaGenome Atlas channel no longer retries inside grpc. connect opens it with grpc.enable_retries = 0 and the vendored service config minus its retryPolicy (atlas_client.channel_options()), and AtlasClient retries each RPC itself under attempt_floor(5) with upstream's backoff. Before, grpc capped the retries at five and JUST_DNA_HTTP_RETRY_ATTEMPTS could not raise them. Now it can, like every other client. The retried statuses are the ones _translate maps to AtlasUnavailable, so INTERNAL is now retried, which upstream's config never did. Exceptions and their types are unchanged. A caller sees AtlasUnavailable after the last attempt, as before. The client still has no pacing gate (RM307).

2026-09-28 — ga4gh.vrs capped below 2.4 until RM304 records the VRS version (patch, on main)

enricher — The dependency is now ga4gh.vrs>=2.3.3,<2.4. VRS 2.1.0 (2026-09-01) normalizes an ambiguous insertion to the smallest repeat unit where 2.0 took the greatest, which moves that allele's ga4gh:VA. id. vrs-python has not implemented it yet (issue #637), and a stored vrs_id does not say which VRS minted it. Without the cap, the first release that ships the rule would make a fresh enrich mint ids that disagree with resolution.csv, gnomAD and ClinVar-GKM, with nothing to say why. The locked version stays 2.3.3, so nothing mints differently today. A consumer that needs ga4gh.vrs 2.4 (metakb's server pins a 2.4 alpha) conflicts until RM304 lifts the cap.

2026-09-28 — RM297: the ClinPGx currency check no longer refuses on an annotation a row never cited (patch, on main)

enricher — enrich_clinpgx (clinpgx check) stops at a row's cited annotation_id when the snapshot does not hold it, instead of comparing the row against another annotation that shares its category (S122: a withdrawn 1A read as "ClinPGx says 3"). The warning names three readings (withdrawn, mistyped, never a ClinPGx accession) and what the snapshot holds there. A row citing no annotation_id whose category or triple match carries another level is still reported. Both land in the new ClinPgxResult.withheld and in warnings, coded clinpgx_annotation_not_in_snapshot and clinpgx_level_differs_from_uncited_annotation. strict is loosened for these rows: they no longer refuse. A difference reached through the row's own cited annotation still refuses, and its text is unchanged. verification.json's pgx_evidence_level counts both lists in findings and names the codes in detail, so an unchanged module does not publish a lower number. A cited id with no match at all used to be an unmatched miss and is now a finding. RM298 is the per-row source column that would let a withdrawal refuse again.

2026-09-28 — RM289: two author-facing sentences say what is true (patch, on main)

format — PgsRow.training_ancestry's description said the score was validated in those superpopulations. It now says developed and evaluated in, naming the PGS Catalog's development and evaluation samples, which are the two stages the enricher's pgs_metadata_agreement check compares the cell against. No value, field or digest moves.

enricher — The declared-use gate's skip for a source whose terms are unknown kept its leading phrase ("terms could not be established, so the data is not used") and now says that no declared use changes the outcome, then quotes the terms' recorded notice and license_url. For PubMind that is why its terms are unsettleable, not a constant left unfilled (S99).

2026-09-27 — RM301: a client reads its credential from .env without exporting the file (patch, on main)

enricher — Constructing EutilsSettings, CrossrefClient, PmcIdConverterClient or PharmVarClient no longer writes the working directory's .env into os.environ. The same holds for the retry floor, the PharmVar rebuild guard, the AlphaGenome key and the HuggingFace token reads. Each reads its one variable through the new locations.env_value: an exported value wins, and the .env beside the working directory is read without being exported. A credential kept only in .env still arrives, whatever the call order (RM100). Behaviour change a host may notice: code that read a .env-only variable from os.environ after constructing one of these clients, relying on the enricher to have exported it, now sees nothing. The cache resolvers still export, and load_dotenv_file=False still declines that. Reopens RM102's credential half, from S124.

2026-09-27 — RM299: the stats ClinVar counts say what they count (patch, on main)

format — Stats.pathogenic_count and Stats.benign_count now say they count authored variants.csv rows, one per genotype, whose legacy boolean folds a tier pair (pathogenic + likely_pathogenic, benign + likely_benign). clinvar_count says the same about grain. Before, the description read "Rows flagged ClinVar-pathogenic", and a catalog showed just-dna-seq/pathogenic's 617,822 as that many pathogenic variants (S123). No value moves, and no field is added. Per-tier counts are RM300, a minor.

2026-09-27 — release numbering: one number per cut, from v0.7.4

Process only. From the next cut, every package that changed since the previous tag takes the tag's number (a skipped number is legal), unchanged packages keep theirs, and each floor names its dependency's current version, so it is the release number exactly when that dependency moved. So v0.7.4 will publish each moved package as 0.7.4. v0.7.3 (format 0.7.1, compiler 0.7.2, enricher 0.7.3) is published and stays as it is. The rule is in RELEASE_CYCLE § Patches.

2026-09-27 — v0.7.3: format 0.7.1, compiler 0.7.2, enricher 0.7.3 (a patch, cut from main)

All three packages move, each by a patch, and the intra-workspace floors move with them (compiler needs format >=0.7.1; enricher needs format >=0.7.1 and compiler >=0.7.2), because the workspace pins each floor to its dependency's current version. Nothing adds a field, table, parameter, public function or warning code. The sweep gate over the 16 reference modules, compiler 0.7.1 against this cut from one spec root, moved nothing on any axis (content_signature, manifest fields, parquet bytes, parquet schema, warnings), so a compiled module needs no recompile for the compiler's sake. The corrections are in what the enricher derives, and a module picks them up by re-running it.

  • format 0.7.1 — RM264: the quote-counter descriptions.
  • compiler 0.7.2 — RM264's docstring and RM277's quote_counter_stale remedy text.
  • enricher 0.7.3 — RM268, RM271 (the Ensembl rungs), RM273 (MITOMAP events, CIViC left-alignment, PubMind), RM274 (same-size indels settled with the reference), RM276 (STRchive's own grade), RM277 (stale literature pins re-fetched), RM293 (MITOMAP : deletions). Each entry below says what moves.
  • Tests and examples — RM291 (a consumer-join test against a normalized ClinVar slice) and RM295 (cyp2d6_structural's CYP2D6*4 coordinate).

After upgrading: just-dna-enricher enrich <spec> --rederive for resolution.csv, the literature pass online for a stale literature.csv, and just-dna-enricher cache rebuild --only mitomap_miss (a derived lane, never published). No HuggingFace snapshot changed.

2026-09-27 — RM293: MITOMAP's : deletions are anchored on rCRS and joined (patch, on main)

just-dna-enricher only; a corrected derivation of a derived lane. MITOMAP writes a deletion with an empty alt (m.7402del, refna=C, regna=:), and the miss lane called those 39 rows unmintable, citing Principle 2. The lane has vendored rCRS since RM273, so it now anchors each on the base before it (after checking MITOMAP's bases against rCRS) and joins it. On the local parents: unmintable 47 → 8, photocopy 663 → 686, rated_miss 1 → 2, unrated_miss 385 → 400. Rebuild the lane (just-dna-enricher cache rebuild --only mitomap_miss); it is never published. RM293.

2026-09-27 — RM295: cyp2d6_structural places CYP2D6*4 where GRCh38 has it (patch, on main)

A reference-example correction; no package changes. The example stated CYP2D6*4's splice variant at 22:42127941 C>T, where GRCh38 has G; it is rs3892097 at 22:42128945. Both rows moved and the example was re-closed. Anyone who copied the old coordinate from the example should move it too. Found by RM291's consumer-join test.

2026-09-27 — RM291: a test that joins every reference example the way a consumer does (patch, on main)

Tests and fixtures only; nothing in any package moves. compiler/tests/test_consumer_join.py compiles every GRCh38 reference example and joins each allele row on (chrom, pos, ref, alt) against a bcftools-normalized ClinVar slice (assets/consumer_join/). 614 distinct allele keys ClinVar carries join; the 2 misses (SHOX, spelled one base right) are pinned to RM270. It also found RM295, below. RM291, postmortem M4.

2026-09-27 — RM274: an indel the allele strings cannot decide is settled with the reference (patch, on main)

just-dna-enricher only; a corrected derivation of the fill, declared here (Principle 3). When an rsID-only row's genotype and the resolved locus name indels of the same size with different content, hosting_verdict withholds and the locus used to be kept regardless. enrich now reads the GRCh38 reference around the locus and slides its indel across the repeat: if the genotype's payload is one of its spellings the locus is kept (same event), and if not it is dropped from resolution.csv and named in one warning, because the genotype belongs to a different variant. Offline or unreadable, it stays undecided and kept, as before. What moves: a module whose genotype named a different same-size indel than its rsID's locus now resolves that row to nothing instead of to the wrong coordinate. RM274, postmortem P2.

2026-09-27 — RM276: draft-repeats names a locus STRchive itself doubts (patch, on main)

just-dna-enricher only; notes, no schema or field. STRchive grades every locus, and a drafted row for DMD (graded Refuted) looked exactly like HTT's. draft-repeats now names a drafted locus graded Refuted or Disputed (DMD, DIP2B, NIPA1, POLG) and counts Provisional apart as not yet curated. Both modes, never a gate. Carrying the grade into the module is RM294, a minor. RM276, postmortem P4.

2026-09-27 — RM273: three enricher surfaces stop keying an indel on its spelling (patch, on main)

just-dna-enricher only; corrected derivations, declared here (Principle 3).

  • The MITOMAP-minus-ClinVar lane joins on events. Both sides' indels are left-aligned against a vendored rCRS (NC_012920.1) before the join. Every machine's mitomap_miss lane is stale until it is rebuilt (the lane is derived and never published; rebuild it with cache rebuild --only mitomap_miss): on the same parents, photocopy 655 → 663 and rated_miss 6 → 1, because five of the six "new" calls were ClinVar's own at another anchor. A MITOMAP draft from an older build still carries them; rebuild the lane.
  • A CIViC-drafted indel is left-aligned. The ClinGen registry's anchor is HGVS's 3′-most point, so a drafted indel in a repeat was right of every caller's spelling (rs72613567: 4:87310241 A>AA, now 87310240 T>TA). Rows drafted earlier keep their spelling; drafting only appends.
  • PubMind's withheld-indel warning states a count: its indels are not left-normalized, 5,158 of 20,006 provably so from the row alone.

No new field, parameter or public function. RM273, postmortem P1; RM293 filed for the : deletions.

2026-09-27 — RM277: a stale literature pin is re-fetched, so "re-run the pass" works (patch, on main)

just-dna-enricher + just-dna-compiler (warning text only). The compiler's quote_counter_stale told the author to re-run the literature pass, and the pass never refetched a pinned citation, so the remedy could not work; the enricher CLI said to delete the sidecar instead. Now the pass re-fetches a pinned citation whose quote count no longer matches studies.csv and re-checks its quotes, and both tiers say the same thing: re-run online (offline keeps the pin). A corrected derivation of a derived row (Principle 3): a re-run over a sidecar that predates its quotes rewrites those rows. No schema, field or code change. RM277, postmortem P5.

2026-09-27 — RM271: neither Ensembl rung serves a non-base allele as a locus (patch, on main)

just-dna-enricher only; a corrected derivation, declared here (Principle 3). Both Ensembl rungs wrote allele strings that are not bases into resolution.csv as status=resolved: live REST's literal dbSNP_novariation as a ref (rs2100212723), and the snapshot's empty alt, <.>, IUPAC placeholders (<R>, <Y>), N runs and N-masked chrY refs. Every read now keeps only the ^[ACGT]+$ alleles: a non-base ref withholds the locus, a non-base alt is dropped, and a locus with no alt left is withheld.

  • What moves: alts narrows where a placeholder or N allele sat beside real ones (rs33946775, HBB: CA,CC,CG,<R> → CA,CC,CG); loci with no base allele leave the table; an rsID REST answers with none is now not_found (Ensembl records no variation there), with a warning. An existing sidecar keeps its old rows: delete it or run enrich --rederive.
  • The snapshot no longer calls such an rsID "not in the injected snapshot"; one warning names them.
  • No new public surface; a structured result field for these rsIDs is RM272, a minor.

2026-09-27 — RM268: the live Ensembl REST rung anchors a one-sided indel (patch, on main)

just-dna-enricher only; a corrected derivation, declared here (Principle 3). Nothing added to any schema. The live REST link used to copy Ensembl's -/C and interbase start straight into resolution.csv, so rs8176719 (ABO O1) resolved to 9:133257522 ref='-' alts='C': a ref the allele grammar refuses, one base right of every VCF record. It now reads the base at start - 1 and prefixes every allele, so the same rsID resolves to 9:133257521 T>TC, the ClinVar/dbSNP spelling. A one-sided deletion (AGTAAG/-) anchors the same way; strings with no - are untouched.

  • What moves: any resolution.csv row the live REST link wrote with ref='-' or an alts member - gets a new start, ref and alts. Only rsIDs the snapshot misses reach that link. An existing sidecar keeps its old row (merge, never clobber): delete it or run enrich --rederive.
  • When the anchor base cannot be read (sequence service down, or a contig outside the refget table) the locus is withheld. An answer with every locus withheld is resolve_rsid → (None, "ensembl-rest"): no row, the key stays unresolved, and the rsID is named by its own warning, apart from unreachable_rsids.
  • The GraphQL rung withholds a one-sided node and lets REST answer; its convention is unprobed.
  • No new public surface. The first commit added EnrichmentResult.unanchored_rsids, EnsemblResolver.read_base and sequences.grch38_base_reader, each minor-class and so not legal on main; the fix-forward commit made the reader private and dropped the field. A structured list is RM272, for the 0.8 branch.

RM268; RM267 (the snapshot's one-base-early insertions) is still open.

2026-09-27 — triage S117–S119: the haplotype convention written down, three items filed

Docs and one code comment; no behaviour change. TABLES.md § haplotypes.csv now states the rule the phase-ambiguity check already computes with: a haplotype carries the reference at every site of its gene that it does not list, closed-world per module. It also says the *1 exemption silences only the used-but-not-defined warning and infers no definition, so a gene whose reference has another name (NAT2 *4) defines it with reference-matching rows. The _REFERENCE_HAPLOTYPE comment claimed *1 can never appear in haplotypes.csv. That was probed false and is corrected. S118.

Filed, not built: RM267 (Ensembl's VCF dump anchors a class of insertions one base early, so resolution.csv can serve a different event; rs8176719, ABO O1, is one), RM268 (the live REST rung writes ref='-' at the interbase start), and RM269 (a hemizygous diplotype has no row, since haplotype_b is required; a 1.0 item under P8). S117, S119.

2026-09-25 — RM264: the 0.7.x manifest's quote counters say what an abstract-only miss is (patch, on main)

just-dna-format (next patch), descriptions only — nothing added, removed or retyped. Literature.quotes_found no longer claims to count fulltext hits only, and it and quotes_unchecked now send a reader to abstract_only_count: a manifest reading quotes_found: 0, quotes_unchecked: 0 beside abstract_only_count > 0 checked nothing against a body. A FAQ entry gives the per-row reading (quote_source = abstract beside quotes_found = 0 is unsettled). The field that settles it outright is RM256's quotes_checked, a minor's, on the 0.8 branch. S109.

2026-09-25 — RM262: a DOI lookup names the paper it found

just-dna-enricher only, inside the uncut 0.8 line. Additive: a new class and method beside CrossrefClient.exists, whose signature is unchanged. A minor.

  • lookup_citation(doi=…) / hint citation --doi now fills title, journal, year and first_author from the Crossref record it was already fetching, and adds an info finding DOI <doi> names: '<title>' (…) — existence is not identity, …. A record with no title says so. With a PMID given too, PubMed's record fills the fields and the DOI's title is reported beside it (S113).
  • CrossrefClient.work(doi) → CrossrefWork; literature.crossref_bibliographic parses a /works message the way bibliographic parses an esummary record.
  • Behaviour change on a failure leg: a 200 from Crossref whose body is not a Crossref record now withholds (exists=None) where it used to answer True. A maintenance page is not evidence that a DOI exists.

DOI → PMID is filed as RM263.

2026-09-21 — just-dna-lite adopted the 0.7 line (format 0.7.0 / compiler 0.7.1 / enricher 0.7.1)

Consumer-side record, no change to any package here. Written by the just-dna-lite side under the working agreement that cross-repo integration changes are noted in this file so parallel work in the other repos is not surprised. What INTEGRATION_0_7 § 3 asked of the reference consumer, and what it did:

  • Change 0 — cache pull → cache prepare: shipped as pipelines prepare-caches over caches.prepare_caches() rather than the enricher's console script, because that script cannot be imported in a workspace that also installs dagster (S107, filed today in CONSUMER_SUGGESTIONS.md: the Atlas gencode is protobuf 7.35, dagster pins protobuf<7, and google.protobuf.runtime_version.VersionError is neither of the two types RM247's guards catch). just-dna-lite mounts the enricher's Typer app behind a guarded import that catches that type and mounts a stub enrich status naming the reason when it fires.
  • Changes 1–3 — nothing to do, and recorded as such: no warning-prose matching on the consumer side, no derived-CSV reads on the consumer path, no reader of manifest.verification.
  • Check 4 — statistical_test and confidence/confidence_unit render in the report's study table, each column only when some study on the variant carries it; confidence renders with its unit or not at all.
  • Check 5 — pharm_variants.pmid renders as a PubMed-linked Citation row beside Evidence level.
  • Check 6 — requires_callable=True now excludes a row from reference-genotype restoration (restoration.hom_ref_rows); False and null are both kept, so the shipped corpus is unmoved.
  • Check 7 — clin_sig_concordance.parquet is discovered and joined on (variant_key, genotype); a discordant row renders an Authorities disagree note naming opposed. Nothing resolves the split. No reference example carries the table, so the consumer's test builds one from ClinSigConcordanceRow.
  • Check 8 — not read. Check 9 — contested passes through _effective_direction with sign 0 and no colour, and the word reaches the reader.
  • Pins: format >=0.7.0, compiler/enricher >=0.7.1, registry >=0.26.1. Both public registry instances answered registry 0.25.2 / format 0.7.0 / compiler 0.7.0 on the day.

One thing found on the way that is not the format's: the same relock floated agno 2.9 → 3.0.10, whose MCP toolkit imports the mcp 2.x names, against a consumer-side mcp<2.0 cap written for agno 2. Unrelated to this repo; recorded because the symptom (the whole CLI dead at import) looked like S107 until the traceback was read.

2026-09-25 — RM260: the three largest modules are packages, under the same import paths

just-dna-compiler and just-dna-enricher, inside the uncut 0.8 line. No surface change: every dotted path importable before resolves to the same object, --help is byte-identical, and every reference example compiles to identical bytes (compiled_at aside).

  • just_dna_compiler.compiler (was one 492 KB file) is 16 submodules: tables, load, the per-concern checks, validate, pipeline, manifest, parquets, reverse.
  • just_dna_enricher.cli (244 KB) has one *_commands module per source or command group, with the app assembled in __init__ in its original order. The just_dna_enricher.cli:app entry point is unchanged, and python -m just_dna_enricher.cli still works.
  • just_dna_enricher.enrich (152 KB) is subjects, outcome, build_declaration, orchestration, verification_records and resolution_csv.

Each __init__.py is a re-export shell kept for the published paths; RM260 removes them at 1.0. A consumer that monkeypatches a name on one of these modules must now patch the submodule where it is looked up. Two tests here had been reaching the live network that way without failing. Every source-reading guard walks the packages, and each was proved red on a planted defect.

2026-09-25 — just-dna-enricher 0.7.2: the CLI imports beside protobuf<7, and the frequencies leg asks about every allele

just-dna-enricher moves alone; just-dna-format stays at 0.7.0 and just-dna-compiler at 0.7.1. A patch cut from the 0.7 branch rather than from main: RM256, which landed in between, adds a manifest field and is therefore a minor, so it waits for 0.8 with RM257 (a new bioc= parameter) and RM259 (a new public function). What the patch carries is the two entries below plus a test that now arranges the absent snapshot it asserts. Neither touches a model, a parquet or the manifest, and the enricher's floors are unchanged.

  • RM254 — the release blocker for just-dna-lite (S107): every command died at import beside dagster's protobuf<7. See its entry below.
  • RM255 — hint variant --frequencies asked gnomAD nothing for a multi-allelic locus (S108). See its entry below.

2026-09-24 — RM259: a GRCh38 position to the genes worth asking the Atlas about

just-dna-enricher only, inside the uncut 0.8 line. Additive: one public function, so a minor.

gene_spans.genes_covering(chrom, position) returns every MANE gene whose span plus the 512 kb attribution horizon covers the position, or says why there are none (no_snapshot, no_gene_within_horizon). It is a query hint for a row that authors no gene (S112: 252 of 1,033 longevitymap rows). Each candidate goes to enrich_expression(spec, gene, chrom=, start=, end=), and the Atlas names the gene on what it returns. Nothing writes a candidate into variants.csv. Expect dozens of candidates: 50 at HFE H63D and 34 at APOE.

2026-09-24 — RM257: the literature pass reads an author manuscript through PMC's BioC service

just-dna-enricher only, inside the uncut 0.8 line (a minor). Additive: one new client and one new enrich_literature(bioc=) parameter, with no schema or compiled-output change.

A record Europe PMC flags isOpenAccess: N, which covers most NIH author manuscripts, was never asked for fulltext, so its quotes were checked against the abstract only (S110: Kunkle 2019, whose rows live in two body tables). The pass now asks PMC's BioC text-mining service for any citation with a PMCID that Europe PMC did not serve. Tables are kept and the reference list is dropped. It runs on the E-utilities pacing gate, and the licence columns stay what Europe PMC said. Rows already in literature.csv are not re-asked: delete it and re-run to pick this up. That a 5xx also leaves a pinned abstract-only row is filed as RM258.

2026-09-24 — RM256: the manifest published an abstract-only miss as a checked quote

just-dna-format, just-dna-compiler and just-dna-enricher, inside the uncut 0.8 line. A minor: one new optional manifest field.

An abstract-only citation stores quotes_found=0, quote_source=abstract, which is not a verdict, and the enricher's pass report already counted it as unchecked. The manifest's literature block counted only null rows as unchecked, so 24 quotes on one paywalled paper published as quotes_found: 0, quotes_unchecked: 0 (S109). The block now carries quotes_checked, in quote units: the quotes a retrieved text settled, found or missed. Read quotes_found against it; quotes_authored - quotes_checked is what nothing established. The rule is LiteratureRow.quotes_checked() and the enricher's report uses the same one. quotes_unchecked keeps its meaning (a count of citations), and its description now says so. Null on manifests compiled before this release.

2026-09-24 — RM255: hint variant --frequencies asked gnomAD nothing for a multi-allelic locus, and said nothing about it

just-dna-enricher only, shipped in 0.7.2. A patch: an advisory surface, no authored schema, no compiled output.

_lookup_frequencies kept only loci whose alts held one allele and returned silently when none did — so for a common GWAS lead SNP (rs3752246 is 19:1056493 G>C,T) no request was made and no finding appended, and an empty populations read as "gnomAD has no data" (S108). Every allele of every resolved locus is now asked about in one batched call, a caller's alts= filters that set, and the three ways the question cannot be put each say so. gnomAD has no record for … is per allele.

One additive key on an advisory payload: a populations row carries allele, variant_id and vrs_id, since a multi-allelic locus answers with one row per ancestry group per allele; the CLI's population line leads with the allele. hint.vrs_id is filled only when exactly one allele answered (@vrsid-per-alt).

2026-09-21 — RM254: the CLI died again beside anything pinning protobuf<7, and the guard that was meant to hold it caught two of three types

just-dna-enricher only, shipped in 0.7.2. A patch, and a release blocker for the reference consumer: just-dna-lite (S107) installs the enricher beside dagster, which pins protobuf<7; the Atlas bindings RM247 regenerated are stamped protobuf 7.35.1 and refuse an older runtime with google.protobuf.runtime_version.VersionError, which is neither ImportError nor RuntimeError, so both RM247 guards let it through and every command in the tier died at import.

  • atlas_protos.ATLAS_IMPORT_FAILURES — (ImportError, RuntimeError, VersionError), bound once in the stdlib-only module and named by all three guards (alphagenome_check, expression, and the lazy one inside _atlas_client_or_none). The RM247 walk asserts the name, not a per-guard set.
  • [atlas] floors protobuf>=7.35.1, the gencode stamp, read off the generated file by a test as the grpcio stamp already was. A co-install that cannot satisfy it now fails to resolve instead of at import; one that does not ask for the extra imports cleanly.
  • A third absence: client_absence() names a runtime older than the gencode with both numbers.
  • A subprocess test simulates the gencode refusal and asserts the console entrypoint still imports.

2026-09-20 — S102–S106: a drafter that claimed absence, a recipe that refused on its own stubs, a table that photocopied the module, and a gate that ignored the file it read

just-dna-enricher and just-dna-compiler, shipped in 0.7.1 (they landed after the cut commit and before the tag, which is what PyPI carries). Five consumer reports from one session building a PGx panel: four answered as shipped and one filed as a design item with its message half shipped. None needed a schema change — every fix is a message, a lookup, or a derived table's contents, so the class is a patch in both tiers.

  • RM249 (S102) — draft --drug azathioprine on TPMT said the snapshot had no row for it, and the snapshot had 35, every one keyed on TPMT and NUDT15. recommendations() rightly keeps single-gene rows only; the explanation for the empty result had three arms and the true one was unreachable from a snapshot. Both CPIC clients gain partner_genes(gene, drug), asked first when a drug comes back empty, and the message names the partner and RM28. Warfarin, which has no recommendation row at any arity, keeps its own arm, reworded to say so. Measured on the snapshot and recorded on RM28: 18 of 103 drugs are pair-keyed, six pairs, arity never above 2.
  • RM250 (S103) — scaffold then draft, the reference README's own recipe, refused on <<REPLACE>> in three fields the draft never reads, and offered genome_build= as the remedy, which is enrich()'s parameter and no drafter's flag. spec_genome_build now looks past the scaffold's stubs — only those, and only outside the build cell; a misspelt key or a placeholder in genome_build still refuses, quoting the diagnosis the placeholder sentence had been hiding. The compiler's append_rows names a scaffold stub row by its line and says to delete it or scaffold without --kind, instead of calling the file broken.
  • RM251 (S104) — a row authoring both an rsID and a coordinate was copied into resolution.csv as source=authored with no ref/alts, so every CPIC-drafted module compiled with "VRS allele identity covers 0/N". The same run had looked the rsID up for the pair check and discarded the answer. Such a row now takes the forward branch when the reference knows its rsID: the link's loci are recorded, an id mints, and the compiler's _verify compares two independent values instead of the module against itself. This changes what a re-run of enrich writes into a derived sidecar, not what the compiler derives — a recompile of an untouched module is byte-identical, so the release record (RM126) has nothing to declare. The caveat is for authors: a table written before this keeps its authored rows under merge-not-clobber — delete resolution.csv or run --rederive to record the reference's answer — and a module whose authored pair disagrees with Ensembl then warns in best_effort and refuses in strict, which is the finding the old branch could not produce.

  • RM252 (S105) — pgx said "cpic forbids sale and no use was declared" about a module whose licence table recorded declared_use=non_commercial for CPIC, and asked the author to assert it again. licensing.effective_declared_use reads the module's own recorded declaration for that source at that layer when the flag states none, at all nine gates that have a module. The flag outranks the file in both directions, unstated on disk is not a declaration, and a leg with no row (PharmVar) still asks. PgxResult.recorded_use and the summary line say where a declaration came from.

  • RM253 (S106), filed — a repeat-count star allele has no home a diplotype can name. UGT1A1 *28 is TA(8) at rs3064744; <CNV:TR:n> is a length, repeat_alleles.csv is a bin, and neither is a name. Two message fixes shipped with the filing: the notation warning names the symbolic spellings the format holds and says the drafter does not translate CPIC's DELTCT/INSCGGG/TA(8) into them, and DPYD's HGVS-named diplotypes are no longer reported as copy-number notation.

2026-09-20 — just-dna-enricher + just-dna-compiler 0.7.1: four false-pass checks, and a CLI that could not import itself

just-dna-enricher and just-dna-compiler move; just-dna-format stays at 0.7.0 — a partial cut, the shape v0.6.3 and v0.6.4 already used, where the lower tiers sat at 0.6.1 and the enricher's floors named them. The compiler moves because the enricher needs it to: resolver imports just_dna_compiler.resolution_findings, a module RM244 added after v0.7.0 was published, so an enricher wheel floored at >=0.7.0 raised ModuleNotFoundError in a clean venv. Nothing new crossed into just-dna-format, so it does not move.

Three defects, all the same class: a check that could not run answered as a check that passed.

  • AcmgReport.clean returned True for a run that consulted no ACMG SF list. clean was not self.mismatches, and mismatches selects only the verdicts not_listed and denied; a run that obtained no list gives every row unchecked, so the set came back empty. version=None checked=0 clean=True and version=3.3 checked=13 clean=True were both True and meant opposite things, with if report.clean: taking the first for a pass. Reachable on a fresh install — every cache lane empty is exactly that. clean is bool | None now and withholds on both non-comparison arms.
  • alphagenome check with no ALPHAGENOME_API_KEY raised ValidationError where it was supposed to attest nobody asked (RM242). The branch wrote skipped(CHECK, "unchecked"), and "unchecked" is not a member of VALID_VERIFICATION_SKIPS. Repaired to offline, guarded by an AST walk over every literal skipped() reason in the workspace rather than by pinning the one branch.
  • A gate answered yes or no without carrying why it was a no (RM235): a registry outage reported as a broken identifier, and an unreachable EFO reported as clean. One property over four registries.

Also in this release, with no behaviour change: RM244 (the nine resolution findings both tiers emit now speak through one builder — four sentences reworded, listed in that entry), RM246 (the mode ladder is one mechanism), and RM245 (the scenario corpus points at symbols rather than line numbers).

RM247 — the version that generates the Atlas bindings and the version that runs them are one number. [build-system] requires resolves in an environment uv.lock does not constrain, so grpcio-tools>=1.68.0 meant newest on PyPI; protoc stamps its version into the generated bindings, and those raise at import on an older runtime grpcio. With the runtime floored at >=1.68.0 the wheel could not import itself, and because two modules import atlas_client at module scope, enrich, draft and literature — which touch no Atlas code — died with it. The build tool is now pinned at ==1.83.1 with the runtime floored at >=1.83.1 (the pair RM192 measured, and the version uv.lock already resolved, so not a floor bump in RM192's sense), both guards catch RuntimeError as well as ImportError, and a test imports the console entrypoint — which nothing did before, so this used to read as thirty-eight broken test modules rather than as a dead command.

Two guards were corrected, and both corrections are improvements rather than concessions. test_an_intra_workspace_floor_names_the_version_being_cut asserted that a floor equals the depender's own version, which is only the same thing as the truth while all three packages move together. It was written on 2026-08-31 in the commit that repaired a real floor lag, and it took the cadence of the three cuts before it for an invariant — outlawing, retroactively, the enricher-only shape this repository had already shipped as v0.6.3 and v0.6.4. The rule is now a floor names the current version of the tier it imports, which catches the original incident exactly as well (format at 0.7.0, compiler declaring >=0.6.6) and costs nothing when one tier is patched alone. Its sibling test_every_member_carries_the_same_version asserted the cadence directly and is gone.

And that restated rule turned out to be necessary but not sufficient, which only a clean-venv install showed: it compares a declaration against the local tree, where every module exists by construction, so it cannot see an import of a module the published dependency does not carry. test_a_cross_tier_import_exists_in_the_version_its_floor_names reads the dependency's modules off the v<floor> tag and fails on exactly that — the same incident the floor rule was written for (0.7.0's compiler importing just_dna_format.overrides under a >=0.6.6 floor) arriving by a route the rule did not watch.

2026-09-18 — RM245: the corpus points at symbols now, because 53 of its line numbers rotted in one session

Internal corpus only, no code and no dependency. A consumer needs to do nothing.

Every scenario in features/ named its emission site as # source: <path>:<line>. RM244 and RM246 were ordinary code changes — an import block, a shared module, a dataclass field — and afterwards 53 scenarios' lines pointed outside the function they had been written against. None was wrong about the code; every one was wrong about where the code is, and the registry-tagged half had to be realigned four separate times during that work.

# source: now names the file, # anchor: names the symbol, and an anchored scenario carries no line number at all — there is nothing left to decay, since a symbol moves with the thing it names. The symbol is resolved through ast rather than grep, because the name appears in its own docstring and in every caller. The alignment check improved with it: the emission site is inside this symbol's range, with no slack constant, rather than within three lines of this number.

225 of 228 scenarios are anchored. The three line-only ones are the sites no symbol names — two module-level vocab.py entries for @reserved verification-check members, and a module docstring — and a scenario with neither a line nor an anchor is refused rather than pointing at a whole file.

The recovery is worth a line of its own: deriving each anchor from its current line would have written the existing drift in permanently, so each was resolved against the revision its number belonged to — a realigned tagged scenario against today's tree, an untouched structural one against the pre-session commit — and then resolved forward. The conversion went green on the first run over both alignment checks, which is the evidence it was right.

2026-09-18 — RM246: the mode ladder is one mechanism, and the @ladder set grew from four codes to seven

No behaviour change and no message change. Every check escalates in exactly the modes it did before; what moved is how a check says that it escalates.

strict means reproducible artifact, and a check whose severity is the mode said so in three unrelated shapes: (errors if strict else warnings_out) for four codes; an if strict: that moved the same messages into errors and also suppressed the symbolic-allele drop; and ResolutionOutcome.strict_errors, where three codes pair a warning with a different, longer refusal. The third is not a variant of the first — quoting the warning describes a compile that succeeds, and quoting the refusal quotes a sentence best_effort never emits — which RM149's first pass could write down and nothing could enforce.

just_dna_compiler.ladder is now the one mechanism: LadderFinding(warning, refusal=None) and route(findings, strict=…). The pairing case is the general one and a plain escalation is it with refusal=None, which makes the two modes say the same thing a claim rather than an inference from which spelling a function used. ResolutionOutcome derives both warnings and strict_errors from one list of pairs, where two separate fields could each be populated without the other.

What stays two things, deliberately: what the compile does with a refusal. Resolution aborts immediately under a strict resolution: prefix; the allele-membership refusals accumulate because an author wants all of them. That is caller policy, not severity.

The corpus guard is the proof rather than the bookkeeping. @ladder was a judgement written four times; it is now an equality over seven codes, and the three that had been @ladder in prose only are members of the same walked set. The walk also had to be corrected from per-function to per-construction attribution — the first cut credited nine plain warnings as ladder members because they share resolve_from_table with two real ones — and a second guard refuses the retired (… if strict else …) spelling by name, since a walk keyed on one constructor cannot see a check that picks a list instead.

2026-09-18 — the one # DRIFT: the corpus was carrying is repaired

COMPILER.md's validate-by-redundancy table gave the rsid↔coordinate check the severity warning with no qualifier; its own mishap matrix, a thousand lines further down, gave the same check ⚠️ warning / ❌ refuses. Two of our own tables disagreeing about whether a strict compile builds on a contradiction between an authored pair and the injected table — and the matrix is the half the code agrees with, because resolution._verify appends to strict_errors.

The table row now reads warning / error in strict, which is how the two rows directly under it already spelled the same behaviour. The # DRIFT: marker in features/compiler/mode_ladder.feature is gone and the scenario keeps the history in prose: the finding came from writing the severity down as a scenario and noticing there were two places to read it from, which is the corpus's argument in one line.

Marking it rather than fixing it was deliberate — a round that silently repairs what it finds cannot be audited — and features/README.md now says that the marker is for the interval between finding and deciding, and that none stands open.

2026-09-18 — RM244: the nine findings both resolution tiers emit now speak through one builder

No schema change, no dependency change, no new tier. Four warning sentences moved; the two the S61 consumer thread quotes did not.

The compiler's resolution.py and the enricher's resolver.py are twins that deliberately share their warning codes — same finding, same remedy, one channel — and shared none of their words. Nine codes are emitted by both and seven pairs were a different sentence: rsid_unresolved read "not found in resolution table, position remains unset" in one tier and "not in the injected Ensembl snapshot" in the other, and a consumer grepping either one found half the emitters.

just_dna_compiler.resolution_findings is now the nine builders both tiers call. They live in the compiler rather than the format tier because the enricher already imports just_dna_compiler.resolution.genotype_fits — that package is already the shared resolution vocabulary — and because the enricher's deprecated ensembl_cache route goes away at 1.0. They return str rather than a CodedWarning, so the literal CodedWarning("<code>", …) stays at each tier's own emission site.

A None argument omits its clause, and four of the omissions are recorded decisions rather than conveniences. Twice S61 — the enricher's snapshot leg may not say position remains unset, nor not an error, because lookup_variant's live leg has not run and may still place the variant. Once S33 — the compiler's per-rsID expansion accumulator is not ported into the route that is being removed. Once structural — only the positional fill can name which tables a cross-build skip left unjoined. And ambiguous_pick takes the event as a slot, because the compiler reports a label the table already carries while the enricher is at the choice, where that table state does not exist yet.

What moved, declared rather than silent: the compiler's rsid_unresolved, its rsid_without_resolution_label and its positional cross-build skip, and the enricher's resolution_not_injected remedy. Two tests that pinned a sentence now derive it from the builder.

The corpus guard is part of the story rather than a bystander: it caught a broken message-keyed dedupe, 77 # source: anchors an import block had shifted, and a narrowing introduced by the widening RM244 required. # text: now searches beside the emission site rather than instead of it, because a sentence assembled from a shared skeleton and a caller's literals legitimately lives in two files.

2026-09-13 — RM149 second pass: a code is not a text, and the algebra had no values in it

Corpus and guard only, no code change, no dependency. A consumer needs to do nothing.

The first pass closed three registries as equalities and the second pass asked whether that was the same thing as coverage. It was not, in four ways.

A code is not a text. Nine resolution findings are emitted by the compiler's resolution.py and by the enricher's resolver.py, and seven of the nine pairs are a different sentence under the same code — rsid_unresolved reads "not found in resolution table, position remains unset" in one tier and "not in the injected Ensembl snapshot" in the other. The per-code equality accounted for all nine while the corpus held none of the enricher's words for any of them, and a warning's text is what a consumer greps. The equality now keys on (code, tier) — the tier read off the # source: path, not off the feature directory — and features/enricher/resolution_warnings.feature is the nine in the enricher's own words. Exactly one of the nine — rsid_no_hosting_locus — agrees with the compiler verbatim, and one more is a prefix of it; the scenarios say so, because that is a fact about the pair rather than a reason to write nothing.

The house algebra was covered as a mechanism and not as a recorded value. tri_state.feature had Kleene OR, the withhold, classify and restate, and none of the columns that carry a three-valued answer into a published artifact: five of the nine members of VALID_FREQUENCY_STATUS, VALID_RESOLUTION_STATUS and VALID_AUTHORITY_CALL_STATUS appeared nowhere in the corpus — including not_covered, the member the gnomAD Y-PAR probe exists to justify. Six scenarios now cover them.

@ladder was a hand-kept set over a walkable one. The mode ladder's first mechanism is (errors if strict else warnings_out), so the codes reachable from a function carrying that shape are derivable, and they are exactly the four that were tagged. Asserted as the fourth equality, beside a guard that no CodedWarning/ran/skipped is called through an attribute — every AST walk in the file matches an ast.Name callee and would skip one in silence.

Ten # source: lines were imprecise and two were wrong. Reading the 105 structural scenarios — the ones carrying only the line is inside the file — found anchors on a blank line, a bare """, a bare return [, and two that named the wrong subject entirely: a total function cannot decide a three-valued answer pointed at a comment about a length constant when the rule is mitomap.vcep_clin_sig, and an absent input is the unknown arm pointed at the concordance block in vocab.py when the rule is needs_recompile. Not rot — no source file here has changed since the corpus was written — so they were authored that way and nothing could say so.

The corpus is now 13 files and 228 scenarios, 118 registry-tagged and 110 structural, with 182 quoted phrases checked against real string literals. RM149 stays open: still a draft nobody has read, and the second addendum in ROADMAP_0_8.md carries the two scoping questions it answered with a measurement — why refusals without a code have no registry to walk, and why the phrase check's file scope is right.

2026-09-13 — RM149 drafted: a scenario corpus at features/, and the guard that makes it a registry

New top-level directory, no code change, no dependency in any tier. A consumer or downstream repo needs to do nothing; the corpus is internal and does not publish.

RM149 asked for our described scenarios as Gherkin and had stood parked on four questions. Two are answered by the ask itself and are recorded as assumptions rather than decisions: the direction is code → Gherkin (the files describe what the code does, there are no step definitions, and behave/pytest-bdd are not added anywhere), and the corpus lives at the repository root rather than under docs/, which keeps the P3 publishing question open instead of answering it by side effect.

features/ is 12 .feature files and 213 scenarios across the three tiers. Scope is decided by registry rather than by taste, which is the third question: three walked sets are covered in full — VALID_WARNING_CODES (73), VALID_VERIFICATION_CHECKS (26) and VALID_VERIFICATION_SKIPS (8). Round-trip fixed points are deliberately out, on RM149's own argument that prose adds nothing to an equality a test already computes.

New: schema/tests/test_feature_corpus.py, beside the other guards that walk the repository rather than a tier. It asserts set equality over each registry and reports a symmetric difference, so a new warning code without a scenario fails — @registry-completeness. Two of its assertions are what make the corpus worth more than the prose it stands beside, and both were proven by reintroducing the defect:

  • a @code:X scenario's # source: must sit within three lines of X's own CodedWarning call, and a @check:/@skip: one within three of its ran/skipped site;
  • every phrase a step quotes on a Then/And/But step must be a real substring of a real string literal in the module named, because a warning's text is an API and a paraphrase is a different claim.

Two of those were narrower than they read on the first pass and were widened after measuring: the phrase check was keyed on a verb before the quote and covered 46 of 162 phrases, and the @check:/ @skip: alignment did not exist, so eleven # source: lines silently went stale when this session's own two fixes inserted comment lines above them. Both widenings found real defects. Of 213 scenarios, 108 carry a registry tag and are aligned and 105 are structural — measured 2026-09-13 — and the structural ones have no call site to check against, which the roadmap addendum states as a known cost rather than leaving implied.

A message built by one module and coded by another needs a # text: line naming where the words live, so the code is checked against the emission site and the phrase against the module that holds the sentence — the boundary where a code goes missing.

RM149 stays open: this is a draft nobody has read. Its dated addendum in ROADMAP_0_8.md carries the measurement the item was waiting for — four findings in the first pass, two of them filed and fixed as RM242 and RM243 — and says what a review should push on.

2026-09-13 — RM243: four per-command counts in the paragraph that refuses to state a total

Documentation plus two walking guards. No code behaviour changes, nothing to do on any consumer side.

ENRICHER.md's which of these attest paragraph refuses to state a total — a number in prose is a registry nothing iterates — and then gave a count for each of four commands. They summed to 17 of the 24 emitting members of VALID_VERIFICATION_CHECKS: enrich() said six and emits eight (it gained published_refutation with RM170 and evidence_status_currency with RM160), check-identifiers said three and emits five (the two PGS members, RM163), and clinpgx check-labels, litvar coverage and alphagenome check were never in the sentence at all. enrich._verification_records' own docstring carried the same wrong six.

Repaired the way RM218 repaired the identical class inside compiler.py: state the rule, not the figure, and assert the property the figure stood in for. The counts are gone from both places, the attribution lives in the check table where it was already richer per row, and ENRICHER.md names the shape of what went wrong rather than reproducing the four sentences — RM218's own guard caught its repair's prose for exactly that, because a stale figure in quotation marks reads like the rule.

New in test_enricher_doc_registries.py: every emitting member's row has a Where cell whose backticked module.symbol pointers all resolve to real modules, and no attests <number> may appear in the reference. Both proven red by reintroducing the defect. The first guard is documented in the roadmap entry as an invariant that was wrong on the first attempt — Where names the site of the comparison, not the module that writes the record, so reference_allele is compared in sequences and attested by enrich, and the obvious assertion fails on five correct rows.

2026-09-13 — RM242: alphagenome check with no API key raised instead of recording a skip

Enricher only, one literal and one new walking guard. No model, parquet or manifest change, and no consumer surface moves. Nothing to do on any consumer side.

alphagenome_check.check_variant_impact's no-client branch wrote skipped(CHECK, "unchecked", …) and "unchecked" is not a member of VALID_VERIFICATION_SKIPS — it is one of the per-pass spellings that vocabulary replaced, surviving in a single call site. VerificationRecord validates the field, so that branch raised ValidationError rather than attesting that nobody asked. The path is the ordinary one: a local AVI snapshot, a variant whose score straddles the threshold, no --offline, and no ALPHAGENOME_API_KEY. The --offline branch three lines above it attested correctly, so the two absences the module is careful to distinguish behaved completely differently.

Repaired to offline — the check needs egress and the run had none, which is what the member means — with client_absence()'s sentence continuing to say which absence it is in detail, per the vocabulary's own rule that the human sentence travels beside the key and never instead of it. A ninth member for the case was considered and refused: the set exists to collapse six such spellings onto one axis.

New: enricher/tests/test_skip_reasons_are_vocabulary_members.py, an AST walk over every literal skipped(check, "…") in the workspace asserting each reason is a member. Red on the live defect before the fix. Pinning the one branch was refused — it would pass the day another pass writes no_snapshot. Found by the RM149 Gherkin drafting round rather than by a failing run, which is the round's first code finding.

2026-09-13 — Genomi survey: what a genomics runtime annotates, and the seven gaps it exposes

Documentation only — a probe round, no code, no RMn. New: probes/GENOMI_SURVEY.md, a competitor survey of exon-research/genomi at commit 1df4f5b (version 0.1.0, Apache-2.0), read from its source rather than its README.

The short finding is that genomi is a runtime — an agent's local index of one person's genome, plus live fetchers — and not a competing artifact format, so its ~100 operations partition into reference annotation (comparable), sample compute (consumer-side here, by the data-agnostic goal) and agent runtime (out of scope). Only the first bucket is compared. Its declared source roster is 29 entries, but adapter_status splits it 15 wired against 14 record_via_research — an agent reading a page by hand — and its whole authored annotation corpus is thirteen records: ten hardcoded Python dictionaries in capabilities/nutrigenomics/catalog.py, which is a variants.csv written in the wrong language, plus three CYP2C19 star-allele markers in a JSON file. §7 was run, not sketched: their MTHFR record is authored as a spec and validate refuses it — studies.csv line 2 [pmid]: Field required, and studies.csv is missing. Grounding evidence is mandatory when the file is dropped instead — then passes once a real PMID (7647779) is supplied. That refusal is the gate working, not a gap: the folate claim has primary literature behind it and genomi cites the CDC's summary of it, which is a curation shortcut rather than something we owe a slot for.

The findings route to RM188, and all of them are 0.8. A derived sidecar is a row model plus a VALID_SOURCE_LAYERS member plus a parquet plus an enricher pass, so it touches format and compiler and is minor-legal-and-minor-required under P3/P8; the same holds for out_of_scope_claims as a new optional authored column. The one item that could still be a just-dna-enricher patch on 0.7.x is Tier 3's PGxDB, and only if its probe finds something the CPIC/PharmGKB/FDA lanes miss and that reduces to a check over existing tables, RM166's check-labels being the shape. Their record also carries one effect_allele where we carry a row per genotype, and encodes the APOE ε2/ε3/ε4 haplotype as an English sentence in a string field, against reference_examples/apoe_epsilon, which already ships the same two variants as six haplotypes.csv rows and six diplotypes.

Nine gap rows, seven real, five of the seven gene-keyed (the other two are locus-keyed and join on variant_key): Open Targets target–disease and L2G, drug target/mechanism (ChEMBL), pathway membership (Reactome/MSigDB), baseline tissue and cell-type expression (HPA), regulatory-feature overlap (ENCODE cCRE), and perturbation screens (DepMap/ORCS); PGxDB is the small ninth and the 1000 Genomes ancestry panel is charter-blocked. Every candidate prices as a derived sidecar via an enricher pass on the gene_validity.csv template — half cost under the 0.6 amendment — never as an authored table kind, because no author decides those cells. The one authored-layer candidate is genomi's out_of_scope_claims, which is a third thing beside conclusion and negatives: the popular claim a row exists to contradict.

HPO stays refused for the reasons already on record in ENRICHER.md (its licence cannot be established, and genes_to_phenotype.txt is a different grain), and the survey records what has no analogue on their side either — heteroplasmy, repeat alleles, copy number, ACMG SF, MANE, constraint, CIViC, AlphaGenome, the literature pack, licence-as-data, signing, the overlay and the round trip.

2026-09-13 — the table reference says who fills each table, and the derived half gets a registry to say it with

Docs plus a new just_dna_enricher.producers registry — no wire change, no schema change, no behaviour change. The site's per-table nav was one flat alphabetical run of CSVs, and no page said which surface writes the table it describes.

Two groups in the nav, read off the page rather than recomputed. Authored — a person writes them and Machined — an enricher pass writes them. The section a table lands in is literally the string its own identity card prints, so the nav and the page cannot disagree about what a table is.

Two new rows on every card. An authored table gets Drafted by — every DRAFT_PROVIDERS entry targeting it, with its kind, its match_on and a link to the module. A machined one gets Written and checked by — the command as an operator types it (linked to the generated CLI page's own anchor), the module, the sources, and the VALID_VERIFICATION_CHECKS members that pass attests while it writes.

The derived half needed a registry, because it could not be answered at all. A drafter has always declared its table, so who drafts me was already walkable. For a derived table the writer's name lived in a comment on VALID_VERIFICATION_CHECKS, the source was a SourceRow literal inside ten passes, and the CLI help is prose — leaving the generator two options, hand-keep a mapping (@registry-not-a-list, and the exact rot its own docstring refuses) or say nothing. producers.py declares DerivedProducer(table, command, module, sources, checks) and every field is checked against the registry that owns it: tables against _FACT_TABLES, sources against licensing's SourceTerms, checks against the vocabulary, commands against the live Typer tree — the same pinning CacheLane.build_command has, after cache status printed two commands that do not exist.

Two things the guards found while being written. pubmed and europepmc have no SourceTerms, which is @per-article-terms working as designed — declared as SOURCES_WITHOUT_TERMS with a reason each, and asserted as an equality so a third one cannot join them by being typed, rather than the guard being loosened. And the first draft of the authored fallback read every row is written by hand, which sat on licensing.csv directly above every pass that consults a source writes one — two rows of one card contradicting each other. It now says what it knows: no drafting provider targets the table.

A #anchor in TABLES.md is a same-page link there and a dead one on every spliced page, reported by the build at INFO and by no test. The S101 fix added the first one. _rewrite now retargets a #<name>csv to that table's page and refuses anything else, since a heading inside TABLES.md cannot survive the split.

_check asserts both new rows against the registries at build time — a generated page needs an assertion, not a green build, which this file has two recorded incidents of. Watched failing by dropping a drafter from the rendered row.

2026-09-13 — RM235: a gate's verdict carries the reasons it is a no, and RM234 is retrofitted to it

just-dna-enricher only — a new verdict module, two properties retyped, two CLI branches; no parquet, model or manifest field changes. In the tree and in no version anyone can install. Two behaviour changes, both named below.

RM235 as filed was refuted; what survives is S100's shape. The item claimed that an unreachable dbSNP or HGNC is counted as a broken identifier while an unreachable OLS4 reports clean. It was measured on hand-built IdentifierReport objects carrying states the pipeline does not write for an outage — RsidStatus.state is live|merged|absent|withdrawn and has no unchecked member, and check_identifiers never populates report.rsids at all. Run against a port nothing listens on, the ontology leg raises IdentifierUnavailable, and the CLI exits 1 with an unreachable VerificationRecord for all five checks. A registry outage was already a hard refusal carrying a structured reason (@a-disagreement-with-a-document-may-be-in-the-instrument).

What was real: clean is vacuously True when nothing was asked. --strict printed the unreadable table and then exited 0 beneath all identifiers current. A terminal reader saw both halves; a library caller reading the dataclass — which is how this arrived, as an MCP tool — saw only the verdict.

The shape is a result type, not a third truth value. A gate has to choose an exit code, so a build that cannot be certified is a no; the unknown rides beside the verdict rather than inside it. Verdict is a frozen dataclass over a closed code vocabulary, falsy exactly when it carries codes, and if report.clean: is correct across every shape this property has had. That is the one documented exception to the house tri-state, and verdict.py carries the argument.

The set holds errors, not non-answers. offline, tables_unreadable, stale_identifiers and mismatched_assertions are members. not_requested and nothing_to_check are not — a check the caller switched off, or a module with no row a check applies to, is an honest non-answer reported through the denominators that already exist, so --strict --no-traits does not start failing builds for doing as it was told. unreachable is not a member and could not be: no path could set it.

Behaviour change 1 — check-identifiers --strict newly exits 1 when a table carrying identifiers is present and will not parse. It exited 0 with the diagnosis printed above it.

Behaviour change 2 — AcmgReport.clean on nothing_to_check goes from None to a pass. RM234 gave None to both arms of not_consulted; only one is an error. offline means no list was obtained and fails; a list read against a module stating no acmg_sf cell is a module with nothing to disagree about, and check-acmg now says so with checked beside it instead of withholding. The attestation still records a skip on both arms, and the test asserting the two agree was narrowed to the predicate they actually share, because an equality across that split would force one of them to lie.

A prose filter was the defect underneath. unreadable_tables kept everything in *_tables_not_read except the literal "not present", so the row-taking call form's reason counted as eight tables that would not parse — a property whose docstring promised never merely absent ones and had stopped being true on a path the CLI never takes. Found by the existing identifier tests going red against the new gate arm. Both benign reasons are one constant each now, written by the producer and read by the filter; the structural repair (carry the roster's read_errors onto the report) is a field and stays out of a patch.

test_verdict.py walks both call sites' Verdict.of(...) keywords with an AST pass and asserts set equality against the vocabulary, so a member wired nowhere fails rather than describing a format nobody has (@registry-completeness). Watched failing with a deliberately unwired member before it was kept.

2026-09-13 — S101: pgs.csv's page claimed a compile gate, and read a calibration term as a licence one

just-dna-format and the docs — a Field(description=…) and two TABLES.md sections; no behaviour changed, and nothing in the gate moved. In the tree and in no version anyone can install.

TABLES.md's ## pgs.csv said "a module citing an academic-research-only score cannot compile without a declared use … that is the one place this table reaches the compile gate". It reaches no gate at all. Measured on a scaffolded module carrying PGS000001 at research_tier=research_only, through validate_spec(strict=True):

A. no licensing.csv at all                  valid=True
B. commercial_use=false, declared_use empty valid=False   # the gate, and it is licensing.csv's
C. declared_use=non-commercial              valid=True

The conflation under it is the part worth recording. research_tier is a calibration axis, not a licence one — research_only pins as data that a score is a within-reference Z/percentile rather than an ancestry-calibrated absolute risk, which is what pgs.py has said since the field shipped in 0.4. Two unrelated senses of "research only" met in one paragraph. Wiring it to the gate is refused rather than deferred: one field carrying a statistical frame and a licence term is the overloading Principle 5 forbids.

The field's own description was silent in the same way and is the third surface fixed. It read research_only | calibrated (VALID_RESEARCH_TIERS) — a member list naming no axis, which is what let the prose read it as terms (@field-description-is-a-claim). It now names the axis and says it reaches no gate.

Also stated, and settled rather than new: layer names what a source fed, not which table it fed, so every authored table is annotation and a PGS Catalog row reads as annotation-layer content in the gate's error. There is deliberately no per-table member — VALID_SOURCE_LAYERS is a wire vocabulary, and the last request for one was refused on the reporter's own argument (S82, RM147).

Reported by just-module-creator, who had the diagnosis right and asked for the behaviour reading if we disagreed; their own dossier was teaching the correct thing before ours was.

2026-09-12 — RM234: AcmgReport.clean is three-valued, because it was True on a run that compared nothing

just-dna-enricher only — no parquet, model or manifest field changes, and the release class is the maintainer's call. Landed after the 0.7.0 cut, so it is in the tree and in no version anyone can install.

AcmgReport.clean was not self.mismatches, and mismatches selects the verdicts not_listed and denied. A run that obtained no ACMG SF list gives every row the verdict unchecked, so mismatches was empty and clean answered True — a comparison that never happened reporting as one in which everything agreed. version=None checked=0 clean=True and version=3.3 checked=13 clean=True are the two readings a caller could not tell apart, and if report.clean: took the first for a pass.

clean is now bool | None and withholds on both arms of a non-comparison: no list obtained, and a list obtained that no row could be looked up in. It is None rather than False, because answering False would say the module disagrees with a list nobody read.

The attestation was already correct and that shaped the fix. verification_record has always returned a skipped record on both arms, so the persisted record never claimed a pass while the in-memory property did. Rather than testing the same two conditions in two places, a single not_consulted property names the arm and both read it; a test asserts clean is None holds exactly where the record is a skip, across all five arms.

For callers. None is falsy, so if report.clean: was correct before and stays correct — the CLI's green line uses that spelling and its now-redundant and report.version guard is gone. if not report.clean: newly fires on an unconsulted run, which is the intended change. check-acmg --strict gates on mismatches and never on clean, so no offline run newly refuses.

Reported as S100 by just-module-creator, who had already written the caller-side guard and filed it so the next consumer would not have to. One existing test was asserting the defect.

Filed alongside: RM235, open and more serious. The same shape in IdentifierReport.clean, which combines four registries and gates check-identifiers --strict's exit code: an unreachable dbSNP or HGNC is counted as a broken identifier (so a third party's outage fails your build), while an unreachable OLS4 is dropped and reports clean. Not fixed here — its unknown arm is per registry and combines under Kleene, which is a design rather than a one-line change.

2026-09-12 — the docs site gets a user-facing front (no version bump: nothing shipped in a package changed)

Documentation only — no package version moves and nothing a consumer holds changes. The site went up carrying the material that already existed, which is almost entirely development record and tier reference written for whoever maintains that tier; its home page was the README, whose first three sections are Develop, The docs site and Design docs. A reader who has a compiled module and wants to report on a genotype had nowhere to land.

Four new pages, written for a reader rather than an author, sitting in front of SCHEMAS / COMPILER / ENRICHER and citing them: index.md (what this is, and which tier answers your question), GETTING_STARTED.md (install, compile a reference example, verify it), CONSUMING.md (what is in an artifact, the two digests, the join contract's two obligations, the licence axes) and COMPILING.md + ENRICHING.md (the operator paths). They restate no rule — where a guide and a tier reference disagree, the guide is the one that is wrong. Every command in them was run before it was written and the output is quoted as it came back.

The CLI reference is generated (scripts/gen_cli_pages.py, a second gen-files script), walked off the live Typer command trees at build time: of the order of a hundred commands and several hundred flags across the two tools, which is exactly the shape CLAUDE.md already refuses to keep by hand. It reads parameters through param_type_name rather than isinstance, and that is load-bearing — Typer vendors its own copy of click, so TyperOption is not a click.Option and the first version wrote every page with its argument and option tables silently empty.

Two nav changes. navigation.sections opens the top level, because Material folds every group until the reader is inside it and the sidebar therefore read as a site documenting one tier — Reference with nothing under it. And the roadmap ledgers (ROADMAP, ROADMAP_HISTORY, the two deferred files, RM_TOC) leave not_in_nav for a Releases group: they were excluded so a reader was not wading through shipped-roadmap rationale to reach the schema reference, which is true of a sidebar where they sit beside it and false of a group somebody opens deliberately — and every reference page cites RMn by number with nowhere to resolve one.

schema/tests/test_docs_site_nav.py now derives the generated-page prefixes from the gen-files scripts: list rather than one hardcoded path, so a second generator is visible to the partition guard instead of unknown to it.

2026-09-12 — RM233: a red main that was a colour code

Test infrastructure only — no package version moves and nothing a consumer holds changes. main was red at 2001215 on both Python jobs while every local run was green. Two tests match a string inside Typer's rich-rendered output; rich decides colour from the environment, so on a runner --use arrives as three ANSI spans and the assertion reports a command missing a flag it declares. The second failure, a diagnostic wrapped across the error box, was never environment-dependent — CliRunner pins its own width, so no environment variable can widen it. Fixed in the two places the two causes live: TERM=dumb in a root conftest.py, and the box-collapsing helper promoted out of one test module into a shared cli_text fixture. The transferable half is that the helper already existed under a private name, so the second caller that needed it wrote the raw match instead.

2026-09-12 — docs/ renders as a site (no version bump: nothing shipped in a package changed)

docs/ is now a documentation site, built in place. mkdocs.yml at the repository root, a docs dependency group (uv run --group docs properdocs build --strict -f mkdocs.yml), and a CI job that fails on a build warning and uploads the rendered site as a PR artifact. Nothing moved: every path CLAUDE.md, schema/tests/test_doc_links.py and the .claude/ tools cite is unchanged, and no file was added to docs/ — the home page (this repository's README, its links repointed) and the 119 API pages come from the packages' own docstrings and are written into the build by mkdocs-gen-files.

Nothing about a compiled module, a model, a manifest field or a CLI changed, so no package version moves and there is nothing for a consumer to do. Two things are worth knowing anyway:

  • The site's navigation is curated, not filtered. The charter, the three tier references, the cross-tier and using-it documents, the two integration guides and this changelog are in the sidebar; the development records (the roadmap ledgers, the consumer inbox and its archive, the agent notes, the probe rounds, the design threads, the code-first audit) are built, linkable and searchable but absent from it.
  • It is published to https://just-dna.life/just-dna-compiler/ by .github/workflows/pages.yml on every push to main, through the GitHub Actions Pages source — no gh-pages branch, so the repository never carries a copy of its own output. That job holds pages: write and an OIDC token, which is why it is separate from ci.yml's docs job: the PR-side build uploads a plain artifact for a reviewer and must stay runnable on a fork. The URL is a path beneath the organisation's custom apex domain, which a project site inherits rather than serving from github.io, and the path segment is the repository's canonical name — just-dna-compiler, this repository having been renamed, with just-dna-format redirecting. A docs./format. subdomain was weighed and declined: a path costs no DNS and leaves one for every sibling repo. schema/tests/test_docs_site_nav.py asserts the two halves partition docs/ exactly, so a page added to docs/ without a place in one of them fails the suite rather than quietly rendering without a sidebar entry.
  • The builder is ProperDocs rather than mkdocs. MkDocs 1.x upstream is unmaintained and the announced 2.0 removes the plugin system with no migration path; ProperDocs is a 1.x continuation over the same mkdocs.yml, and mkdocs-gen-files and mkdocs-literate-nav already depend on it. The config keeps the mkdocs.yml name so that Zensical — the Material team's successor, Material itself having entered maintenance mode — stays a change of build command away. Zensical was tried, not assumed: it renders all 60 real docs/ pages and mkdocstrings works under it, but it accepts mkdocs-gen-files and then ignores it, so the home page and every API page vanish with no warning. That one plugin is the whole distance to the switch, and docs/AGENT_NOTES.md carries the ten-minute probe and its pass criterion rather than an opinion to re-argue.

Heading anchors are slugified GitHub's way (pymdownx.slugs.slugify), which is what test_doc_links.py has always verified them against; the first build without it reported 540 broken anchors that were not broken.

2026-09-12 (latest) — RM232: a drafted row commits with its licence row too

Cut as 0.7.0 across all three packages, with v0.7.0 now at 2001215 on 2026-09-12 — the maintainer settled the line and retagged at 65b9913, so this entry is inside the cut rather than the first of the next number, and the patch-versus-minor question its first draft left open does not arise. The tag's history, kept because the hashes in INTEGRATION_0_7 § 5 only mean something against it: first at 8b981f6 (2026-09-11), re-cut at 83b1674 when RM231 and a release-record declaration landed after it, re-cut at 65b9913 when RM232 did, and moved once more to 2001215 after S99 was answered, which is where it stands here and on origin. dist/ is built from 65b9913 and stays there: the delta to the tag is five documents and no code, no sdist carries docs/, and a rebuild at 2001215 reproduces all six artifacts byte-identically. So the readiness commit no longer sits past the tag — that sentence held for the first three positions and this entry asserted it for a fourth it could not see, which is why INTEGRATION_0_7 § 5 now carries the position as a measurement rather than a plan.

Published to PyPI on 2026-09-12, 15:49–15:50 UTC, all three packages. The six uploaded sha256 equal the six in INTEGRATION_0_7 § 5, so what is out is what that table describes.

RM231 closed the two-step tail in the eight enrichment passes and had to name five exemptions to close its guard's roster. Two of them were one defect rather than two local ones, and this closes it: a drafter's rows do not go through a writer of its own, they go through the compiler's draft.append_rows / append_partial_rows, which had no before_commit to hand a callback to. So civic_citations.draft_civic_citations and the drafting scaffold's record_draft_provenance recorded the licence row only once the drafted rows were already renamed into place — and a refused merge (a scaffolded module's unreplaced <<REPLACE>> row is enough, which is S98's own trigger) left drafted rows in the author's tables with nothing recording what licensed them. sources.csv is the only file the compile gate reads.

  • Both draft writers take before_commit and thread it to all three of their atomic_writer sites. Optional, defaulted to None, so a caller that passes nothing gets exactly what it got before — see the release-class note above for why that is not by itself the number.
  • drafting.licence_commit is the merge half of record_draft_provenance as a closure factory — one body, handed to every append a drafter makes. It fires per table that actually writes, which is the grain the row needs: hoisting the merge ahead of the appends cannot work, because covered is not knowable until the append is attempted (a run whose rows all differ covers nothing and must write no row), and binding it to only the first or only the last append leaves a table committed unlicensed whichever one turns out to be the no-op. The merge is never-clobber, so N firings record one row.
  • record_draft_provenance keeps its tail call for the run that covered something and wrote nothing — every row already present, so no append reached a writer — along with the stale-label withdrawal and the projection restamp, both of which are answers about the run and not about one table.
  • The pre-flight is inherited, not remembered: building the closure reads the licence table, so a placeholder refuses before the first append rather than after it.
  • The guard is an equality from both ends. Nine committing passes now (civic_citations stops being an exemption), the two closures named as being the callback, and a new walk over every append_* call under just_dna_enricher — 11 of them, all 11 unbound before this change and all 11 bound after.
  • The residual is unchanged and still stated: two files are two renames, so a table rename failing after its callback has returned leaves a licence row for data that never arrived. Conservative, and the OSError names what landed.
  • One behaviour change beyond the fix, and it is intended: the pre-flight runs when the closure is built, and a drafter builds it unconditionally, so draft --dry-run against a module whose licensing.csv does not load now refuses where it used to report. A dry run that hides the refusal the real run will make is the worse of the two, and RM231 made the same trade for the same reason. Pinned by a test.

Suite 4,639 passed, 29 skipped.

Filed before it was fixed, and that is the point. The gap was confirmed and then recorded only as two exemption reasons in a passing test — honest, and invisible to a release check that reads no open RMs, all green. It went into ROADMAP as an open item in its own commit before any code moved.

2026-09-12 — RM231: a licence row is part of its table's commit, in every pass

just-dna-format and just-dna-enricher, inside the uncut 0.7.0. alphagenome expression wrote 12,003 non-commercial rows, then refused to record its licence row on a scaffold's <<REPLACE>> placeholder, and printed FAILED (S98) — and the compile gate, keyed on the licence table alone, then passed the module as unrestricted. Eight passes had the same two-step tail.

  • layout.atomic_writer(before_commit=…) runs a callback after the temp file is fsynced and before the rename. Every fact pass now merges its SourceRow there, so a refused merge removes the temp and a table that fails to serialize never reaches the merge: neither file exists without the other. A rename failing after the callback — two files are two renames — raises an OSError naming what landed.
  • licensing.require_sources_file, the strict read factored out of merge_sources_file, runs before each pass's fetch, so a placeholder row fails in a second rather than after a whole-gene query. The gentle read_sources_file for readers is unchanged.
  • A guard walks every function that records a licence row: eight commit through the seam and pre-read the table; five are named exempt with reasons, two of which are the same gap in the drafting scaffold, RM228's to close.

Refused: writing the licence row first, because a row for a pass that then contributed nothing is a false statement in a published artifact. The if write and result.written gate is unchanged.

The release gate, re-run at the cut, found one undeclared manifest field. expression_effects — the top-level block RM200 added for the tenth derived sidecar — appears (as null) on every module compiled under 0.7.0 and on none under 0.6.6, and the 0.7.0 record did not list it: the block landed after the round's last sweep. Declared as an addition, and the gate reads covers the measurement again: content_signature 0/15, manifest_fields 15/15, parquet_bytes 14/15, parquet_schema 14/15, warnings 3/15, cyp2c9_warfarin_grch37 unmeasured.

2026-09-11 — RM230: a leak an exemption hid, a remedy no flag could reach, and a debt that was not owed

EuropePmcClient.lookup leaked all three failure legs — httpx.HTTPStatusError, httpx.ConnectError and json.JSONDecodeError — into enrich_literature, whose try: has only a finally: and no except. It was hidden because the contract suite exempted the class behind a note about what one method promises: fulltext really does withhold correctly, and that sentence silently covered a sibling nobody had read. Third form of one shape this round — RM208 was a guard inheriting a roster's exemptions, RM225 a comment outliving the code it described. Removing it also showed covered was keyed on the module and could not express "one class here is covered, two are not".

check-identifiers gains --use. PGS rows were built with declared_use="unstated", hardcoded, and the academic_research_only class is commercial_use=False at the annotation layer — exactly where the compile gate reads. So the compile refused and told the operator to "re-run the enricher with a declared use (--use non-commercial)", a flag that did not exist, on a file merge_sources_csv will never clobber. The audit left "is any live score in that class" undetermined, so it was measured: 6 of the first 250 Catalog scores, PGS000013–PGS000017 among them. Reachable, not latent.

And a debt disproved by writing its test first. RM228 recorded that clinvar_draft and pubmind_draft write their licence row unconditionally. They do not — both return early at "nothing matched" before the write. The test that would have proved the bug passes unchanged, so it is kept as a pin on the early return, which is the thing actually holding the rule and which nothing else asserted.

DRAFTING_MATURITY.md is new: the post-RM228 read on what drafting still needs to be a mechanism rather than seven grassroots implementations, written after the migration because four of its findings were invisible until the copies sat side by side.

2026-09-11 — RM228: drafting stops being seven grassroots implementations of one mechanism

Seven *_draft.py providers turn a snapshot into authored rows, and they grew one at a time. By 0.7 each carried its own copy of the same four decisions and the copies had drifted: clinpgx_draft and pgx_draft recorded a release label and never withdrew a stale one, so a module widened from a newer snapshot kept a licence row naming the older release; pubmind_draft imported clinvar_draft._MATCH_ON across modules; civic_draft branched on pydantic's rendered error text; and DRAFT_PROJECTIONS was a hand-kept copy of the drafters' own match_on, its comment pointing at clinvar_draft._MATCH_ON by name.

The obvious repair was refuted by measuring it. Four providers restated the model's skip rule and two derived it, so migrating everyone onto the derived one is the instinct — and it is wrong. authoring_requirements("variants.csv") answers any_of: [['rsid'], ['chrom','start']], which cannot express VariantRow's third clause, ref/alts require chrom and start. The derived implementation therefore accepts {"rsid": "rs1", "alts": "G"} — a partial coordinate the model and a compile both refuse. Constructing the model is the only complete oracle, and it deletes the message parsing for free.

A new drafting.py splits the derived model requirement from a declared source precondition that carries its reason as a field. That distinction is what makes mitomap_draft's clin_sig clause legible: it gates identity on a column the model does not require, and it is correct — a rated_miss carries one by construction — but the reason lived three lines away in a comment, so a real constraint and a misread read identically. This item predicted mitomap would be its one behaviour change; it was not.

The import cycle the refactor hit was the diagnosis rather than an obstacle: draft_digest, stamp_draft_digest and drafted_unchanged had been drafting code sitting in provenance.py, a boundary only holdable while the registry was a copy. Two guards — registry equality and an AST walk — fail on all seven providers pre-migration. Enricher suite: 2108 passed, 25 skipped.

2026-09-11 — RM229: a cache lane declares its size, and a derived lane prices its parents

just-dna-enricher only, inside the uncut 0.7.0. A first-run offer had to du a provisioned box to price a lane (S97), because CacheLane said everything about whether and how and nothing about how much. approx_mb is now on every lane — an order of magnitude in whole megabytes, measured 2026-09-11 and rounded up, None legal and meaning unmeasured — with a test that re-measures it against every snapshot the running machine holds, so a stale number fails a suite instead of a consumer. LaneStatus.size_bytes measures a present lane and cache status prints it. provisioning_closure(lane) walks parents transitively in registry order: mitomap_miss is under a megabyte and its closure is a ClinVar download, which is what a blank box actually pays.

2026-09-11 — RM224: either spelling of a sidecar is a key

just-dna-format only, inside the uncut 0.7.0. layout.SIDECAR_SPELLINGS is keyed on the table key — sources.csv, the spelling the parquet and the manifest keep — and the preferred filename was not a key, so sidecar_spellings("licensing.csv") answered a one-tuple and sidecar_write_path created the preferred copy beside a deprecated one: the collision it documents itself as preventing, reached by a consumer holding a tar member named derived/licensing.csv (S96). sidecar_key(name) reads the map backwards, derived from it, and sidecar_spellings normalises through it, so every helper answers the same for either spelling. Nothing is refused, nothing added or removed; a test walks the map. AGENT_NOTES @sidecar-name-and-place carries the read-side lesson: a helper whose wrong answer is a plausible path fails quietly in both directions.

2026-09-11 — RM207–RM227, a second and blind derivation of the docs found in the code

The docs were re-derived from the code by three agents that had never read them, one per tier, in worktrees with docs/ and CLAUDE.md deleted — the method is now written down as BLIND_REDERIVATION.md rather than living as its first output's preamble. The maintained references gained what they were missing (twenty-six modules, a check row, an AlphaGenome section, the eighteen-source licence roster, all with walking tests). What follows is the other half: RM207–RM227, eighteen shipped and two filed for 1.0 — places where the code was wrong, or where a registry had a hand-kept copy of itself, each reproduced or measured before it was repaired.

RM225 — four stale claims a reader acts on, and one had drifted three times. The schema tier's long-tail candidates, walked item by item instead of batch-closed. actionability is a closed vocabulary and three separate places said otherwise: vocab.py's comment ("the field is not built yet"), base.vocabulary's own docstring listing it among the open sets — inside the helper that defines what closed means — and an earlier reference.py incident. The constant's name carried it too, so ACTIONABILITY_SEED is now VALID_ACTIONABILITY with the old name a working alias. CANONICAL_MT_REFERENCE_SEQUENCES looked like dead code and is not: it is correctly unenforced (an allow-list would refuse future refs), but the refusal beside it spelled NC_012920.1 as a literal, so the set had one job and was not doing it. A pgx.py comment named StudyRow.chrom as validated at two sites, and it has no validator. And is_multi_valued_number's docstring promised a tri-state its signature cannot express — checked every caller first: there is one, it gates whether to raise a warning, so False on unknown is the withhold. The deliverable is the guard, which measures each vocabulary's closedness against its validator and walks the constant names out of base.vocabulary's docstring; run against the pre-fix text it reports the third instance, the one no tool could catch because only the prose was wrong.

RM226 and RM227 filed for 1.0. RM226 collects three warts the code declares in its own comments and no tracker held — the grandfathered content_signature asymmetry, the hardcoded likely_pathogenic/likely_benign, and _freeze_identity's build-blind key — so the 1.0 cleanup has a list rather than a grep. RM227 is RM215's other half: derive_variant_key's coordinate fallback still does not fold allele case, which splits joins rather than dedup and moves a stored cell.

RM215 — one heterozygote had four content identities. ALLELE_PATTERN carries re.IGNORECASE, so a lowercase allele is legal, and the cell is stored verbatim — so A/G, a/G, A/g and a/g are one genotype that hashed four different ways, which for a content-dedup key is the wrong answer. content_signature now upper-cases a cell whose grammar is case-insensitive, and the four collapse to the value every existing module already had. This was filed for 1.0 and the filing was wrong: it sized the change by citing RM81, a parquet retype, where folding at hash time retypes nothing, leaves the authored cell untouched and leaves the round trip alone. That is P3's corrected derivation, which may ship in any release but never silently — and RM36 is the precedent in this very function, having made genome_build feed the hash back in 0.5. The fold is driven by a marker over the four columns whose validator is that grammar, found by probing all 57 models rather than grepping for "allele"; ref/alts are excluded because they are not grammar-checked at all. Measured across the corpus: 536 such cells, none lowercase, so no published signature moves. The variant_key coordinate fallback still splits on case (1:100:a:g,t) and is surfaced rather than fixed, since that one moves a stored cell.

RM223 — the upgrade guide was the one maintained doc nothing walked. A consumer measured four of INTEGRATION_0_7.md's numbers against the installed packages and found all four wrong: 22 parquets where ARTIFACT_PARQUETS holds 23, 72 and then 71 warning codes where the vocabulary holds 73, and 31 models where the authoring reference renders 32. Three had moved when the AlphaGenome round landed expression_effects after the numbers were taken. The document's own § 8 forbids exactly this and §§ 2.2 and 2.3 already said to derive from the constants, so the repair is the guard rather than the four words — the counted-prose test walked SCHEMAS.md and COMPILER.md and stopped. The document now states no registry size at all, and a new guard refuses the shape.

RM207 — the refusal that is fatal in both modes was asked of the wrong key, on the wrong side. resolve_from_table walked the post-expansion rows and looked each one's variant_key up in a table keyed by the key the author wrote. Those strings are equal for a row the fill merely completes and different for a row the table expands — the expansion mints the locus's ga4gh:VA.… id — so a module carrying a withdrawn rsID on an expanded variant compiled clean, success=True. And the check existed only inside resolve_from_table, which only compile_module calls, so where it did fire validate was green in both modes and a plain compile refused: the @validate-refuses-all sequence. Both reproduced on reference_examples/hfe_hemochromatosis with one column changed. The predicates are shared now; the published resolution: / strict resolution: prefixes are unchanged, because that text is API.

RM208 — two clients put the translation inside the retry, so one never retried and one never translated. CrossrefClient.exists caught httpx.HTTPError, the superclass of both types its own @retry matched, and turned a ConnectError into a withhold before tenacity could see it: one upstream request measured where attempt_floor(3) asked for three, and the knob a deployment is meant to raise moved nothing. GwasCatalogClient had the mirror image — the transport leg re-raised bare for the decorator and nothing translated it once the attempts ran out, so associations_for raised a raw httpx.ConnectError while its own docstring said "Both legs are translated". Both got the _request/_get split every other client in the tier already had.

The guard walks the package, not the roster: both clients sat in the contract suite's exempt set, and a guard that iterates a roster inherits its exemptions — the RM101 blind spot one file over. GWAS's exemption is removed rather than re-argued.

RM209 — the publish half walked the root-file registry and the pull half did not. upload.py walks locations.SNAPSHOT_ROOT_FILENAMES; download._provision_snapshot iterated a hand-kept (release.json, LICENSE.txt) pair. So RM198's avi_knots.parquet was published and never pulled — and since the AVI lane stores no PHRED, a pulled lane held scores nobody can rank and alphagenome check refuses it. Two docstrings over that code said the registry was walked; they described the design, and the copy beside it was what ran.

RM210 — one finding, two counts. _cross_check_literature runs on both sides and every message it builds embeds a count; the pre-flight was handed the tables as loaded and the compile the tables after the symbolic drop. pharm_variants.csv is both droppable and a citation site, so a row that is both made the two sentences differ by a number — and a dedup that compares strings kept both. One manifest carried "1 citation(s) … ['99999999']" beside "2 citation(s) … ['29165669', '99999999']" with warnings_summary: {literature_row_uncited: 2} for one finding. The repair is to make the inputs agree rather than to stop re-running: the post-drop view was already computed three lines earlier.

RM211 — @parity-by-check, on the sibling RM93 left behind. _check_gene_metrics_arithmetic is the exact structural analogue of the check RM93 moved into the pre-flight, and stayed compile-only for two releases, so a module whose oe_lof disagreed with obs_lof / exp_lof passed a green validate and warned at compile. Four checks moved — the arithmetic one, the two gene-keyed orphan checks (gene is authored and nothing fills it), and the declared-licence check. Five deliberately did not, and that is recorded rather than left implicit: four are keyed on position or variant_key, so asking them before resolution would report every rsID-only row as an orphan, and _source_checks needs a set that is complete only after the last sidecar is read.

RM212 — the AlphaGenome key in a .env was invisible to the two paths that read it. @credential-where-read has two clauses and only the first was kept: expression._connect — the one function whose docstring cites the rule — read os.environ without load_env(), and nothing else on alphagenome expression's path loads a .env. So the pass refused with is not set while the key sat in the working directory. cli._atlas_client_or_none had the same gap, failing quietly into the knot interval. Both refusals now carry missing_credential_reason, because export VAR= is strictly stronger than never setting it and "not set" sent an operator to the wrong fix. Same incident as the PharmVar lane's, whose comment already says a pre-check answering differently from the code it guards is worse than no pre-check.

RM213 — merge_key collapsed silently on an empty key. Its docstring says a silent () would merge every row into one; it raised for a model declaring no key and returned () for one declaring an empty key, which is MeasureBinRow's base-class default. Latent — the table is authored and every subclass overrides — and fixed because the next kind to inherit the default would find the collapse in a merge pass instead.

RM214 — the allele grammar is case-insensitive and the sort beside it was ASCII, which orders every uppercase letter before every lowercase one. So A/g validated and a/G did not: one unordered pair, two answers. Sorted on str.casefold, stably, so every previously-valid value still validates — a loosening, minor-legal. RM215 is the half it leaves open and is filed for 1.0: the cell is stored verbatim, so those two spellings still hash to two content_signatures, and normalizing allele case moves an identity key.

RM216 + RM217 — the superset sweep the round owed. Take each *_FROM_CODE.md, pull every backticked identifier, keep the ones the package defines at top level, grep them against the maintained doc: 341 real surfaces absent across the three tiers. Most are legitimately delegated — per-field and per-column names belong in the dated snapshot, not in a hand-kept table that loses one (@fieldnames-from-model). Two classes are not.

ENRICHER.md's exception-contract § carried a ten-row table against a tier defining 83 error classes, 51 of them named nowhere — in a section titled what a caller catches. The complete roster is now grouped by what raises them, because the groups are the contract, with the narrowing column asserted to be a real subclass ladder (except order is load-bearing). And eight of vocab's 29 vocabularies were missing from SCHEMAS.md, two of them from every maintained file; the roster there carries count, openness and purpose but deliberately not members.

RM218–RM221 — the long tail, and what it says about where drift lives. The round's second pass worked through the candidates the snapshots recorded but nobody had verified. Four more items came out of it, and three of the four are the same defect class one layer out from where the rule is written down:

  • RM218 — test_counted_prose.py enforces "no number beside a registry" for docs/ and stops there; compiler.py's own module docstring said "twelve" against 23 parquets and _vrs_gap_reason said "six reasons" against 8 arms. Neither number was re-counted: both sentences state the rule, and the guard asserts what each stood in for — for the reasons, that the arms are pairwise distinct, which a count never checked. A 15-name issubset floor over a 39-column schema became an equality, and validate_spec's docstring stopped claiming strict "changes severity only".
  • RM219 — download.py documents that every fetch stages through .part, and one of its own fetches did not, so a repo publishing no release.json got a 0-byte one. That converts nobody said into the description is corrupt.
  • RM220 — --offline had two readings and expression had the wrong one for a licence-gated source, so an injected client fetched under a flag documented as making no egress. The axis the two readings differ on is now written down; gwas keeps its behaviour, deliberately, because the GWAS Catalog is ungated.
  • RM221 — Rich ate [atlas] and [dev] out of three help texts, the extra's size disagreed across five places, "vendored" survived RM196 in two, and VALID_VERIFICATION_CHECKS said enrich writes "six" against eight.
  • RM222 — the CIViC drafter wrote a licence row on a run that drafted nothing (the rule was a comment above a gate that only checked dry_run), and the row carried no dataset because record_source_terms had no parameter for one — so a CIViC-drafted module sat outside the currency check --verify-datasets runs.

What the round says about itself. All three are shapes this repository had already written a rule against — a hand-kept list beside a derivable one, a check on one side of the validate/compile pair, a translation on the wrong side of a retry. That is the same finding the 2026-08-18 round produced with RM93–RM100, and it is the argument for the method rather than for any of the fixes: reading a reference against its code confirms sentences, deriving it again asks which of two documents is wrong.

2026-09-11 — the axis the AVI artifact threw away, as the tenth derived-fact sidecar

RM194 + RM200 shipped together, because they were always one build. expression_effects.csv → expression_effects.parquet: one row per (variant, gene) saying which way a variant moves that gene's predicted expression, how many of AlphaGenome's 371 tissue tracks agree, and how far the variant sits from the gene. Filled by a new just-dna-enricher alphagenome expression.

The design question RM200 could not settle dissolved rather than being decided. It asked whether non-commercial Atlas Output may become a stored value at all, given RM193's position that it enters as a finding and never as one. The answer is a third route: a derived sidecar, which honours that rule literally — nothing here is authored, nothing enters content_signature, and the compile gate still reads sources.csv and nothing else — while carrying the per-gene direction a finding would have had to flatten into prose. Half cost under Principle 9 rather than full.

One scorer of twenty-two, and the other twenty-one are a measured refusal. RNA_SEQ is the only one with a gene axis, the only one whose disagreement across tracks is meaningful rather than flat, and the only one that attributes its own claim to a gene. CAGE's top-5 of 546 tracks carry 2–7% of the effect and the concentration runs backwards to effect size; CHIP_TF's leading factor is whichever TF happens to be measured once.

distance_to_gene is the column that makes the table usable, and null is one of its values. Distal scores run ~10× lower than scores at the gene, so a flat --min-score keeps only the proximal rows while looking like it filtered on effect — the failure RM194 exists to prevent. The distance comes from the MANE lane, and a deployment without it records the absence rather than an interval edge; the manifest publishes without_distance so a whole table of them is visible before a join.

What the schema gained: ExpressionEffectRow, expression_effect_signature, a manifest.expression_effects block, an overrides.csv target keyed (variant_key, gene), and the expression_effect source layer. alphagenome_atlas is a second source name because one (source, layer) key cannot carry two licence classes — the AVI artifact is Permissive Use while this scorer's output is non-commercial, and that one is documented rather than read off a page.

Three bugs only real execution could find, each one a case where every offline fixture agreed with the code that built it: the Atlas wants chr22 on the wire where a module stores 22; a snapshot's summary.parquet lives under data/ rather than at its root; and a lane directory existing is not the same as a lane being built. The first is now fixed by the client owning the spelling, as it already owns the 0-based/1-based conversion.

Verified end to end against the live service and a real MANE lane: 303 rows compiled into a module, compile → reverse → compile byte-identical with artifact.digest, content_signature and expression_effect_signature all stable.

2026-09-11 — the registry's field notes: five things a caching proxy had to work around

just-dna-enricher only, inside the uncut 0.7.0. just-dna-registry filed S91–S95 while building its 0.25 caching-proxy surface over this package — GET /caches, a hosted lookup_*, egress metering — and none of the five blocked them. Four shipped the same day; the fifth is an idea, filed as one.

  • RM204 — the status half of the cache registry is a function (S91). caches.lane_status() returns one LaneStatus per lane and cache status renders it, so a consumer serving the same answer stops re-deriving the loop. A third state, occupied — the place the lane looks is non-empty and holds no snapshot, the target prepare refuses — used to print as absent, which sends an operator to run a pull that will decline. looked_in names which directory the verdict is about, because status reads the override and prepare's refusal reads the default.
  • RM206 — LookupClients has one lazy path (S92). ensure(name, factory) builds a client under the bundle's lock on first use and keeps it; six legs had been building a per-request client and closing it, discarding exactly the pacing state the bundle exists to keep, and two assigned back without a lock. close() walks the derived CLIENT_FIELDS; a lookup_* call given no bundle closes the one it built. The CPIC half stays consumer-side: draft_gene(client=) already shares pacing.
  • RM205 — the hint payload names lanes, not paths (S93). VariantHint.checked holds labels only, the unreadable-snapshot finding interpolates the label, and the new VariantHint.snapshots (label → path, every snapshot opened or tried) is the one field carrying a filesystem path — drop it and audit nothing else.
  • RM203 — PacingGate.spent (S95). Admissions so far, one per wait() that returned, which is one upstream attempt because the clients wait inside their retry loop. A host meters egress from it instead of charging by request shape.
  • S94, a resolver rung that consults a configured peer, is in ROADMAP's 0.7 idea-book, not an RMn: the consumer called it a suggestion and had not designed it, and the one thing the entry keeps is the gate — a peer serving answers from a licence-gated snapshot is neither a fetch nor a read of an operator-built snapshot, and the licence table has no row shape for it (RM27's undesigned axis).

All four shipped items are additive on the enricher: new fields, a new function, a derived tuple, one new rendered line beside two byte-identical ones. One value changed on a read field — checked members that were paths are lane names now — which is what the reporter asked for and what the live leg's ensembl-rest had already set the pattern for.

2026-09-11 — RM201: a declared correction now says which modules it can reach

just-dna-format only, inside the uncut 0.7.0. A registry adopting needs_recompile for its re-publish sweep reported (S90) that a correction declared for gene_metrics.parquet re-published every module in its catalogue, because nothing in DeclaredChange could say the change reaches only modules carrying that block — and it refused, correctly, to read the rule out of target's first segment, since that is a consumer re-deriving our grammar from a field's spelling.

  • DeclaredChange.requires — an optional tuple of dotted manifest paths a module must carry non-null for the change to apply, a necessary condition rather than the exact reach. () means every module and None means unstated; the three scoped corrections in the 0.7.0 record now carry ("gene_validity",) (RM108) and ("gene_metrics",) (RM110), and RM121's stats.genes pair stays None on purpose, because every module carries the field and presence cannot spell "the modules whose lead table named no gene".
  • The predicate is ours. DeclaredChange.reaches(manifest) is three-valued — False is the only answer a consumer acts on — over the pydantic manifest or the mapping json.load returns alike (manifest_carries is the walk); RecompileAnswer.declared_for(manifest) keeps everything but a certain miss, so an unstated reach costs a version number rather than a wrong value left standing.
  • Forced, not defaulted. A test asserts as an equality which shipped corrections leave requires unstated, so a future correction added without deciding its reach fails in the suite. Every required path is walked against the manifest models, the same way manifest_fields already is.

Additive: a new optional field on a package-level record, minor-legal, and a consumer that ignores it keeps the behaviour it has. Docs: SCHEMAS § The release record, INTEGRATION_0_7 § 2.8.

2026-09-11 — RM200: the API has expression direction, and the scorer that looked most useful does not survive being measured

The shipped AVI artifact is one number per variant with the sign discarded. The Atlas serves twenty-two scorers, and nobody had asked which of the other twenty-one a module could take. Measured against the live service.

A correction to our own probe first. ALPHAGENOME_ATLAS.md § 4.7 concludes "there is no expression direction anywhere in the model". That is true of the AVI aggregate, whose eighteen features are all MAX_ABS_*, and false of the API: eight scorers report is_signed=True, RNA_SEQ among them. AVI throws the sign away; the others never had it thrown away.

RNA_SEQ is the one to adopt, and the reason is structural. It is the only scorer with a gene axis — its shape is (genes, tracks) and the gene count varies with the window, 61 at one variant and 48 at another, against (1, N) for all twenty others. So AlphaGenome makes the gene attribution itself, which is exactly what @gene-map-is-another-sources-attribution demands: a gene claim comes from a source's own per-record attribution, never from a span the caller drew. Signed, 371 tissue tracks, 22,631 values at one variant — expression, per gene, per tissue, with direction.

CHIP_TF was the most promising candidate and it is refused. All 1,617 tracks carry a real transcription factor code, so "this variant disrupts CTCF binding" looks sayable. Aggregated by TF over 751 factors, the top three carry 1–3% of the mass, the concentration does not track effect size — the PHRED 84 variant is no more concentrated than the PHRED 1 one — and every leading factor is a singleton track, so the "top TF" is whichever one happens to be measured once. Naming a TF from that would publish a sampling artefact as a mechanism.

And *_ACTIVE is not what its name suggests, which is the finding worth keeping. It is an activity level, not a variant effect: raw assay units, barely moved by swapping the ALT. So it does not describe the variant, it describes the locus — "this position sits in open chromatin in these cell types", which is a claim the format holds nothing like today. Every track carries a nonzero_mean, so a level normalises to fold-over-typical; without that a raw 31 and a raw 0.7 are incomparable, which is why a first pass found all 167 tracks "above threshold" everywhere.

Filed with what still has to be decided: what one authored cell would record, whether this is a check or a column, and that CAGE — 546/546 tracks negative at the extreme variant, the only signal so far that looked like a claim — has not been put to the same test and should be before anyone believes it. All of it is non-commercial Output and needs the second source name alphagenome_atlas.

2026-09-10 — RM197 and RM199: the ALT column a proof removed, and the description a publish sends last

RM197 — wide by position. The AVI lane stores one row per locus with three ALT-score columns and no alt column at all: chrom, pos, ref, alt0, alt1, alt2. 3.371 B/row against 3.882, so 29.7 GB rather than 34.2 — taken now because RM198 publishes the lane, and transfer size binds where local disk never did.

The column disappears because of a proof, not an assumption. Which base each column means is {A,C,G,T} − ref ascending, a function of ref alone, so nothing travels beside the data — and that is only well-defined if every locus really carries all three. Measured over the whole corpus first: across 24 contigs and 8,812,917,339 rows, rows == 3 × 2,937,639,113 positions exactly, pos sorted, alt strictly ascending within a position, no alt equal to ref. Three distinct non-ref bases must be all three. Zero violations.

Two earlier attempts at that check were OOM-killed — group_by(pos) and n_unique both hold billions of keys. Counting boundaries in a sorted column holds nothing and did the genome in 178 seconds; the technique is the transferable part.

It is not a schema break. to_long() recovers (chrom, pos, ref, alt, score) rows from the stored ones with no lookup table, so RM193's join is unchanged — it converts the handful of rows it already filtered to. The round trip is tested against the published source text, not against another derivation of the same parquet.

And the builder re-proves the property on every build, refusing a locus that breaks it. Padding a missing ALT with a null would silently redefine what alt1 means at that locus. That refusal immediately caught a defect of ours: _stream_lines yields whole lines, and a locus is three lines, so a chunk boundary split one — the genome-wide build stopped at chr1:1196920, a locus the file carries in full. Every test passed while that was true, because the 4 MB fixture fits in one 64 MB chunk. The regression test builds the same slice at 64 KB and demands an identical table.

RM199 — release.json goes last. It is what a puller reads to learn which release it holds, so a publish that lands the description and then fails leaves a snapshot that reads as provisioned and is not. Payload first, description second, on every path — and the ordering lives in SNAPSHOT_ROOT_FILENAMES rather than in the publisher, so a --dry-run prints the plan in the order the upload sends it.

Above 5 GB the payload goes through upload_large_folder. That is an atomicity choice rather than a speed one: upload_folder is a single commit with no resumption, right at megabytes and wrong at 32 GB where a failure at 30 GB starts over. The large uploader chunks, retries and resumes, at the cost of not being atomic — which is why there is a threshold instead of always using it.

A declared retirement on that path is refused, naming both rules: RM186 promises the arrival and the departure are one commit, and the large uploader cannot give one. Doing the deletion quietly in a separate commit would leave exactly the window RM186 exists to close.

2026-09-10 — RM198: the AVI lane joins the cache surfaces, and the publisher nearly dropped the file that makes it usable

With RM195 settled, the AVI lane is publishable, and wiring it into cache status / pull / prepare / upload took three fields — the lane registry means no per-lane branch anywhere. It is the first lane that is pullable without being buildable: this tier may not fetch the gated 88.5 GB source, but the 32 GB re-encoding is Permissive-Use output, so an operator who cannot download the artifact can still provision the lane. That is the whole reason to publish it.

The defect wiring it found is the third of its exact shape. plan_reference_snapshot collected data/*.parquet, the sidecar directories, and then a hardcoded pair: release.json and LICENSE.txt. avi_knots.parquet is a root-level sibling of data/ — one small parquet rather than a directory of them — so it would have been dropped silently. The published snapshot would have looked complete, every score present and release.json valid, and nothing on the other side could reconstruct a PHRED, because the artifact deliberately does not store one.

That pair was itself a repair: LICENSE.txt is only in it because a share-alike snapshot had already gone out without the terms it exists to carry. So the names moved to locations.SNAPSHOT_ROOT_FILENAMES and the publisher walks them — the same move CACHE_LANES is, one layer down — and a fourth such file added to a lane and not to the registry now fails a test rather than a publish.

Filed rather than improvised: RM199. publish_reference_snapshot uses upload_folder, which is one atomic commit with no resumption, and every lane published so far is at most ~200 MB. The AVI snapshot is 32 GB over 26 files with a 2.7 GB largest member. upload_large_folder is what the Hub documents for that shape, but it is not atomic — and a publish that lands the parquets then fails before release.json leaves a snapshot that looks provisioned and cannot name its release. Commit ordering is a design decision, not a parameter, so it is written down rather than guessed at.

The dataset repo exists at just-dna-seq/alphagenome_avi with a card and no data. Nothing publishes without an explicit --publish.

2026-09-10 — RM195 and RM196: a pinned page settles the licence, and the repository stops vendoring somebody else's source

Two of the AlphaGenome round's open items closed the same evening, and both are about what a repository should carry as evidence.

RM195 — AVI may be used commercially, and now the repository can show why. The Additional Terms define a "Permissive Use Downloadable Artifact" class and grant it commercial use, then delegate membership to a section of the Atlas website. None of the four pinned terms documents named a single artifact, and that page is a sign-in-gated single-page app. So the most consequential claim about this source shipped as commercial_use=None — unknown is a value, None is never False, and unknown commercial terms warn rather than gate.

The maintainer saved the page. It carries its content as embedded JSON rather than markup, which is why every attempt to fetch it had returned navigation chrome, and it says: "Permissive Use Downloadable artifacts for commercial and non-commercial use" → AVI SNV scores; "Downloadable artifacts for non-commercial use only" → merged splicing scores and feature importance scores. So commercial_use=True, and it independently confirms why the builder refuses the other two artifacts by name. The whole page is pinned (7.1 MB → 1.1 MB gzipped) with a greppable extraction; the test asserts against those bytes rather than a constant.

redistribution=True is a reading, not a clause, and the entry says whose. Prohibition 1 permits sharing "indirectly via … open source release", and an openly published snapshot — carrying the Use restrictions inside it as LICENSE.txt, which restriction 3b requires — was read as one. The three permission axes now rest on three different kinds of ground and sources.csv shows only booleans, so a test records which is documented and which is decided.

RM196 — pip install just-dna-enricher[atlas] used to install two packages and then fail to import. The protos lived in docs/, the bindings were git-ignored, and neither reaches a wheel.

What shipped is not "commit the generated code" or "vendor harder": the repository now carries a pin rather than a copy. A commit id and a sha256 per file; atlas_protos.fetch_protos() downloads from google-deepmind/alphagenome and refuses anything that does not match; enricher/hatch_build.py runs that and protoc at build time. Both trees are git-ignored and deliberately not build-ignored, so the sdist and wheel carry the files while the history does not. A commit id proves what git had; a digest proves what arrived.

Only the enricher moved to hatchling — build backends are declared per package, so just-dna-format and just-dna-compiler are untouched. hatch-protobuf was measured and cannot do the job: it has no import rewriting, and that rewrite is what stops the generated package being named alphagenome and shadowing the real wheel.

Three things only the real build found: upstream's Apache-2.0 LICENSE is at the repository root, not beside the protos; hatchling globs LICEN[CS]E* for the package's own licence metadata, so a fetched LICENSE was added twice and mis-advertised as ours; and force_include duplicates files that live inside the declared package, where artifacts is the right mechanism. Verified from a clean venv: installs with grpcio and protobuf alone, imports, and find_spec("alphagenome") is None.

docs/vendor/ keeps the terms documents and the download page, and its README now says why — those are evidence about licensing, which should be frozen in the repository. Upstream's source code is the opposite kind of file.

2026-09-10 — the AVI lane built genome-wide: 8.8 billion rows, 34.3 GB, and two defects only the real thing found

just-dna-enricher alphagenome build has now been run over the whole artifact. 88.5 GB of tabix TSV becomes 34,291,319,173 bytes over 24 parquets — 3.891 B/row — in 85 minutes on twelve streams.

Every published number cross-checks against a measurement taken independently: 8,812,917,339 rows and 672,931 genuine zeros against the probe's own counts, 49.30% negative against its §1.4, and 41,474 knots against a table a different session built from a different pass over the same bytes — compared knot by knot, not by count, with zero raw values in one and not the other, zero n disagreements and zero phred_lo disagreements. Exactly one knot straddles any integer PHRED threshold from 1 to 50: 0.00076, 676,356 rows, spanning 2.99961 to 3.00027.

Two defects the real artifact found that no fixture could, and both are the same shape — a thing that is correct at every size a test can reach and wrong at the size that matters.

pl.len() is UInt32. Every contig fits it — chr2, the largest, is 721 M rows — and the corpus does not, so summing the per-contig knot tables wrapped 8,812,917,339 to 222,982,747, which is − 2·2³² exactly. Nothing else in the pass would have noticed: the parquets were right, the per-contig counts were right, and the curve would have looked plausible while describing a fortieth of the data. The reconciliation check that requires sum(n) to equal the rows written is what caught it, 65 minutes in.

And alphagenome check joined the module's variants against the snapshot before filtering it. On a fixture that is merely inelegant; on 34 GB it was fatal — the first smoke test was killed by the OOM killer on a module with twelve variants. It now picks parquets by contig from the filename and filters pos inside the scan, where parquet's row-group statistics skip almost everything before a row is decoded. The same query takes 4 seconds.

A third, smaller and caught in the same run: the check's subjects was decided + unanswered, and a variant whose knot straddles the threshold is legitimately in both — so a three-variant module published a denominator of four.

Per-contig knot aggregates are now parked until the merge succeeds. They are built from PHRED, which the artifact deliberately does not store, so they cannot be recovered from the finished snapshot — which is why the overflow cost a full re-read rather than a retry.

2026-09-10 — the artifact is 34.2 GB, and the 43 GB was the tape measure

A correction to the RM191 entries below, and the reason it gets its own heading is that the mistake generalises further than the number does.

The first genome-wide build measured 4.878 B/row → 43.0 GB, against the 34.4 GB the design had projected. That went into RM191's history entry, into a new RM197, and into a sentence saying the design's figure "reproduces in neither layout". All of it was wrong, and the defect was in the thing doing the measuring.

The builder assembled each contig with scan_parquet(...).sink_parquet(...) — the memory-cheap way to concatenate — which fragments the result into one arrow chunk per morsel: 1,432 of them for chr22. Parquet writes at least one row group per chunk, and a sorted pos column only delta-encodes within a row group. Varying nothing else on the same frame:

arrow chunks B/row genome-wide pos alone
1–64 3.871 34.1 GB 1.249
128 3.904 34.4 GB —
1,432 4.892 43.1 GB 2.067

Assembly now reads the chunk files in bounded groups, rechunks each group — one small copy at a time, rather than a 14 GB rechunk of chr1 for no gain — and the same contig writes at 3.882 B/row, 34.2 GB.

Two things worth carrying out of it. Setting row_group_size explicitly is worse at every value tried (4.7–5.1 B/row against 3.88): the default adaptive sizing is what responds to the data, so the obvious knob moves it the wrong way. And what actually governs is rows per chunk, not chunk count — measured on two different slices, the cliff sits near twenty-odd thousand rows per chunk either way, so a cap expressed in chunks is data-dependent. The builder carries both.

The lesson is the one this round keeps finding, pointed the other way: a measurement disagreed with a document, the document was assumed wrong, and the disagreement was in the instrument.

2026-09-10 — the Atlas interval RPC works hand-built, and the reason it looked broken was one enum

AtlasClient.score_interval lands, which is the half of the client RM192 deferred and the thing RM194 was waiting on. It is worth its own entry because what it cost was not what anyone expected.

The blueprint recorded the interval RPC as unreachable from a hand-built client, needing "an x-goog-fieldmask header and 32 bp chunking". Measured against the live service:

  • The field mask is optional. The same interval answers identically without it — asserted as equality of the two answers, not as "both calls succeeded".
  • 32 bp chunking is not a requirement. A 128 bp interval answers in one call; the SDK's 32 bp sub-intervals are how it parallelises, not what the protocol demands.
  • Interval.strand was the whole thing. The Strand enum has no zero member — STRAND_UNSPECIFIED = 0 is the proto3 default — so an Interval that omits strand goes on the wire carrying a value the server rejects, and it rejects it as a bare INVALID_ARGUMENT: Request contains an invalid argument. naming no field at all.
  • A filter is effectively required, which is a size limit rather than a rule: unfiltered, 32 bp comes back as a 43 MB message against a 4 MB receive default.

And an upstream bug, in the pagination. The server returns a next_page_token on an exactly-full final page, and following it fails. A 1,000 bp interval — 3,000 variants, a short last page — correctly omits the token; 1,024 bp is 3,072, exactly six pages of 512, and page six carries one whose seventh request 400s. AIP-158 says an omitted token means there are no further pages, so a client that believes the token crashes on precisely the interval widths that divide evenly. The SDK's loop has the same shape and never trips it, because 32 bp cannot fill a page. score_interval follows the token and stops once the interval it asked for is covered, with an offline test that reproduces the lying token so this cannot come back as a live-only surprise.

Interval.start is 0-based while Variant.position next door is the 1-based VCF one, so the conversion happens at that boundary and callers of this module pass VCF positions throughout.

2026-09-10 — RM193: the Atlas as a resolver, and a check that refuses to run unbounded

just-dna-enricher alphagenome check <spec> cross-checks a module's variants against AlphaGenome's AVI scores. It reports and never repairs, and it exists for the three questions the nine-billion-row snapshot on your own disk provably cannot answer.

Most of it is offline, by design rather than as a fallback. Without --threshold there is no question the local artifact cannot settle, so nothing is asked. With one, the knot table names the candidate set — the variants whose printed score spans a PHRED interval containing the cut — from 466 KB, before a single request is spent. threshold_is_safe() answers the prior question for the same 466 KB, and genome-wide it is yes at every integer threshold from 1 to 50 except 3.

The check refuses to run unbounded. Rebuilding the PHRED column by RPC is 272 days and ~92 million requests at the measured rate, so a caller whose candidate set exceeds --refinement-cap gets a refusal that costs zero calls and names the cheaper answer they already hold.

Three states, kept apart, and none of them a zero. A REF that disagrees with GRCh38 becomes a finding carrying the base the server named — the one thing a local lookup cannot produce, since a file simply misses and a miss looks like an uncovered position. An indel, or a quantile that saturated off the top of the float32 scale, is recorded as no answer exists. A transport failure is recorded as could not ask and deliberately produces no finding at all: a bad minute at Google is not a claim about the caller's data. --offline gives the same third state rather than a silent skip.

It emits its own variant_impact_agreement check rather than a second reference_allele. The Atlas answers the REF question too, but that check belongs to enrich and compares against the reference sequence — letting an Atlas outage write a skip against it would make one registry's availability speak for another's question.

A bug worth naming, because it was silent. VariantRow normalizes chrom and stores 22; AlphaGenome ships chr22. Joining one onto the other matched nothing and raised nothing — every variant came back as "absent from the snapshot", which reads exactly like an artifact that does not cover them. Caught only because the test asserted a positive count rather than the absence of a crash. The conversion now happens at the boundary, and a test asserts both spellings give the same answer rather than that neither is empty.

The [atlas] extra stays optional, and a subprocess test with grpc blocked proves it: the offline half of this check, and the whole CLI, must not acquire a 19 MB dependency by the back door.

2026-09-10 — RM191: AlphaGenome's AVI scores as a cache lane, and a 466 KB curve instead of a 24.7 GB column

just-dna-enricher alphagenome build --input <the file you downloaded> re-encodes AlphaGenome's Variant Impact scores — 8,812,917,339 SNVs — into a fifteenth cache lane.

It never downloads, and that is not the usual inject-only rule. This is the network tier, so it is allowed to fetch; the reason it does not is that the artifact sits behind a sign-in whose eligibility clause bars classes of holder rather than classes of use. So --input is required and there is no default URL: acquisition is the operator's act under their own acceptance.

PHRED is not stored. Measured over every row, PHRED ≥ p keeps 10^(-p/10) of the corpus to four significant figures across four decades — it is an exact within-corpus rank and therefore a function of raw_score. Storing it costs 24.7 GB. A 466 KB knot table beside the data carries the reconstruction curve instead, and carries an interval per printed score rather than a point, because the artifact prints raw_score to four significant digits and PHRED to six. That interval is the useful part: a threshold is unsafe iff it lands inside a knot's span, which is checkable in advance from 466 KB without reading a data row. Genome-wide, exactly one knot straddles any integer threshold between 1 and 50 — 0.00076, 676,356 rows, spanning 2.99961 to 3.00027. Every other threshold is decided.

raw_score is Int32 at a scale of 10⁵, and the losslessness is checked rather than asserted. Both published columns print at most five decimals, so an integer scale is exact where Float32 is both larger and lossy. The builder refuses a value that does not land on the grid instead of rounding it, so if upstream ever widens the column the build stops rather than silently disagreeing with its own source.

One consumer-facing caveat, measured: the exactness is about the decimal. Recovering a float with raw_score_e5 / 1e5 disagrees with float(printed) on 53% of rows — the division rounds a second time and lands one ulp away. Compare in the integer domain (score >= 0.1 is raw_score_e5 >= 10_000) and it never arises.

The artifact is 34.2 GB, which is the figure the design quoted — see the entry above for why this said 43.0 GB for several hours and what the difference was.

Absence stays row-absence. AVI covers about 95% of the assembly and writes 672,931 genuine zeros, so an unscored position has no row while a scored-zero position has a row holding zero.

The snapshot carries the Output Terms' "Use restrictions" section as a LICENSE.txt beside the data, because restriction 3b requires it to travel inside a derivative rather than as a link when the distributor attaches terms of their own — which a module's sources.csv is. release.json records the artifact's own timestamp, which is legally load-bearing rather than provenance hygiene: the Output Terms pin the applicable version to the date the Output was generated.

commercial_use is recorded as unknown, not permitted — RM195. The Additional Terms define a Permissive Use class and grant it commercial use, then delegate membership to a sign-in-gated page that nothing in docs/vendor/ pins. None is never False, and unknown commercial terms warn rather than gate.

2026-09-10 — RM192: the AlphaGenome Atlas client, on two packages

just-dna-enricher gains a working client for the AlphaGenome Atlas — Google DeepMind's precomputed variant scores, served over gRPC — and it costs two packages, not the SDK's forty-seven.

The measurement that decided it. uv add alphagenome resolves to 550 MB and 47 packages (re-measured into a clean venv on 2026-09-11; this entry first said 255 MB and 81 — the count was never run at all, and the size does not reproduce) (anndata, pandas, scipy, zarr, h5py, numcodecs, pyarrow, matplotlib, seaborn, pyfaidx, absl-py, fsspec) against a tier whose entire runtime list is httpx/tenacity/huggingface-hub/typer/ga4gh.vrs. Six of the twenty dependencies that wheel declares are never imported on any scoring path, and atlas.py imports anndata at module level so even the SDK's own import path costs 242 MB. The Atlas score fields are plain bytes and the request filter is an AIP-160 string, so grpcio + protobuf reach every RPC and struct.unpack from the standard library decodes the scores — 19 MB and +2 packages, measured in a clean venv on 2026-09-10 at grpcio 1.83.1 / protobuf 7.36.1.

What a consumer does. pip install just-dna-enricher[atlas], from a checkout, then just-dna-enricher atlas generate once. The bindings are generated from the three Apache-2.0 .proto sources vendored in docs/vendor/alphagenome_protos/ rather than committed, so what the repository carries is reproducible input rather than machine-written output. The extra is checkout-only for now — an installed wheel has neither the sources nor the bindings, which is filed as RM196 rather than papered over: the client's import is guarded and names the command to run.

The error contract is the part worth reading. The service says no in three ways with three different remedies, and the types keep them apart: AtlasUnavailable (retry), AtlasRefMismatch (your REF disagrees with GRCh38 — and the server names the real base, which @va-omits-ref says only this tier can discover), and AtlasNotScored (an indel, or a quantile that saturated off the top of the float32 scale). AtlasNotScored deliberately does not derive from AtlasRefused, so an except AtlasRefused cannot swallow it and record a zero for a variant the service never claimed to have scored.

What is deliberately absent, filed rather than improvised: no tenacity layer over upstream's vendored retry policy, no shared pacing gate, and no interval RPC — that one is RM194's.

The alphagenome extra is deleted, and the comment block in enricher/pyproject.toml that argued the light client "is not declarable here" went with it: commit 1f9a84a had already refuted it by vendoring the sources, and an argument against what the file now declares is worse than no comment at all. Twenty-five tests moved into testpaths with the code and all twenty-five are green, including the four live ones against the real service.

2026-09-10 — PROPOSAL 0.7 PT4: adopting AlphaGenome as RM191–RM195

The first live proposal since PT3 closed on 2026-09-03, and the point at which eleven rounds of measurement in probes/ALPHAGENOME_ATLAS.md become a build. PROPOSAL_0_7_PT4 wins over the roadmap files until its items land. Additive throughout — a cache lane, an optional extra, enricher checks and a drafting provider add no authored column at all, so the round fits the uncut 0.7.0.

  • RM191 — the AVI artifact as a derived cache lane. 88.5 GB of tabix TSV to 34.4 GB of parquet, raw_score as Int32×10⁵ (exactly lossless; Float32 is larger and lossy), and PHRED not stored at all — the 466 KB knot table reconstructs it with zero threshold misclassifications. Operator-built, never fetched: the source is 88.5 GB behind a sign-in.
  • RM192 — the Atlas client on two packages, and the alphagenome extra deleted. 22 MB in a new [atlas] extra against 550 MB and 47 packages for the wheel, six of whose declared dependencies are never imported on any scoring path. Already built and tested as probes/alphagenome_poc/.
  • RM193 — the Atlas as a resolver, not a source. Knot-straddle refinement (the API's raw_score carries ~7 significant digits against the file's 4), REF validation that names the real base, and UNIMPLEMENTED recorded as the third state. Reports, never repairs, and refuses an unbounded refinement.
  • RM194 — gene-scoped subslices, and the ±512 kb horizon. Measured: gene-filtered scores reach +500 kb and vanish at +700 kb. The filter is required rather than optional — unfiltered fails RESOURCE_EXHAUSTED — and at ~50 minutes per gene this is a panel tool, never genome-wide.
  • RM195 — commercial_use=None, not True. The Terms define the Permissive class but delegate membership to a page nothing in docs/vendor/ pins, so the claim that AVI is commercially usable rests on a reading the repository cannot verify. Unknown is a value; it resolves by saving one page.

The round's standing rule is measured-not-inferred, and it is not rhetoric: four of the probe document's own claims were refuted by later measurement, every one a plausible reading of upstream prose that the bytes then contradicted.

2026-09-10 — a 502 KB knot table replaces both the PHRED column and the API refinement

Eleventh round on probes/ALPHAGENOME_ATLAS.md § 4.7.2. No RMn, no adoption.

  • Rebuilding PHRED through the API is 272 days. 8.81 billion variants at the measured 375 variants/s (the interval RPC, 10 workers) is 23.5 M seconds in ~92 million RPCs; single-variant calls would be 45 years, and even the 14.2% of rows on an ambiguous knot is 39 days. It is also the wrong shape of request: prohibition 3 bars republishing the Services, the licence is revocable, and credentials are personal — 92 M calls on a personal key to reconstruct a file that is a download.
  • The raw_score value set saturates, so the ambiguity is enumerable. Four-significant-digit printing makes it a fixed grid: chr22 alone shows 40,204 distinct values, and adding ~400 M more rows across five other regions grows it to 40,888 — 1.7% — while the 2,001 ambiguous knots and the 0.000700 widest span do not move at all.
  • So ship the knots, not the column. (raw_score, phred_lo, phred_hi, n) over ~41,000 rows is 501,592 bytes as parquet (381,432 without counts), and it carries three things at once: the exact reconstruction curve, the per-knot ambiguity interval — so an ambiguous row reports [lo, hi] rather than a point, which is @tri-state-is-the-house-algebra rather than a workaround — and threshold safety decidable by scanning 41,000 rows instead of 8.8 billion.
  • API refinement then shrinks to a last resort: only rows on a knot straddling the threshold actually in use — ~633,000 genome-wide at threshold 3, zero at every other integer threshold from 1 to 50 — and only if a caller insists on a point estimate where the data supports an interval.

2026-09-10 — the precision the file lost is on the API, and neither surface is a superset

Tenth round on probes/ALPHAGENOME_ATLAS.md § 4.7.1. No RMn, no adoption.

  • The 3.56e-4 residual is a publishing artefact of the TSV, not a model limit. The Atlas returns raw_score as a float32 — about seven significant digits against the file's four. Tested at the exact atom that causes every threshold-3 flip: six rows the file prints identically as 0.00076 come back as 0.000758832 … 0.000764675, in an order that makes the file's own PHRED perfectly monotone, and the PHRED derived from them reproduces the published column exactly at all five decimals. The tie was never a tie in the model.
  • Neither surface is a superset of the other, which is @two-surfaces-two-denominators as sharp as it gets. The API wins on raw_score everywhere; the file wins on PHRED above 72.247, where the API's float32 quantile saturates at exactly 1.0 and the file still carries values to 89.451. There is no single surface carrying the artifact at full fidelity.
  • No other download helps. The splicing artifact publishes at the same four significant digits, and the SHAP artifact carries the eighteen model features — also four — and no AVI score column at all. AVI_SCORE_FEATURE_IMPORTANCE, which does sum to the score, exists only as an Atlas scorer.
  • So the shape this suggests is not a choice between surfaces but a bulk build plus targeted API refinement, with §4.8.1's knot rule saying exactly which rows need it — a threshold inside a knot's span — and 161 ms per variant making a few thousand of them a matter of minutes.

2026-09-10 — sixteen bits for PHRED is dominated by not storing it at all

Ninth round on probes/ALPHAGENOME_ATLAS.md § 4.9.1, one question: PHRED rescaled into UInt16 rather than Int32×10⁵. No RMn, no adoption.

  • It loses to the free option on both axes. A rescaled UInt16 costs 16.3 GB genome-wide to deliver max |err| 7.63e-4 and 6,615 threshold flips, while dropping the column entirely and rebuilding it from raw_score costs nothing and delivers 3.56e-4 and 1,218 flips. Twice as inaccurate and five times as flip-prone, for 16 GB. There is no operating point where it wins.
  • And it is not a tuning problem. Beating the reconstruction's 3.56e-4 needs a uniform scale of at least 1,404, putting the top of the 89.45 range at 125,632 — nearly twice what UInt16 holds. 65,536 codes over that range is a mean spacing of 1.37e-3 however they are assigned, so no uniform 16-bit encoding can match a column that is free. A density-weighted non-uniform code could in principle do better in the dense low-PHRED region; it was not measured, and it would still have to beat free.
  • So the roster is two entries, not three: Int32×10⁵ when the rank must be exact, nothing when it need not be. Float32 remains worse than both — larger than UInt16 and lossy where Int32 is not.

2026-09-10 — the reconstruction residual reranks, but only inside ties the rounding made

Eighth round on probes/ALPHAGENOME_ATLAS.md, following the previous round's 3.56e-4 residual to its consequences. No RMn, no adoption.

  • Threshold flips are lumpy, not smooth, and the lumps are diagnosable (§4.8.1). Across a 36 M-row slice, every integer threshold from 1 to 50 flips zero rows except 3, which flips 1,218 at once — and all of them come from a single printed raw_score. 0.00076 occurs 8,443 times on chr22 and spans PHRED 2.99961–3.00027, straddling 3.0, so the curve's mean lands a hair below and the whole atom moves together. That turns an error bar into a decidable rule: a threshold is unsafe iff it falls inside a knot's span, 2,001 of 40,204 knots span more than one PHRED, the widest span is 0.000700, and exactly one integer threshold in 1–50 is unsafe. A "sharp 50" is safe.
  • Rounding does not mend it and makes the worst case worse (§4.8.2). Rounding to 10⁻³ makes 97.4% of rows compare equal, but the maximum residual rises from 3.6e-4 to 1.0e-3, because two values 3.6e-4 apart can land on opposite sides of a grid line. It hides the common case and worsens the rare one, and it cannot recover what the fourth significant digit already lost.
  • Reranking is pervasive but microscopic (§4.8.3): 15.74% of rows change rank, mean shift 143 places, but the largest move anywhere is 3,168 of 36 M — 0.0088 of a percentile. So "the ordering is preserved" is false and "the ordering is usable" is true; the churn is entirely inside ties that raw_score's rounding created.
  • Integer packing is strictly better than Float32 (§4.9). Both columns print at most 5 decimals (measured over all 117 M chr22 rows), so Int32×10⁵ is exactly lossless while Float32 is larger and lossy. Whole rows: 59.2 GB against 67.9 GB, and 34.4 GB keeping raw_score alone. Float64 beats Float32 on raw_score for the same reason — a 5-decimal value has long runs of zero mantissa bits in f64 that f32 rounding scatters into noise. The lesson generalises past this source: where a source publishes fixed decimals, a float is the wrong container.

2026-09-10 — where AVI's bytes go, whether PHRED is droppable, and what a negative means

Seventh round on probes/ALPHAGENOME_ATLAS.md, three measurements answering three questions the size table raised. No RMn, no adoption.

  • The 88.5 GB → 69 GB gap is not container overhead (§4.5). bgzip's block headers are 0.03% and the .tbi is 3.2 MB. Uncompressed the TSV is 310 GB, so DEFLATE already gets 3.5× on it. What parquet cannot beat is the payload: written in isolation the two float columns are 54.7 of 67.9 GB — 80% — while pos is 11.0 GB and ref/alt/chrom together 2.1 GB. Text holds up because the values only ever had 4–6 significant digits and DEFLATE squeezes the rest, where an IEEE float stores a full mantissa of noise the source never had. Quantising raw_score to Int32×10⁵ — its real precision — costs 2.41 B/row against Float32's 3.15, and Float64 beat Float32 for the same reason. An earlier draft of §4.4.3 labelled its float column f32 when it was f64; corrected.
  • PHRED is monotone in raw_score but not recomputable from it exactly (§4.6). Zero negative steps across chr22's 40,204 distinct values, so there is one 1-D empirical curve. Applied to three disjoint regions totalling 59.9 M rows, reconstruction gives mean |Δ| 2.2e-5, max 3.56e-4, and Σ|Δ| of 322–765 against an ε·N budget of ~3e-9 — the machine-epsilon test fails by eleven orders of magnitude. The cause is not float error: the file prints raw_score to four significant digits and PHRED to six, so 2,001 raw values carry up to 68 distinct PHREDs each. It reclassifies zero rows at ≥10 or ≥20, so it is exact for triage and lossy for reporting a rank — and the framing inverts: raw_score is published too coarsely to regenerate PHRED.
  • A negative AVI is low conservation, not down-regulation (§4.7). AVI_SCORE_MODEL_FEATURES and AVI_SCORE_FEATURE_IMPORTANCE are Atlas scorers returning 18 values each, and they are the SHAP file's columns in order — verified by matching feature 0 against the splicing artifact at two loci (4.0252 vs 4.025, 0.0345 vs 0.03451) and against its absence at a third. At both negative variants probed the whole score is one feature, the signed CACTUS_241_WAY at −20.0 and −11.5, with attributions of −1.23 and −0.70 and everything else near zero. The nine regulatory features are all MAX_ABS_* — sign discarded before the model sees them — so no expression direction exists anywhere in the feature set. The axis is benign ↔ damaging. The positive extreme decomposes as a stop-gain (attribution 1.63), a splicing effect (1.29) and AlphaMissense 0.92 (0.83). Also worth noting: ALPHAMISSENSE comes back nan on non-coding variants, so one input distinguishes unmeasured from zero where MERGED_SPLICING does not.

2026-09-10 — AVI measured whole, and its PHRED turns out to be a size dial

Sixth round on probes/ALPHAGENOME_ATLAS.md: the 88.5 GB AVI artifact extracted and passed over in 46 minutes, 24 contigs, 12-way. No RMn, no adoption.

  • AVI is genome-wide where splicing was not: 8,812,917,339 rows over 2,937,639,113 positions, ~95% of the primary assembly against splicing's 42%, and 2.25× the rows. raw_score is signed and 49.30% negative, so an ingest calling abs() discards what half the corpus says, and it writes 672,931 exact zeros where the splicing artifact had none — so 0.0 cannot be an absent sentinel here either.
  • PHRED ≥ p keeps exactly 10^(-p/10) of the corpus — 79.4321% at ≥1, 10.0002% at ≥10, 0.9999% at ≥20, 0.0100% at ≥40, matching the transform to four significant figures across four orders of magnitude. The column is the variant's exact percentile rank among all possible SNVs, so its histogram is a straight line by construction: no shoulder, no natural cut, nothing to discover. What it is is an unusually honest dial — state the budget, read off the threshold.
  • That refutes an inference this document had already drawn. From the upstream FAQ alone it looked as though the rarity axis was already inside the calibrated score, the background being ~300 K common variants at MAF > 0.01. An all-possible-SNV corpus scored against a common-variant background would be visibly shifted; it is not shifted at all. The FAQ describes the API's quantile_score for the recommended scorers, and AVI's published column behaves differently. PHRED is not a rarity comparison and cannot stand in for §5's missing allele-frequency axis.
  • Thresholding on PHRED drops every negative-scored variant. Every row above PHRED 5 is positive; at ≥1, 36.17% are still negative. A rank discards sign, and nothing in the column names warns of it — a design decision disguised as a filter. Negative-scored, not down-regulating: AVI_SCORE is not among the seven scorers the Atlas marks is_signed and its quantile is [0, 1), so the source states no direction and reading one in would be @field-description-is-a-claim.
  • Sizes at each cut, measured on chr22 and scaled: 69.0 GB for everything with both columns as f32, 34.4 GB keeping raw_score alone, 6.7 GB at ≥10, 662 MB at ≥20, 70 MB at ≥30. AVI still fits whole, but only one way: a both-column build needs a cut between ≥1 and ≥5 to reach 40 GB, while raw_score alone holds every row inside the budget with the sign intact — and drops nothing unrecoverable, since PHRED is the rank and recomputes from a complete raw_score column. The information lives in raw_score, whose magnitude histogram is a genuine lognormal-ish hump peaking at 0.032–0.056 with 20.5% of rows.
  • The API saturates where the download does not. calibrated_scores is a float32, so the largest quantile it can carry caps a derived Phred at 72.247 — and the artifact reaches 89.451. At chr22:30339156 C>A (file: PHRED 84.10132) the API returns calibrated = 1.0 exactly and the rank is gone, while raw_score still agrees to five decimals. About 1,300 rows genome-wide, and they are the highest-impact ones, so @two-surfaces-two-denominators applies to AVI after all — narrowly, on the rank and not the magnitude. The PoC client had clamped at the FAQ's Phred 50 and would have rewritten those as fifty; the clamp is gone and saturation now raises.

2026-09-10 — the light Atlas client is declarable after all, and now it is built

Fifth round on probes/ALPHAGENOME_ATLAS.md, after reading the upstream repository rather than the rendered docs. Still no RMn, still nothing imported by a shipped package.

  • §6.2 said the 22 MB client was "not declarable". That was wrong. github.com/google-deepmind/alphagenome is Apache-2.0 and ships the four .proto sources its own wheel generates bindings from (hatch_build.py calling grpc_tools.protoc). The Atlas surface needs three of them — 28 KB — so vendoring those and declaring grpcio + protobuf is a real dependency set. Corrected in place, and the option list went from three shapes to four.
  • docs/probes/alphagenome_poc/ is the blueprint, test-proven. A working client with 23 passing tests, 20 needing no network. An AST walk asserts its third-party imports are exactly {grpc, docs}, which makes the 22 MB figure a property of the code rather than a claim in prose; the bindings regenerate from the committed sources; and a live test asserts the RPC returns what the 88.5 GB AVI artifact contains. Scores decode with struct.unpack — numpy is not needed either. It is outside testpaths on purpose, so the 4,332-test suite is unchanged.
  • One real mistake, caught by writing it. protoc bakes the staged path into every cross-import, so staging at upstream's own alphagenome/protos/ produces a package literally named alphagenome that shadows the real wheel for anyone with both installed. Staging under the full package path fixes it and removes the sys.path insertion at the same time; test_the_generated_bindings_do_not_shadow_the_upstream_package pins it.
  • The error contract is where the house rules land. Three refusals with three different remedies — AtlasUnavailable (retry), AtlasRefMismatch (the caller's REF disagrees with GRCh38, and the server names the real base), AtlasNotScored (an indel: the answer does not exist). The third deliberately does not derive from the second, so except AtlasRefused cannot swallow it and record a zero for a variant nobody scored.
  • The upstream FAQ defines calibrated_scores: the quantile score, an empirical rank against a background of common variants, MAF > 0.01 in any gnomAD v3 population, ~300 K of them, capped at ±0.999990 — so PHRED cannot exceed 50. That matters for §5: the rarity axis is already inside the calibrated score, and pairing PHRED with gnomAD MAF double-counts the same reference set.
  • The merged-splicing discrepancy is not an abs-versus-signed slip. Upstream's own compute_merged_splicing_score takes a signed max; re-running with it changed nothing, so §6.3's finding stands with one candidate cause ruled out rather than assumed away. And the only mention of "motif" in the whole repository is the quick-start's note that contribution scores are what discover them, naming tfmodisco-lite, tangermeme and tomtom — outside tools. §6.5 confirmed.

2026-09-10 — what the API adds over the files is resolution, and motifs are not a dataset

Fourth round on probes/ALPHAGENOME_ATLAS.md, measuring the network surface against the offline copies now that all three artifacts are on disk. No RMn, no adoption.

  • Three surfaces, not two (§6.4). The bulk files, the Atlas service (precomputed) and the model service (computed on demand) differ in a way that matters: indels work on the model API and are UNIMPLEMENTED on Atlas, exactly as they are absent from the files. Atlas also validates REF against the assembly and names the real base in the error, where a file lookup just misses — @va-omits-ref answered by the source rather than by us.
  • The API's addition is resolution, not more scores. One variant unfiltered returns 36,152 values across 22 scorer blocks against the files' 20 columns. The breakdown is the finding: CHIP_TF comes back as 1,617 named tracks where the SHAP file carries one MAX_ABS_CHIP_TF, CHIP_HISTONE 1,116 against one, RNA_SEQ 37 genes × 371 tracks against one. MAX_ABS_ is a lossy aggregate and the name is what the file throws away. Median single-variant latency 161 ms, so it is not a bulk surface and does not pretend to be.
  • Motifs are not a dataset at either surface, and the announced "Motif datasets" block nothing (§6.5). The string "motif" appears nowhere in the SDK and none of the 22 scorers is one. What reads a motif is in-silico mutagenesis, and query_interval over a 200 bp window returns a 600 × 1,617 ISM matrix in 1.6 s with a transcription_factor column naming every track. The files hold a genome-wide ISM matrix already — that is what a score for every SNV is — but with the TF identity and cell type aggregated away. So the motif conclusion holds and its reason does not: the API is the only route because a motif needs per-track resolution, not because a motif dataset sits behind it. Motif work is per-locus and network-bound, which puts it on the enricher side as a check about a variant, never as something a compiled module carries.

2026-09-10 — the Output Terms want the licence inside the artifact, and the dep moved to an extra

Third round on probes/ALPHAGENOME_ATLAS.md. Still no RMn and still no adoption; what changed is that the terms are now readable from the repo and the dependency question has a landed answer.

  • alphagenome is an optional extra, not a core dependency of the enricher. A default install is back to its 45-package closure; just-dna-enricher[alphagenome] adds 36 more (anndata, pandas, scipy, zarr, h5py, matplotlib, seaborn, pyarrow, …). The 22 MB protos-only client measured in §6.2 is not declarable — the generated stubs ship only inside the wheel, so the light path is pip install --no-deps alphagenome grpcio protobuf, a deployment recipe rather than a dependency specifier, and declaring it would mean vendoring the protos. Nothing imports either yet.
  • Four terms documents are now pinned in docs/vendor/ with pdftotext extractions and recorded hashes: the Additional Terms (2026-09-08), the Output Terms of Use (Effective 2025-06-25), the Google Terms of Service (2026-07-30) and the Google APIs Terms of Service (2021-11-09). Two binding documents are still missing, and §2.8 names them — the Generative AI Prohibited Use Policy, and the website section that says which artifacts are Permissive, which is the only statement of the single most consequential claim in the whole probe.
  • The Output Terms add three things the Additional Terms did not say (§2.7). The licence text must travel inside a derivative rather than as a link — restriction 3b makes the "Use restrictions" section an enforceable provision anyone attaching their own terms must carry, which is the shape SNAPSHOT_LICENSE_FILENAME already has for ClinPGx. Google may demand deletion on breach, not only on termination. And the applicable version is pinned to the date the Output was generated, which turns an artifact's timestamp into a legally load-bearing field and is the argument for recording license_sha256 and a dated dataset rather than linking a live page. The document also predates the Atlas by fifteen months and carries no AVI carve-out at all.
  • The SHAP artifact is a 20-column feature table, not an opacity (§1.5): MERGED_SPLICING, nine MAX_ABS_* modality scores, and three third-party features — ALPHAMISSENSE, CACTUS_241_WAY, PHASTCONS_470_WAY — so a module carrying those columns stacks terms rather than inheriting one set. Its MERGED_SPLICING is the splicing artifact rescaled by 3.345 ± 0.043, not copied, so neither file recovers the other exactly and a module must say which it read. And it writes 0.0 where the splicing pipeline never scored a position — 2.42 M of 2.95 M rows in the window measured — so the source itself collapses unmeasured into no-effect, @unreachable-not-absent occurring upstream of us. Its IS_INSERTION/IS_DELETION columns are constant 0 across 8 M sampled rows despite the filename, so the SNV-only finding survives being checked.

2026-09-10 — the AVI artifact is a different corpus, and a 22 MB client reaches all of it

Second round on probes/ALPHAGENOME_ATLAS.md. Still no RMn, still no adoption, and alphagenome>=0.9.0 is deliberately left uncommitted in enricher/pyproject.toml — adding it is the decision §6.2 exists to inform.

  • The sizing in §4 was the splicing artifact and nothing else, which §4.3 now says in its first line rather than its last paragraph. AVI is not a bigger version of it: raw_score is signed, its rows start at chr1:10001 rather than 65,409 (so plausibly genome-wide, ~2.4× the rows), and its PHRED is not a second score — PHRED = −10·log₁₀(1 − calibrated) reproduces the file exactly from the API's calibrated percentile, so a parquet needs two columns and not three.
  • A complete Atlas client costs two packages. The Atlas service (gdmscience.googleapis.com, x-goog-api-key) serves the precomputed scores — 22 scorers, the 283.9 GB SHAP breakdown among them — and the generated protos with grpcio/protobuf reach all of it: the score fields are plain bytes that frombuffer('<f4') decodes and the request filter is an AIP-160 string. Measured as real venvs: 22 MB protos-only, 85 MB with numpy, 242 MB for the SDK's own import path (atlas.py imports anndata at module level, which drags scipy/zarr/h5py), 255 MB for uv add alphagenome — that last figure does not reproduce, and re-measuring it on 2026-09-11 gave 550 MB across 47 packages (probe § 6.5.1). Six declared dependencies — matplotlib, seaborn, pyfaidx, absl-py, fsspec, pyarrow — are never imported on any scoring path. Against a tier whose whole list is httpx/tenacity/huggingface-hub this is the "dependency tiers are sacred" question, not a size one; §6.2 states three shapes (core deps, optional extra, or no client at all) and picks none.
  • The two surfaces agree for AVI and not for splicing — @two-surfaces-two-denominators, measured rather than assumed. The AVI file is the API's float32 printed to five decimals. The splicing file is not reachable from its own documented merge formula (2.81 against a stated 2.735, 1.83 against 2.112), and at chr1:65409 — 10 bp upstream of OR4F5 — all three splice scorers return zero rows at both 128 KB and 1 MB windows while the file scores it 0.003052.
  • The source names its own threshold, which replaces the invented ones: the splicing docs call

    1.0 a substantial effect, and that is 9,761,281 rows, 0.249% of the corpus, ~62 MB as parquet.

2026-09-09 — AlphaGenome Atlas read as an exploration: three artifacts, three licence answers

No code, no RMn, no adoption. Google DeepMind published the AlphaGenome Atlas on 2026-09-08 — precomputed molecular predictions over every possible human SNV — and this is the exploratory read of what it would take to carry any of it. The record is probes/ALPHAGENOME_ATLAS.md; the terms it quotes are pinned in docs/vendor/ as the maintainer's saved PDF plus a pdftotext extraction, so §2's clauses are greppable and hashed rather than re-fetched from a sign-in-gated single-page app.

  • The three bulk artifacts are not one licence, and the one already downloaded is the restricted one. Only AVI SNV scores (88.5 GB) is a "Permissive Use Downloadable Artifact", usable commercially and by commercial organizations; the merged splicing scores (20.6 GB) and the AVI feature-importance/SHAP scores (283.9 GB) are non-commercial-only. The Terms make the split a definition: the AVI Score is carved out of nearly every prohibition and the AVI Score Feature Breakdown is expressly not part of it.
  • Four things the Terms need that SourceRow cannot say (§2.6), the first of which is not a use restriction at all: eligibility is a bar on who may hold the data — "aren't available for any commercial entity, even if conducting non-commercial work" — where every gated source the repo already carries restricts only what may be done. The other three are the no-training-a-similar-model clause, a revocable licence with a delete-and-tell-third-parties termination obligation against P4's frozen digests, and a notice requirement that must travel with a derivative and state the modifications made to it. Two further shapes are existing gotchas rather than gaps: @acquisition-gate-is-not-a-read-gate (eligibility gates the download; commercial use of the downloaded AVI artifact is expressly permitted) and @write-the-sourcerow's (source, layer) key, which one source with two licence classes would collide on.
  • The corpus was measured, not sampled — 3,924,674,451 rows over 1,308,224,817 positions, exactly three ALTs each, chr1–chr22/X/Y and no chrM, covering ~42% of the assembly rather than all of it. The score has no tail toward zero: 86.7% of rows sit between 0.032 and 0.100 and 0.32% fall below 0.01, so discarding "the negligible 90%" is a threshold against a background lump, not a filter of zeros. A first pass over the leading 20 M rows of chr1 would have sized a slice ~55% too large; the whole-file numbers are in §3.
  • Size turned out not to be the constraint. Measured on chr22 and scaled: the entire corpus as parquet (UInt32 at 10⁻⁶, one row per position, three ALT columns — verbatim, since the source states six decimals) is ~11.4 GB with nothing discarded and no digit lost, and a ≥0.1 slice is ~1.1 GB. So the design question is not which slice fits a budget but whether the artifact is a lookup table (absent means unscored) or a finding list (absent means unscored or below threshold, a fresh instance of @unreachable-not-absent). §4.2 states both and picks neither.
  • The rarity axis exists offline and is 6% populated, which is worse than absent. The Ensembl snapshot the enricher already provisions carries MAF/MAC, so no new download is needed — but on chr22 only 6.1% of its SNVs have a value, and after the join only 0.61% of AlphaGenome's ≥0.1 slice has a frequency at all. "Score ≥0.1 and rare" is ~1.6 M rows genome-wide, small because the column is 94% null rather than because the biology says so: MAF absent means unmeasured, never rare, and a build that filters on it answers a different question than the one asked. §5 states the four honest options. Separately, only 9.95% of AlphaGenome's SNVs are observed variants at all, so intersecting changes the corpus kind. And the bulk artifacts are SNV-only, so rare indels are not a slice of them — they exist only through the API, which is per-request, returns 367-track matrices rather than scalars, and carries no AVI carve-out.
  • ALPHAGENOME_API_KEY joins .env.template with the PharmVar-shaped warning: personal under prohibition 7a, never in a module, fixture or snapshot.

2026-09-09 — the pre-cut audit: what a green suite and a passing sweep did not see

Found by a code audit against the house rules rather than by a failure, run after the 2026-09-01 readiness measurement had gone 138 commits stale. Every gate was re-measured first and held (the table in INTEGRATION_0_7 § 5 carries the commit); the items below are what reading the code found behind the green. Each landed as its own commit with the test that pins it.

  • resolution.csv's writer kept its column list by hand, twice (just-dna-enricher, no output change today). enrich._FIELDNAMES was a literal in sync with ResolutionRow by coincidence, paired with a per-column dict literal in _write_resolution_csv — the shape SOURCES_FIELDNAMES had when it lost redistribution, and quieter here because DictWriter raises nothing for a model field the dict simply omits. The next optional column on ResolutionRow would have been dropped by every enrich run, and a fact column among them would have made resolution_signature disagree between a hand-filled and an enricher-filled table. Now list(ResolutionRow.model_fields) and a generic renderer, the idiom every sibling writer already uses; reproduced on the old writer by subclassing the model with one extra field, and pinned by the equality plus a full-row read-back.

  • An overlay spelling one key two ways defeated its own coherence rule (just-dna-format; a check gains reach, no schema change — INTEGRATION_0_7 § 1 lists it as the fourth check that can newly refuse). _canonical_key_cell had repaired the match — member=AFR reaches the afr row — but apply_overrides and update_targets still grouped by the raw spelling, so two spellings were two key groups with one operation each and "one operation per key group" had nothing to refuse. Reproduced: an update under AFR beside a suppress under afr corrected the row and then deleted it with no error, and two updates of faf95 under the two spellings kept the later value with the author's first correction silently absent. Grouping now runs through the model (_key_groups), and a merged group with more than one spelling is refused on a mixed operation or a twice-stated field, naming both spellings and the stored key. Two spellings carrying one operation on different fields stay one coherent group. The pre-flight overlay_coherence_errors is unchanged (it has no table to canonicalize against), which is why the refusal lives in the function both validate and compile call.

  • Three clients leaked on a 200 that is not JSON (just-dna-enricher; no schema change). The contract test drove a 5xx, a transport failure and a 404, and a maintenance page served with a 200 is none of those: raise_for_status() passes it and .json() raises a bare JSONDecodeError. identifiers (OLS4/HGNC) let it reach cli.py's except ValueError arm, which sits before except IdentifierUnavailable, so the run was filed as "rows will not load" and skipped the unreachable attestation; ensembl's GraphQL leg answering HTML never fell through to REST and aborted enrich mid-loop; grch37.variants_at raised out of a three-valued method. All three now translate (IdentifierUnavailable, EnsemblError → the (None, None) withhold, None). The contract test gains the leg over every client, a WITHHOLDING_CLIENTS table pins the withheld value across all three legs for the two clients whose contract is a value rather than an exception, and the discovery guard now keys on owning an httpx.Client rather than on a Client suffix — EnsemblResolver had been invisible to it since RM101. Fourth appearance of @client-exception-contract in AGENT_NOTES.

  • A directory is not a snapshot — three cache-side repairs (just-dna-enricher; no schema change). (1) The derived-lane parent guard judged a parent by is_dir(), and every adapter mkdirs before it downloads, so a ClinVar fetch cut mid-body left an empty out/clinvar/ that the guard accepted; mitomap_miss then ran, its join found no parquet, and the child was the lane reported FAILED — the arm the guard's own docstring forbids. Parents are judged by their lane's resolver now (_snapshot_at), and a supplied path with no payload is reported missing by name rather than silently swapped for the machine's older cache. (2) prepare_lane read resolve() is None as "absent" and renamed its staging directory onto a target that could exist with no payload, so Directory not empty escaped the whole cache prepare with every later lane unattempted; it refuses before the build now, naming cache prune, since provisioning never deletes, and prepare_caches isolates each lane the way cache pull always has. (3) net.stream_to_file removed its .part only on httpx.HTTPError, so a disk that filled mid-body left the partial its docstring promised was gone; removed on every failure now, translated only for the transport.

  • A rejected partial row widened the author's header, and both drafting writers truncated the author's file in place (just-dna-compiler, drafting surface only; no artifact change). append_partial_rows computed what the batch fills over every partial before the loop rejected any, so a column only an invalid row filled — or only an already-present row, or a raw "" — was added to a hand-authored CSV, empty in every row; append_rows had always decided it over the rows it writes. Now decided over the accepted rows, with a blank counting as unfilled, and a dry run reports the extension it would have made rather than the one it was asked for. Separately, both writers opened the author's CSV with open(path, "w") after reading it, so a kill mid-rewrite left a valid short file; both go through atomic_writer now, the append path copying the author's bytes through verbatim, and an AST walk pins it the way the enricher's sidecar guard does.

  • Ten builders wrote release.json in place, and acmg build its whole snapshot CSV (just-dna-enricher; no schema change). The nine spec-dir sidecar writers have been guarded since S66; the builders were not, and the three newest (mane, drug_labels, strchive) had adopted atomic_write_text while the ten older ones kept write_text — the next builder would inherit whichever neighbour it copied. A truncated release.json is at least invalid JSON that read_release degrades to None; a truncated ACMG CSV parses cleanly and is simply short, so verify_acmg_sf would have reported a gene as "not on the list" with nothing failing. Every builder text write goes through atomic_write_text / atomic_writer now, and the guard walks every *_build.py by AST rather than a hand-kept function list.

  • The lookup's refusal map was read with a default that is itself a member (just-dna-enricher, latent; no output change). _advisory read _REFUSAL_BY_COLUMN.get(column, "redundancy_bearing"), so the next advisory column would have been handed the mildest refusal silently, where identity_bearing may be the right one, and every output assertion would have passed against it (@lookup-with-a-default-hides-a-new-member). Indexed strictly now — an unknown column is a KeyError at the call — and a test walks every _advisory call site by AST to assert the advised columns are a subset of the map's keys and the map's values are members of REFUSAL_REASONS. The two entries nothing advises on yet (gene, trait_efo_id, since 0.5.0) are decisions recorded ahead of the lookup that would need them, and stay.

  • The readiness table was re-measured at a6f31f8 (INTEGRATION_0_7 § 5): suite 4311/0, lint clean, corpus 16/16 with every digest identical to the pre-audit compile, sweep gate exit 0, the 0.6.6 client at 15/16 on the same one field — its previous row had rested on "nothing touched the manifest surface", which RM160 had already falsified. dist/ was rebuilt from a detached worktree at that commit: the six artifacts it held were 741ec59 bytes under the 0.7.0 name and lacked twenty enricher modules added since. The tag is still the maintainer's.

2026-09-04 — the allocator reserved a number when you asked it for help

Agent tooling only (.claude/rm-next.py), no package, no schema, no CLI surface. Found while clearing a stale 🔷 reserved row that RM187 had left in RM_TOC.md after the item shipped.

  • Reserving was the default path, so an unknown flag reached it. main() handled --dry-run, --note, --release and --list and let everything else fall through to allocate() — so .claude/rm-next.py --help, a flag the tool never had, claimed RM189, and one more typo while repairing it claimed RM190. Both numbers are spent (ids are never reused) and share one tombstone row in RM_TOC.md that names the cause rather than reading as two withdrawn items.
  • The repair: a KNOWN_FLAGS set, a real --help printing the usage, and exit 2 on anything else. --note's value is excluded from the flag scan, so a note may open with a dash.
  • Pinned by running the version without the guard, the idiom the lock probe in the same file already uses: the unguarded copy reserves on --nonsense, the guarded one refuses and leaves the index byte-identical. Three tests on the argument surface, which had none — the tool had eleven on the lock, the tombstone and the anchor, and nothing on how it reads sys.argv.
  • Recorded in AGENT_NOTES under @an-index-is-not-an-allocator, with the sentinel slip the repair passed through on the way (else -1 made index 0 the exempt slot, so the first argument went unchecked — None is the sentinel that cannot collide with a real index).

2026-09-03 — a download that could not fail politely, and one body for eleven of them

just-dna-enricher only, no schema change. RM187, found by a real failure rather than an audit: NCBI closed the connection 180,927,542 bytes into a 193,427,450-byte ClinVar VCF during RM179's republish.

  • A flaky download was a traceback, not an outcome. download_clinvar_vcf raised httpx's own exception and _rebuild_clinvar catches ClinVarBuildError, so the lane could not report built=False — and it escaped rebuild_lane too, which in a full cache rebuild aborts every lane after the flaky one. Four of eleven builder downloads leaked; the other seven translated.
  • None of the eleven retried, on the largest requests this tier makes, while every live client has had attempt_floor since RM42. They now share net.stream_to_file: atomic through .part, translated at the boundary, retried on TransportError (a status error is not retried — a 404 from a mistyped tag is the same 404 four times over), and restarted from byte zero per attempt.
  • Eleven public signatures are unchanged. Each downloader builds its own return type from the StreamedFile it gets back, so the five existing return shapes all stay.
  • constraint_build gains the error type it never had — the reason its download had nothing to translate into. ClinVarUnavailable, ClinPgxUnavailable and ConstraintUnavailable subclass their lane's error, so an existing except still catches them.
  • Two defects the sweep found: pubmind_build was the one handler of eleven that left its .part behind on failure, and four of them computed a sha256 while streaming and only logged it.
  • Consumer-visible: a failed bulk download now raises the lane's own type rather than an httpx one. If you catch httpx.HTTPError around these builders, catch the lane's error instead — the httpx exception is kept as __cause__.

2026-09-03 — what a publish may delete, and what it may stop describing

just-dna-enricher only, no schema change. Two policy decisions taken with the maintainer after the published-artifact audit, and the code that enforces them. The premise being corrected is this repository's own: @snapshot-accumulates had been read as never delete, and a HuggingFace dataset repo is git-backed — a delete is a commit and a superseded revision still resolves. The risk was never lost bytes; it is that a retired file goes on answering 200 to whoever still asks for it, and that a sweep removes what nobody looked at.

  • A publish refuses to leave a sidecar undescribed (RM185). The general form of RM179: a snapshot's release.json describes every half the artifact carries, so a publish carrying one half replaces the whole description while add-never-delete leaves the other half as bytes nothing describes. OrphanedSidecarError now refuses that publish and names the file and the command that builds the missing half. The guard reads the remote tree, not the remote release.json — by the second bad publish the block was already gone while the sidecar was still there, so a guard interrogating the description would have passed it. --dry-run runs the same check and exits non-zero, because a rehearsal that skips what the publish refuses on is a different operation.
  • cache prune is new, and it is the only unnamed deletion route (RM186). It names two kinds of remote file — one under data/ that the lane's own glob excludes, and one a LayoutShift declares retired — and nothing else: README.md, .gitattributes, release.json, LICENSE.txt and sidecar directories are never candidates. Without --yes it reads, prints each file with its size and why it is nameable, and stops. Against the live repos it finds one candidate (just-dna-seq/clinvar's 159 MB single-file clinvar.parquet, from before the per-chromosome split), six lanes clean, STRchive n/a — its snapshot is one JSON at the repo root, so there is no data/ to be outside of, and "cannot look" is not "found nothing".
  • A layout change now carries its own migration (same item). A change that retires a published file and introduces another declares a LayoutShift: if the new spelling is absent from the repo and the old one is present, upload the new and delete the old. The predicate is over the remote, so it fires once per repo and is a no-op afterwards, and it rides the upload's own commit as delete_patterns — the arrival and the removal are one commit, with the retired name in the commit message. This is the only deletion a publish performs.
  • The per-lane file globs became a registry (same item). cache prune asks what a repo carries that its lane is not made of, which can only be asked by lane, so SNAPSHOT_FILE_GLOBS is now the one place each pattern is spelled and the ensure_* closures read it — the provisioner and the pruner cannot come to disagree about what a snapshot is. Walked by test against the publishable lanes.
  • Operator-visible: one new command (cache prune), and a publish can now refuse where it used to proceed. No parquet column, model field or vocabulary member changed. Nothing was deleted from any published repo by this work.

2026-09-03 — a recompile question with no lower bound, and a registry missing one attribute

Two consumer reports from the same preview build against the uncut 0.7, both answered the day they arrived; no schema, parquet or manifest change in either.

just-dna-enricher. RM184, from S89: CacheLane.env_var names the variable that overrides a lane's location — the one attribute of a lane that was still a string literal inside its resolver, so three consumers were hand-keeping fourteen names. The literals are now locations.<LANE>_CACHE_VAR constants read by both the resolver and the registry, beside CACHE_BASE_VAR, the shared base no lane owns; pinned on behaviour per lane and by an equality over every JUST_DNA_* string the module declares. str, not optional — every lane has one.

just-dna-format. RM183, from S88 — a consumer reading § 2.8 before adopting it.

  • needs_recompile(None, current) answers unknown instead of raising AttributeError. Compilation.compiler_version is str | None, so a manifest stamping nothing is well-formed and a registry walking manifests it did not produce will meet one. None, "" and whitespace now answer alike: every axis None, complete=False, compiled_under=None, span=(None, current).
  • A present but unreadable stamp still raises, and the refusal quotes the whole stamp. just-dna-compiler 0.6.6 (marketplace-server) was refused as '(marketplace-server)'; it is refused as itself now. Absent is unknown, malformed is a caller's bug — two states, not one.
  • Two fields widened: RecompileAnswer.compiled_under and span[0] are str | None, None only where the call used to crash.

2026-09-03 — the published artifacts, audited: two lanes that could not describe themselves

just-dna-enricher only, no schema change. An audit of every artifact this project publishes to HuggingFace — nine snapshot repos and the module repo — pulled each publishable lane into a scratch cache and read what actually arrived. Every lane is current against its source and every published module verifies byte-for-byte against its own manifest; what the audit found was two lanes that could not state what they held, and one of them had been publishing a false description.

  • A ClinVar rebuild builds the citations half too, or reports failed (RM179). A ClinVar snapshot is two halves on two cadences — the VCF and var_citations.txt — which is why build_citations merges a citations block into release.json rather than writing its own file. _rebuild_clinvar built the VCF half only, and release.json is written by that half, so cache rebuild clinvar --publish uploaded a citations-free provenance over a repo whose sidecar it had not replaced. The publisher adds and never deletes, so the sidecar outlived its own description: the published snapshot carried records from 2026-08-29 beside citations from 2026-06-27 and said nothing about it. Both halves now, or built=False with no out_dir — and --publish uploads only on built is True. Operator-visible: a fully offline cache rebuild --only clinvar --source clinvar=<vcf> now reports failed instead of producing that snapshot, because the citations file is a separate ClinVar download. The published repo still needs a citations rebuild off 2026-08-29; this stops the next bad publish, it does not repair the last one.
  • cache status names ClinVar's release instead of printing a blank (RM182). The reporter labelled a lane with release.json's dataset; eleven builders write it and clinvar_build writes clinvar_file_date, so the one lane that refreshes weekly was the only one an operator could not read a release off. CacheLane gains release_label, defaulting to the dataset reader, with ClinVar's set to clinvar_dataset_label — the function clinvar_draft and clinical.tautology_reason already share, so there is one spelling of the label rather than two. The override set is asserted as an equality over the registry. Consumer-visible: one more populated column in cache status output; no parquet column, model field or vocabulary member changed.

2026-09-03 — an overlay's reason was part of its content identity

just-dna-format + just-dna-compiler, no parquet column, no vocabulary member, and no published signature moves. RM180, from S87 — a consumer building a preview against the uncut 0.7, decided with the maintainer before the cut because the window in which it moves nothing is the release.

  • overrides.csv's reason, decided_by and decided_at are outside content_signature. Rewording a reason is a patch, as a README caveat is (S25); changing the value an override writes is still a new identity. They stay in overrides.parquet, manifest.inputs, the verification binding and artifact.digest, which all still move.
  • The mechanism is a field marker, not exclude=True. base.OUTSIDE_CONTENT_IDENTITY, walked by content_identity_exclusions, read only by integrity.content_signature, so model_dump() stays complete for every writer. The stamped-column idiom was probed first and emptied reason in draft._authored_dump and the enricher's overlay writers — a drafted row the model then refuses.
  • Filed, not built: RM181 — a byte digest moving beside intact signatures says something changed and not what, and provenance has no shift tracker of its own. The maintainer's note for the 0.8 digest-coverage review.

2026-09-02 — a cache roster that could not be walked, and the three lanes it lost

just-dna-enricher only, no schema change. RM176, from the maintainer's question: do all the caches we build have a common rebuild endpoint, and does each have download, build and upload? No — and the three gaps were one defect, a roster that was a four-tuple list inside cli.py rather than something a test could walk.

  • cache status reported nine caches on a machine that has twelve. acmg_build, strchive_build and drug_labels_build existed with no roster entry, and cache pull refused all three as unknown names. The roster is caches.CACHE_LANES now, walked against the *_build modules on disk in both directions.
  • The same three had no resolver, so the flagless path never found them. ACMG's fell through to scraping NCBI's page, which serves v3.2 while the built snapshot holds v3.3 — a correctly authored row reported as wrong. The other two skipped themselves with no_reference about a snapshot sitting in the cache directory. Each now resolves through $JUST_DNA_ACMG_CACHE / $JUST_DNA_STRCHIVE_CACHE / $JUST_DNA_DRUG_LABELS_CACHE or the shared base, and the tests assert the call rather than the resolver.
  • Three lanes had the licence to publish and no way to. New: strchive publish (MIT), clinpgx publish-labels (CC BY-SA, its own repo because the two ClinPGx archives do not refresh in lockstep), and ensure_civic_snapshot / ensure_strchive_snapshot / ensure_drug_labels_snapshot. The three repos do not exist on HuggingFace yet — the first publish creates each, and until then cache pull says so.
  • cache rebuild is the one endpoint over eleven builders, with a driver at scripts/rebuild-caches.sh. It calls the same download_*/build_* the per-lane commands call, so those stay and nothing forks. Its outcome is three-valued: ACMG needs an Elsevier workbook, PharmVar a personal key, CIViC a pinned release, Ensembl is built by just-dna-pipelines — all print as not run with a reason, and the exit code counts only real failures. Every lane builds into <base>/<lane>/, never in place over a live cache.
  • publish_reference_snapshot derives its allowlist from the plan. The pattern list and the file list were two statements of one thing, which is how citations/ and LICENSE.txt each went a release printed-in-the-dry-run and dropped-on-upload. plan_reference_snapshot also takes a payload filename now, because two snapshots hold no parquet at all.
  • A repo nobody has created yet is absent, not failed. Five of the snapshot repos have never been published, and both provisioners were letting the transport's own error reach cache pull's blanket handler — so the command that provisions a deployment exited 1 on a fresh machine because a snapshot had never been uploaded. SnapshotNotPublished is its own type now, printed in yellow and not counted; a download that breaks still fails and still exits 1.
  • Two credentials were not being read from the .env, found by an operator asking whether the new endpoint handles one. $PHARMVAR_API_KEY: the rebuild guard read os.environ directly to decide no key configured versus a key that failed, while PharmVarClient loads the file before reading the same variable — so a key living only in a .env was visible to the builder and invisible to the guard, and the lane reported "no key" and never built. $HF_TOKEN: upload._hf_api called get_token(), which reads the real environment and the hub's own token file and neither is a .env, so a publish refused outright. Both load at the point the credential is read now; an exported variable still wins.
  • Absent and exported-empty are two states. load_env uses override=False, so export FOO= is strictly stronger than unset FOO — an empty variable is present and the .env cannot replace it. Both credential refusals name which state they are in and give the matching remedy; the empty one says unset.
  • cache prepare is the command a deployment wants, and the complement of cache pull. It leaves the machine with every cache it can have: pulls what is published, builds the four that are not — PharmVar, PubMind, MANE and ACMG, each unpublished for a recorded reason — so a machine that only pulled is no longer four caches short with the checks reading them silently skipping. The route is a property of the lane, never a flag. A present cache is left alone, so it is idempotent and cheap to re-run; a built lane is staged beside its target and moved across, never written into a live cache directory.
  • A Python counterpart for both loops: caches.prepare_caches() and caches.rebuild_caches(), returning one outcome per lane in registry order. PrepareOutcome.route says which route answered — present / pulled / built — because a deployment auditing its caches has to tell a snapshot it fetched from one it made, and release.json names the release but not the route. caches.CACHE_LANES is the registry itself: read it instead of hard-coding which snapshots exist.
  • cache rebuild --only acmg needs no --source from a checkout. assets/acmg_sf_v*.xlsx travels with the repository and is found by walking up from the working directory; a glob rather than a pinned filename, because the list is versioned and a constant would stop finding it the day v3.4 lands. assets/ stays out of the wheel, so a pip install is unaffected and still supplies its own copy — the workbook is Elsevier supplementary material.
  • cache rebuild --source lane=path is validated up front and expands ~. A mistyped path used to reach the lane's builder as a bare [Errno 2], and acmg is the last lane in the registry, so that arrived after every other lane had already downloaded.
  • Consumer-visible: cache status output gains three rows and names each lane's real build command (it composed f"{name} build", which named two commands that do not exist), and cache pull's exit code no longer reflects an unpublished repo. No parquet column, model field or vocabulary member changed, and builder-only dependencies were already [dev]-only.
  • Every builder now writes under data/repro/<lane>/ when the operator names nowhere, and the rule is derived rather than restated (RM177, just-dna-enricher only, no schema change). "Nothing a command generates goes in the repository root" has been the rule since civic reproduce needed its own .gitignore line — and it was prose, so nine --out defaults drifted past it: civic, clinvar, pubmind, gnomad_constraint, mane, strchive, acmg_sf, mitomap, mitomap_miss, each a bare relative name that dropped a snapshot directory beside pyproject.toml. Four more builders (clinpgx build, clinpgx build-labels, cpic build, pharmvar build) required --out with no default at all, which is how the docs came to tell an operator to write --out ./clinpgx into the root. One locations.repro_out now spells it, and an AST walk over the CLI asserts no --out default is written inline — so the next builder inherits the rule instead of repeating it. cache rebuild keeps data/caches/, named the same way. civic reproduce moved from data/repro/civic to data/repro/civic_reproduce, since civic build now takes the plain name. Callers passing --out are unaffected.
  • The .claude/skills/create-module/ authoring skill is deleted. It was the dogfooding predecessor of just-module-creator's /create-module and stage skills, was never invoked, and had become a third command-surface table to keep in step with --help. Authoring guidance is that plugin's; this repository documents the format. Recover the wording from git history (git show dffa03f:.claude/skills/create-module/SKILL.md) if a rule in it turns out to have lived nowhere else.
  • A failed optional fetch no longer leaves a 0-byte LICENSE.txt in the cache, and a blank licence pins nothing (RM178, just-dna-enricher only, no schema change). HfFileSystem.get opens the destination for writing before it resolves the remote path, so cache pull created a phantom licence file for every snapshot whose repo publishes none — four of the nine published ones (clinvar, gnomad_constraint, cpic, mitomap) — and a re-pull of a repo that had dropped the file truncated a good local copy. Both verified against the live hub. The fetch now stages through .part like every other download in that function. An empty licence file is not a smaller absence: absence leaves license_sha256 null and warns, while an empty file used to pin sha256:e3b0c442…b855, the hash of the empty string — so blank now normalizes to None at the sink (SourceTerms.row), with read_license answering None for a present-but-blank archive member and clinpgx_draft keeping its own check for the warning it owes. Consumer-visible: none — no parquet column, model field or vocabulary member changed. An existing cache's 0-byte LICENSE.txt is inert and is not deleted for you.
  • MITOMAP is two more lanes, and one of them is derived from the other two (RM171, just-dna-enricher only, no schema change). mitomap build cuts the source's published pg_dump into a parquet snapshot — both curated mtDNA variant tables, because the repository's one mtDNA module draws from rtmutation and neither of its variants is in mmutation — and mitomap miss joins it against the ClinVar chrMT parquet to publish the increment MITOMAP carries over the cache, never the rows both already have. draft-panel --source mitomap-miss appends the rated half of that increment to variants.csv and studies.csv. CacheLane gains parents, empty for the twelve lanes that shipped yesterday, and a child whose parents are not on disk is built=None naming which — never False, and never an empty increment, which would be the strongest possible claim about a source derived from a comparison that never ran. The child pins both parents in its release.json, so a ClinVar rebuild without a child rebuild is detectable, and the drafter says so rather than refusing. Nothing hardcodes a diff: the number the item was filed about ("sixteen new expert-panel calls") is derived on every rebuild, and against the ClinVar of the build's own day it is six rather than sixteen — thirteen of the sixteen are deletions MITOMAP writes right-anchored, which need an rCRS base at position-1 that Principle 2 forbids these tiers from fetching, so they are counted as unmintable instead. MITOMAP's confirmation token is never mapped onto clin_sig (the source states it is not an assignment of pathogenicity) and neither is its undocumented [VUS*]; the five documented ClinGen mtDNA VCEP abbreviations became keys in the one shared normalizer, so normalize_clin_sig("LP") and normalize_clin_sig("Likely_pathogenic") are the same answer. MITOMAP_TERMS is CC BY 3.0 written as a floor, with commercial and clinical use stated rather than inferred, so the compile gate does not fire. A drafted row's genotype is a placeholder — MITOMAP's homo/hetero are literature-presence flags rather than a called genotype — so a module drafted from this source does not compile until a human writes those cells, and the draft prints one worklist line per row with the flags the source did publish.
  • Consumer-visible, second half: draft-panel --gene is optional under --source mitomap-miss (the increment is asked for as a whole) and still required for the other three sources; a lane name may be written with a hyphen anywhere one is taken (--only mitomap-miss, --source mitomap-miss) and the declared underscore member is what is stored. cache status gains two rows.

The citations ten CIViC records carry and no dated file can reach (RM160)

All three packages. just-dna-format gains two optional studies.csv columns; the compiler carries them through; the enricher gains a client, a command and a check. Additive throughout — a new optional column and a new verification-check member are both minor-legal, so this fits the uncut 0.7.0.

  • civic citations <spec> recovers what the builder cannot see. RM169 widened the CIViC snapshot as far as a dated file goes, and the wider basis is a VCF — which needs a POS. CIViC publishes no GRCh37 coordinate for a variant it names as a class of event or a legacy notation, so the submitted evidence on those records is on the GraphQL API and nowhere else. Ten records' citations were unreachable; one of them is variant 1955, whose only reachable evidence for the numbering convention its identity turns on is EID 9969 (PMID 12202531, free full text).
  • civic build and civic reproduce are untouched. That is the decision, not a side effect: the published snapshot keeps its byte-reproducibility contract and does not grow. Hashing an API capture as a build input was available and not taken — it keeps the word reproducible while changing what it is reproducible against. The read is one request per variant by construction, which is why it fits a command an author runs and would not fit a builder.
  • StudyRow.confidence / confidence_unit (optional, 0.7.0): how far the citing source stands behind an evidence link, in its own units, unconverted. CIViC's accepted and submitted are the case — a module holding both must not render them as the same row — and there is no house grade for "an editor signed this off". A confidence with no confidence_unit is refused at the model, the way ClinSigAuthorityCallRow already refuses it. studies.parquet gains both columns; a module that fills neither hashes exactly as before.
  • evidence_status_currency joins VALID_VERIFICATION_CHECKS, emitted by enrich. It re-asks CIViC about every citation the module recorded from the API and reports what has moved — a status accepted or rejected since the draft, or a citation added since. Warns in both modes and escalates in neither, because a source re-curating its own evidence is not an authoring error, and it is deliberately not dataset_currency: that one asks which release a table came from, this one asks whether a per-item judgement has moved.
  • The pin is on the SourceRow, at the literature layer. Every drafted row records when the API was asked and on what basis (dataset = civic_api:status=ALL), rather than a timestamp in each conclusion — a moment is not a claim about a variant. (civic, annotation) stays the snapshot drafter's row: a second surface of one source may not claim the lane's slot.
  • Withholding, three ways. Rejected evidence is not drafted (counted, never silently dropped); a paper whose live items disagree about their status gets its confidence withheld rather than picked; a non-PubMed citationId withholds rather than becoming a pmid. --offline records every subject as not-asked and writes nothing — never ran, findings=0.
  • Consumer-visible: two new studies.parquet columns, one new verification.checks member, one new command. Nothing was removed, promoted to required, or retyped.

2026-09-02 — two records that contradicted themselves, and a header a draft could not write into

just-dna-enricher + just-dna-compiler, no schema change. Three defects, each found by a probe that was measuring something else, plus the probe round behind RM170.

  • release.json declared the accepted basis while recording the wider one. RM169 added --submitted and every derived field moved with it — status_basis, status_counts, vcf_evidence, the per-row evidence_status — but notice stayed a literal, so a snapshot built on the wider basis published "every row of which is status 'accepted'" beside status_basis: accepted+submitted and 642 submitted rows. A consumer quoting the notice was quoting a false sentence. Derived from the basis now, with the counts it read.
  • A partial draft into a header that predates a column crashed. append_partial_rows re-rendered the existing rows against the model's full field list and then wrote them under the file's narrower header, so csv.DictWriter raised a bare ValueError: dict contains fields not in fieldnames. draft-repeats into the shipped htt_repeat_expansion example hit it on pmid and measure_tiling. The header now grows by exactly the columns the batch fills, settled before anything is rendered — the rule append_rows already had.
  • CONTRADICTION_CORPORA, the RM170 probe: both corpora measured before either is designed against. No refutation in CIViC that stands against a claim is accepted — so that finding's subject count is 0 on the accepted basis and 3 on the wider one, and a hint that does not state the basis cannot be honest. Four adopted sources publish this shape, three already land it, and STRchive's is dropped at parse. Filed RM174 out of it: a combination-genotype refutation reaches the parquet as two single-variant rows because the row builder stamps the variant's profile over the evidence item's.
  • Two probes landed, and each replaced the entry that asked for it. MITOMAP_STATUS (RM171): status is a two-token grammar, not 29 free strings — base × optional bracket covers 568 of 602 — and both positions are documented, the bracket explicitly as a ClinGen mtDNA VCEP rating. So mapping the bracket is a normalization, not a curation decision; but 120 of the 136 bracketed rows are already in the ClinVar chrMT snapshot, all as reviewed_by_expert_panel, 119 agreeing, leaving 16 new calls. The base half must not be mapped: MITOMAP states it is not an assignment of pathogenicity. Also: the sibling rtmutation holds both variants of the repo's only mtDNA module, and mmutation holds neither. rm170_kleene is RM170's design record.
  • RM174 rewritten on a phase measurement, and RM28 gains its first corpus entry. EID 8721's own description reads "heterozygous compound mutation" — the two variants are in trans, and a haplotype is cis, so HaplotypeRow would assert the opposite of what the source observed. CIViC's profile grammar is boolean and counted: 209 multi-variant of 1,964 — AND 141, OR 72, NOT 1, nested. Both of RM28's surviving arguments (economy in trans, open-world negation) appear there with instances. The stamp defect stays RM174's; the representation is RM28's and stays parked.
  • RM160's shape decided (unbuilt): read SUBMITTED at enrich time. civic build / civic reproduce keep byte-reproducibility and the snapshot does not grow.
  • RM174's stamp half shipped: a row names the profile its evidence item actually belongs to. A combination-genotype claim reaching the builder through the VCF was fanned out into one row per variant, each stamped with that variant's profile — so the parquet stated, once per variant, that a two-variant claim was a single-variant one. evidence_molecular_profile_id and evidence_molecular_profile_name now ride beside the join key on every row (the key has to stay the variant's own profile or the row does not join), a composite is the inequality of the two ids, and release.json carries composite_profile_rows. The name is null on a TSV-sourced row because that file publishes none. The representation question is not touched — 8721 is a claim about two variants in trans, HaplotypeRow is cis, and no brick holds it; that is RM28's and stays parked.
  • RM170 shipped: an authored direction beside a refutation the source published. Does Not Support was already withheld rather than negated — but a variant CIViC supports and also rebuts still got a risk row drafted, and nothing then said the rebuttal existed. contested_variants cannot see it (a refutation enters no camp, so that counter is correctly 0 on every basis). Now draft-panel --source civic names the variants it wrote such a row for, and enrich folds in the new published_refutation check whenever a CIViC snapshot resolves, so a hand-authored module meets the same sign. Two finding codes — refutation_beside_claim, refutation_without_claim — because they are two sentences; they key the record's detail rather than joining the compiler's VALID_WARNING_CODES, which a compile restates as verification_findings_recorded. Warns in both modes, escalates in neither, repairs nothing. Consumers pinning VerificationRecord.check should add the member. The record names its status_basis on every run including the empty one: every such pair in CIViC rests on submitted content, so on the accepted basis the class is empty by construction.
  • RM173 closed, and RM175 opened: the PGx lane reads a retired filename. The entry's premise was replaced twice in one day. First: clinicalVariants.zip is not a third source — 96.3% of its 5,190 rows are already in the archive the lane reads, and its type column is Phenotype Category with a different separator. Then the maintainer's investigation (CLINPGX_ARCHIVES) replaced the 13-month gap that finding ended on: the 15-column table was renamed to summaryAnnotations.zip when PharmGKB became ClinPGx on 2025-07-29, and clinicalAnnotations.zip is a frozen S3 object last written 24 days before that post, on no downloads page, still answering 200 — and, until the rebuild two bullets down, this lane's default. So every annotations.parquet ever built here came out of the database as it stood 14 months ago. RM175 is the rebuild: four member renames, one id column, no vocabulary or model change, but 8 rows change Level of Evidence and every URL rehosts, so the digest moves. A retired filename that still 200s is indistinguishable from a live one at the HTTP layer, which is the durable lesson; and a no-JS fetch of any clinpgx.org page is the Javascript Is Disabled! shell, so it is no evidence about what the source lists.
  • RM175 shipped the same day: the PGx lane builds from summaryAnnotations.zip now, and refuses the retired archive by name. The default URL, two member names and the id column moved (summary_annotations.tsv, summary_ann_alleles.tsv, Summary Annotation ID) with no vocabulary member, no model field and no parquet column changed — the other fourteen columns and Phenotype Category's values are identical. The guard is the item: an archive carrying the old member names parses perfectly and yields a plausible fourteen-month-old parquet, so the builder reads the member names first and refuses the retired spelling with the rename, its date and the URL to use instead; a third arm answers for an archive that is neither. Both spellings live in one table the reader takes its names from, and the retired one is returned by nothing, so no path through the module can read a 2025 archive. The data moved and the digest with it: 16,087 → 16,117 snapshot rows over 5,190 annotations, 22 (annotation, genotype) keys gone and 52 new, 30 rows changing evidence_level, 120 drugs, and every shared row rehosting its URL — so a module drafted from this lane can see an evidence level move under it, which is the check working. The builder docstring's "4,618 of 5,113" is gone from all five live files, restated as a relationship rather than swapped for a fresh count. Left unbuilt on purpose: the three currency canaries the entry listed, each of which is its own design — nothing here would notice summaryAnnotations.zip itself going quiet.
  • MITOMAP's terms are read (RM171): CC BY 3.0, commercial use and redistribution permitted, attribution required. Read from a Wayback capture because the live page is behind a Cloudflare interstitial; the CC BY-NC a search surfaces is the article's licence, not the database's.

2026-09-01 — the source-adoption round closes: regulator drug labels (RM166)

just-dna-enricher, plus one VALID_VERIFICATION_CHECKS member. Last of the five items PROPOSAL_0_7_PT2 decided, and with it the whole 2026-09-01 source-adoption round has landed inside the uncut 0.7.0: RM163, RM165, RM166, RM167 and RM168 built; RM164 parked on a measured negative, spinning off RM171.

  • A drugLabels.zip builder beside clinpgx_build — same cache, same payload-read LICENSE.txt, and its own release.json from its own CREATED_*.txt. ClinPGx's archives do not refresh in lockstep; relationships.zip was once a year newer than clinicalAnnotations.zip.
  • A regulator-label cross-check joining at two tiers. Star-allele where the file supplies one, gene otherwise, and the tier is part of the finding — a gene-level agreement and an allele-level agreement are different claims and a consumer must be able to tell them apart.
  • It is five regulators: FDA, Health Canada, EMA, Swissmedic, PMDA. The surface is named for the labels rather than for any agency, so adding a sixth is data rather than a rename.
  • A blank Testing Level is unknown and withholds — about a third of the file states none, and reading that as No Clinical PGx would manufacture a negative regulatory claim.
  • New verification check regulator_label_agreement. Consumers pinning VerificationRecord.check should add it; it warns in both modes and never escalates under --strict.
  • What did not ship, and closed instead: the PGx lane gains no member outside its licence class by this route. ClinPGx is the same CC BY-SA + no-sale gate, and the FDA's own association table is 126 rows of HTML with no stated terms. Diversifying that lane needs its own item, choosing candidates for their terms first.
  • Noticed and filed rather than built: ClinPGx publishes at least twelve archives and this tier reads two of them.

2026-09-01 — the source-adoption round: literature coverage, and the tier that answered (RM167)

just-dna-enricher, plus one VALID_VERIFICATION_CHECKS member. Writes no authored row at all. Fourth of the five items PROPOSAL_0_7_PT2 decided.

  • just-dna-enricher litvar coverage asks LitVar2 which papers name a module's alleles, and records which tier could answer: allele-resolved, position-only, absent, or unchecked. The distinction is the point — APOE rs429358's position node carries 3,945 papers and its allele node 328, so an answer that did not name its tier would understate that locus twelvefold.
  • It writes no row. A PMID list per variant is not a table kind, so the pass reports and attests and changes nothing in the spec directory. That was the entry's pre-authorised outcome.
  • New verification check literature_coverage. Consumers pinning VerificationRecord.check should add it.
  • The bound ships with it, in the lane's own documentation: LitVar answers which papers discuss an already-identified allele, and does not answer which allele a name meant. Measured against the two CIViC legacy insertions, it returns no node for any of their four candidate alleles.
  • clingen_allele.AlleleIdentity.unanchored is now populated on a resolved result too. It was computed and then discarded whenever the registry also served an rs-number, so a caller wanting the allele rather than an identity got nothing — every PALB2 indel read as incomparable. A consumer reading unanchored only under outcome == "needs_anchor" is unaffected.
  • Terms: NCBI publishes a policy rather than a licence, so every gating axis is None. A module carrying LitVar-derived findings records unknown terms rather than permissive ones.

2026-09-01 — the source-adoption round: STRchive, checked and drafted by column (RM165)

just-dna-enricher, plus one VALID_VERIFICATION_CHECKS member. No authored column, no parquet change, just-dna-compiler untouched. Third of the five items PROPOSAL_0_7_PT2 decided.

  • repeat_alleles.csv gains a source, split by column. just-dna-enricher check-repeat-bands compares an authored band table against STRchive's and reports; just-dna-enricher draft-repeats drafts the identity half. New STRCHIVE_TERMS (MIT — the first candidate in this round with terms that are both established and permissive) and a strchive snapshot builder.
  • The bands are checked and never drafted, for a measured reason. STRchive reproduces htt_repeat_expansion's first two bands exactly, and gives FMR1 one 45–200 band where the module has 45–54 and 55–200 — losing 55, the premutation threshold. The finding names the missing boundary rather than reporting that two tables differ.
  • pathogenic_max is never written as measure_max. A catalogue maximum is an observation, not a clinical bound; imported as one, an allele above it would match no bin at all and nothing would say so. It is reported as its own finding kind.
  • New verification check repeat_band_agreement. Consumers pinning VerificationRecord.check should add it. It warns in both modes and never escalates under --strict.
  • The drafting provider writes no band column at any severity, and a drafted row is gene, motif, trait and a placeholder conclusion — RepeatAlleleRow has no column for the coordinates, locus_structure or ref_copies the catalogue also publishes. That gap is RM65/RM87.
  • Known and not fixed here: just_dna_compiler.draft.append_partial_rows crashes on any table whose header is narrower than its model, which reaches all four existing partial-row providers. Found while drafting into a real module, reproduced independently, filed rather than patched — this round changes nothing in just-dna-compiler.

2026-09-01 — the source-adoption round: the PGS Catalog becomes a registry (RM163)

just-dna-enricher, plus two VALID_VERIFICATION_CHECKS members in just-dna-format. No authored column, no parquet change, just-dna-compiler untouched. Second of the five items PROPOSAL_0_7_PT2 decided.

  • check-identifiers asks a fourth registry. pgs_id was the one authored identifier in the format nothing checked, on the column PgsRow is keyed by. New pgs.py client; the roster derives from DRAFTABLE the same way the other three do.
  • The verdict is read off the response body, never the HTTP status. The Catalog answers 200 with {} for a never-assigned accession and for a malformed one, so the status carries no existence information. A consumer relying on raise_for_status to mean "this id is real" would be wrong.
  • The absence message names the typo reading first. Only about a third of the accession range is assigned, so an unrecognised id is overwhelmingly never-assigned rather than withdrawn — the opposite weighting to the dbSNP message, and stated for the measured reason.
  • Two new verification checks, pgs_accession_currency and pgs_metadata_agreement. Consumers validating VerificationRecord.check against a pinned vocabulary should add both.
  • PGS_TERMS is a floor, not the terms. Each score record carries its own license, and the values are not variations on one licence: most generic, some academic-research-use-only, some CC0. The per-score string overrides the constant in that score's SourceRow. A module naming an academic-use-only score is now refused by the compile gate by name — if you have such a module, it will stop compiling under a commercial declared_use, and that refusal is correct.
  • Drift is checked over training_ancestry and training_cohort only. match_rate_floor and research_tier are author judgements the Catalog does not publish, so there is nothing to compare them against.
  • Currency comes from /rest/info, added as currency.default_probes' second member, so --verify-datasets now covers this source.
  • Known gap, recorded rather than fixed: a pgs.csv drift finding cannot be answered in overrides.csv — the overlay applies to derived tables and pgs.csv is authored.

2026-09-01 — the source-adoption round: MANE becomes a cache (RM168)

just-dna-enricher only; no schema change, no new authored column, nothing in just-dna-compiler. First of the five items PROPOSAL_0_7_PT2 decided, all landing inside the uncut 0.7.0.

  • just-dna-enricher mane build — MANE is a source now, not a sentence in a probe document. CIVIC_IDENTITY_PROTOCOL § 3b pinned a numbering frame with a file "downloaded once and cited"; it is now a cache with a location, a recorded release and a currency check, like every other reference table here. New: MANE_TERMS, $JUST_DNA_MANE_CACHE / default_mane_cache_dir / resolve_mane_reference, mane_build.py, a mane sub-app, a cache status row.
  • Three files, under 1.2 MB together. The summary (19,437 rows), changed_select_accessions (120 rows) and protein_coding_genes_not_in_mane (222 genes). The second is the currency check and ships in the same pass deliberately — a cache without the thing that notices it going stale is the defect the item was about.
  • release.json is copied from the source's own README_versions.txt, so it carries the MANE version, the NCBI RefSeq annotation release and the Ensembl release — two of which appear in no filename.
  • MANE_status is a column and is never collapsed. 74 of 19,437 rows are MANE Plus Clinical, and CDKN2A carries two rows with different CDS numbering. A consumer reading one row per gene would not see it.
  • Update_Affects_CDS is carried as a tri-state. Yes on 74 of the 120 changed accessions: a MANE Select change that moves the CDS moves every c. and p. derived in that frame.
  • The negative roster keeps its reasons. 222 genes over a seven-member vocabulary derived from the file, so "MANE has no answer for this gene" is distinguishable from "nobody asked" — and pending MANE review is neither.
  • Terms are None, not permissive. NCBI publishes a policy rather than a licence, so every gating axis is unknown; there is no --use flag, because a gate whose every answer is a skip is a flag that does nothing. Only NCBI's side was read — nothing is asserted about EMBL-EBI's terms for the same tables.
  • Scope is a transcript-identity aid, not HGVS generation, and the lane documents its own bound: MANE is the default, not the answer. It makes CDKN2A's problem visible and is silent on RUNX1's.
  • Two --help-parsing tests that had been failing every PR run since 2026-08-30 are fixed: Typer renders through Rich, CI sets FORCE_COLOR, and a coloured flag is not one token.

2026-09-01 — CIViC: the identity in a variant's name, and the basis in its own VCF (RM159, RM169)

just-dna-enricher only; no schema change, no new column, nothing in just-dna-format or just-dna-compiler.

  • civic build places 33 variants it used to drop. CIViC states an identity in a variant's name for records whose identifier columns are empty — N150fs (c.448delA), IVS2+1G>A, D1709N — and those identities now ship as civic_identities.CIVIC_NAME_IDENTITIES and are emitted with identity_derivation="curated_name". Coverage over the dated 01-Aug-2026 release goes from 237/290 variants (81.7%) to 270/290 (93.1%) and from 474/533 rows to 507/533; unresolvable_identity falls from 59 rows to 26.
  • identity_derivation gains the member curated_name. Additive, and deliberately not folded into rsid/grch38_hgvs: those mean "the source stated this in the column for it". A consumer filtering on the vocabulary should add the member; one that ignores it sees the rows as ordinary placed rows, which they are.
  • release.json gains curated_identities — applied / superseded / renamed / absent, every member present, summing to the table. superseded means CIViC has since published an identity of its own and the curated row stood down; it is also a free currency signal.
  • civic reproduce now reads 57 coordinates against the GRCh38 reference instead of 24, still 0 mismatches. Every hand-read allele is confirmed at its stated position by an unrelated service.
  • Unchanged: the build is still offline and byte-reproducible from the dated TSV pair, and allele_registry_id is still CIViC's verbatim cell — the CAIDs the resolutions went through are provenance on the shipped table, not the source's statement.
  • The procedure behind the identities is docs/probes/CIVIC_IDENTITY_PROTOCOL.md; the per-variant evidence and the classes that did not resolve are docs/probes/CIVIC_UNRESOLVED.md.

RM169 — civic build --submitted and civic reproduce --submitted.

  • The unreviewed majority is now readable from the dated release. CIViC publishes <date>-civic_accepted_and_submitted.vcf beside the three TSVs, so no API read and no reproducibility bargain are needed — RM160 was filed believing otherwise. Opt-in: 507 rows on 270 variants → 1,149 on 397, rebuild still byte-identical, refget cross-check 57 → 129 coordinates with 0 mismatches.
  • New parquet column evidence_status — accepted / submitted, CIViC's own word, unconverted. Present on every row; on the accepted basis it is uniformly accepted.
  • identity_derivation gains vcf_csq. VariantSummaries.tsv is accepted-only as well, so 112 of the 127 new variants have no row in it; their identity comes from the VCF's CSQ block through the same parsers and the same published identifiers. The member names the file, not the route. A consumer filtering on the vocabulary should add it.
  • release.json gains status_basis, status_counts, vcf_evidence and unjoinable_submitted. status_basis is the field to read before comparing any count here with a count from anywhere else.
  • Unchanged: the TSV pair is still primary (the VCF cannot carry a variant with no GRCh37 position, which is exactly the class RM159 resolved by name), and nothing is placed from the VCF's own GRCh37 coordinate.

2026-08-28 — 0.7.0: the items PROPOSAL_0_7 decided

Cut as 0.7.0 across all three packages and tagged v0.7.0 at 2001215 on 2026-09-12; until the tag this entry described the minor as still in progress, with its number decided and no tag, and six replies (S75–S80) cited it as the record of that state. The tag is the record. Every item in this batch is additive under Principles 3 and 8, and a new optional column is what sizes a release. The heading names 0.7.0 because the proposal decided it per item, so unlike the 2026-08-24 batch below — which was also an uncut minor when written and deliberately names no number — there is a version to write down here. The three pyproject.toml files were bumped to 0.7.0 on 2026-08-31 and work kept landing inside this number until the tag; the 2026-08-24 batch ships inside it too. Each entry below names the packages it actually touched.

The twelve entries dated 2026-08-31 are a batch of their own and are worth reading as one. They are the roadmap items that stood open against this release — RM103, RM108, RM110, RM117, RM136, RM137 and RM138 — plus RM146 and RM150 from the same day's consumer reports, and RM151, which was filed and built the same day as RM117's other half, and RM152, a consumer measurement that arrived carrying no release class and acquired one when its probe was finally run, and RM154, which arrived from a consumer report after the rest were built; all taken in one pass so 0.7 cuts with its own backlog cleared rather than carrying it. Eleven are built; RM138 is closed with its numbers measured and no code changed. The items dated 2026-09-01 are deliberately outside that count — RM155 arrived from a consumer the next morning, RM156, RM157 and RM158 from sweeping its shape across the tier, and RM161 from the pre-build gate run those four preceded; all five ship inside the same uncut 0.7.0. The batch stays the thing it describes rather than growing a member whenever another item lands before the tag: the count is of the 2026-08-31 round, and the rule for a later item is to date it and leave the number alone. Two things a reader should take from them together: carried costs 1.06× gzipped rather than the 1.84× raw the item was filed about, so serve manifests compressed; and two derived values change — gene_metrics.constraint_flags and gene_validity.classifications — which are corrections, so a moved signature there is the fix arriving rather than drift.

  • RM161 — a release record's two halves are written at different times, and the second left the first behind. (just-dna-format; additive — two names into an existing list and one new test. No column, no vocabulary member, no signature moves.) The pre-build gate run for this cut exited 1: gene_validity.superseded_count and identity.version_coerced_from "moved and the release record does not list it". Both are real manifest additions from the 2026-08-31 batch, both carry a DeclaredChange written the day they landed, and neither had reached manifest_fields — the two declarations went in at 06:52 and 07:04, after the measurement the list came from, and the readiness table recorded the gate green earlier that day. The shape is the record's own construction: SweepMeasurement.as_record produces the measured half with declared deliberately empty, which is exactly what makes the gate useful and also what lets a later item add its declaration and leave the measured list behind. Nothing in a checkout could notice — the gate needs the previous release installed and is a release-sequence command by design. The guard is an asymmetry: a declared addition must be listed, because a field that did not exist before moves wherever its block appears; a declared correction may be unmeasurable on the corpus, so gene_validity.classifications and gene_metrics.signature stay declared and unlisted and the gate keeps its existing note for the reverse. The record's evidence sentence already carried the current numbers, so only the field list was stale. After the fix: "release record for 0.7.0 covers the measurement", exit 0. (from the 0.7.0 pre-build gates)

  • RM158 — the GWAS pass asked about one table's rsIDs, and the answer already existed in this package. (just-dna-enricher; additive, and no schema change — no column, no vocabulary member, no signature moves.) gwas._module_subjects built its (rsid, variant_key) list from variants.csv while five authored models carry rsid, so a module whose rsIDs live in haplotypes.csv or pharm_variants.csv got no associations and no line saying none had been asked for — a spec carrying one haplotypes.csv row for CYP2C19*2, which the Catalog has associations for, returned [] against the pre-fix code. The third instance in one sweep, and the one worth reading: the fix was already written. enrich.Subject and its collector exist for exactly this question — resolution read variants.csv alone until RM43, when a PGx module "which by design carries no variants.csv" enriched to an empty resolution.csv — and this pass, written afterwards, restated the narrow loop instead of calling it. So grep for the question, not for the bug. _collect_subjects/_Subject are now public (collect_subjects/Subject), since a private name is what kept the second caller from finding the first. studies.csv carries rsid and is deliberately not a subject: a study row references the variant it grounds, which the module already carries. Measured across the corpus before and after — pathogenic_clinvar 301, hboc_palb2 16, mt_heteroplasmy 2, grch37_build 0, identical lists — because variants.csv goes first in the collector and first occurrence wins, the precedence that stops a PGx row taking an identity a SNP row minted. (from the RM155 sweep)

  • RM157 — the gene set three passes take their scope from read one table while nine carry the column. (just-dna-enricher; additive, and no schema change — no column, no vocabulary member, no signature moves, and no ordering change, since gene_metrics sorts its rows by (gene, dataset) before writing.) gene_metrics.module_genes built its list from variants.csv alone. It is not a report but the scope of the constraint-metrics pass, the gene-validity pass and the ClinGen dosage pass — all three call it — so a module whose genes live in its PGx tables had all three quietly do nothing: no rows, no findings, and no line saying a question had not been put. Measured on this repo's own corpus, where cyp2c19_star_alleles, apoe_epsilon, cyp2c9_warfarin_grch37 and hfe_compound_het returned [] while naming CYP2C19, APOE, CYP2C9, VKORC1, CYP4F2 and HFE. Two things worth reading together: the workspace already held two answers to one question, since pgx._module_genes reads two PGx tables and nobody had put them side by side; and RM104 had patched the symptom — an UnboundLocalError on "any module with no variants.csv" — with that sentence sitting in a comment, describing the defect as a shape rather than asking why the list was empty. The set is now derived from the same registry walk the identifier roster uses, and a table that will not parse refuses here rather than being routed to not_read: a reporting surface may narrow, a scope may not. IdentifierRoster gained read_errors so the loader's own message survives into that refusal and gene_validity's variants.csv is invalid diagnosis is unchanged. pgx._GENE_TABLES stays two tables — it decides whether the star-allele cross-check applies, which is a fact about that check's inputs. (from the RM155 sweep)

  • RM156 — the roster RM155 widened was gated behind the one table it had stopped depending on. (just-dna-enricher; additive, and no schema change — no column, no vocabulary member, no signature moves. What moves is which modules the check runs on at all.) Two gates in front of the widened roster were still keyed on variants.csv: check_identifiers(spec_dir=) loaded that table unconditionally and raised variants.csv is invalid: ... not found, and check-identifiers returned "no variants.csv — nothing to check" one call earlier and hid it. The table has never been mandatory, and four of the nine tables carrying gene are the PGx kinds a module is built entirely out of — reproduced on this repo's own corpus, where cyp2c19_star_alleles, apoe_epsilon, cyp2c9_warfarin_grch37 and hfe_compound_het carry no variants.csv, name CYP2C19/APOE/CYP2C9/VKORC1/CYP4F2/HFE between them, and all four exited 0 having asked nothing. S86's unreadable 0, one level above the function that repaired it. The command's guard is now the roster — nothing to check means no id-bearing table was read, and only then is no attestation written, which is the half of the old guard that was right; an absent variants.csv is no rows, a present-and-unparseable one still raises. A third vacuous pass beneath them: with symbols in hand and no rows, _gene_locus_conflicts returned compared=0 with None beside it, the ran(0, 0) its own attestation docstring forbids. (from the RM155 sweep)

  • RM162 — RM_TOC.md is an index, and an index is not an allocator. (tooling only — .claude/rm-next.py plus its test; no package, no schema change, nothing a consumer installs.) The Sn loop has allocated ids since it was built, because an id written into a document collides when it is stale. RMn had no allocator: the number was read by grepping "the highest in use", and the window between that read and writing the entry is where a second session reads. Reproduced by this repo on itself — on 2026-09-01 two sessions sharing one working tree filed different work as RM159 a minute apart, and 741ec59 renumbered one to RM161. The new tool scans every docs/**/*.md and reserves the next number in the same locked write; a tool that only printed the number would be the same defect with a nicer interface. The lock is on docs/, never on RM_TOC.md: a leftover lockfile would block every later run (@flock-not-a-lockfile), and — the half that was measured rather than assumed — flock binds an inode, so an atomic rename-over leaves the holder locking an unlinked file while a second process acquires immediately. A reservation is a visible 🔷 row in the index rather than a side-car it cannot see, and --release leaves a ✖ tombstone because ids are never reused — the first cut deleted the row, the number went invisible to the scan, and a released RM10 was handed straight back out. Pinned by a test that runs eight allocators at once and runs the same eight with flock neutered to watch them collide (5 distinct of 8, one number taken three times). The loop's Step 5 bullet, which told an agent to read the number off the index and named a highest that had been stale for 114 items, is corrected. (the 2026-09-01 collision)

  • RM155 — the identifier roster read one table while eleven carry the column, and reported its blindness as a clean zero. (just-dna-enricher; additive, and no schema change — no column, no vocabulary member, no signature moves; the new report fields default to empty, so an existing caller reads unchanged.) check_identifiers built its trait and gene rosters from variants.csv alone while eleven authored models declare trait_efo_id or gene — StudyRow has carried the trait column since 0.3 — so a 67-variant module carrying its trait id on all 68 studies.csv rows reported nothing checked and nothing flagged, and could ship a retired CURIE with every gate green. The unreadable 0 was the defect rather than the omission: it asserted this module declares no trait and its traits are in a table nobody read in one breath, which is @unreachable-not-absent at a finer grain, so widening the roster alone would have left the hole — a wide roster still returns [] for a module that genuinely declares none. Both halves shipped: the roster is derived from DRAFTABLE (nine tables per column, asserted as an equality over the walked _ALL_MODELS rather than listed, so a kind added later joins by existing), and IdentifierReport gained trait_tables_read/trait_tables_not_read plus the gene pair, with the CLI count naming its own denominator. MeasureBinRow is correctly absent — the abstract base whose four concrete subclasses are each their own entry — and the three derived models carrying these columns are excluded on purpose, since a stale id in a machine-written row is the source's currency and no author can act on it. A third instance one level up, found by the same framing: report.clean is all() over a possibly-empty set, so check-identifiers printed a green "all identifiers current" having asked nothing at all. An absent optional table and one that exists and will not parse are kept apart; only the second warns. (S86)

  • RM154 — an rsID the source HAS was published as an absence, and the warning explaining it was false. (all three packages; additive, and no schema change of any kind — no column, no vocabulary member, no signature moves, and every existing module recompiles byte-identically.) enrich writes status: not_found — this source has no record of your rsID — when the source answered and the allele-aware filter rejected every locus. Reported over five subjects of a 64-variant longevity module authored from a GRCh37/hg19 supplementary: the paper spells the submitted strand, so its G/A meets GRCh38's C/T, and Ensembl returns all five immediately. Reproduced offline against the real path — a snapshot that has the rsID with complemented alleles and one that genuinely lacks it write byte-identical rows. This is the fourth state in the family RM98 built (unreachable_rsids = the request failed, unconsulted_rsids = nobody looked, unresolved = no position and silent about why): here the asking succeeded and the answer did not match. New EnrichmentResult.allele_mismatches, carrying AlleleMismatch(rsid, genotype, loci, offered, strand_flip), plus one aggregated warning in both modes saying the source has them. The row is deliberately unchanged: a new VALID_RESOLUTION_STATUS member is a wire change, and deleting the row moves resolution_signature (variant_key/rsid are fact fields, status is not — measured, not reasoned), so what moved is the reason rather than the row. Second defect, in the sentence the reporter quoted: hosting_verdict's two False arms shared one explanation, so a strand-flipped SNV was diagnosed as "The event sizes differ, which re-anchoring cannot change" about two 1 bp substitutions — compiler.resolution.contradiction_reason is now undecided_reason's twin on the False side, with a test asserting the arms' reasons stay pairwise distinct. strand_flip_explains and reverse_complement land in format (pure string work, and the compiler's twin site needs them); both withhold rather than guess — a degenerate code cannot be complemented, and a palindromic SNV that already fits is never reported as a flip. (S85)

  • RM153 — the identity CIViC does not publish, recovered through a registry rather than by lifting a coordinate. (just-dna-enricher; additive, no schema change — one client, one snapshot derivation, three withhold reasons, one licence row.) RM152's residue. CIViC publishes GRCh37 coordinates or none, and after every published identifier is read some variants still have no route to a GRCh38 identity. Two questions, answered in opposite directions. Taken: the ClinGen Allele Registry. A CAID is build-independent; the registry serves an rs-number and a GRCh38 coordinate, needs no key, and answered 102 probe requests with zero failures. Recovery over the direction set goes from 138/290 (48%) to 237/290 (82%) — 52 via an rs-number, 12 via a coordinate, and 35 one-sided indels anchored. The rs-number is preferred because ClinGen supplies it and Ensembl verifies it: two authorities, so the check is real, which is precisely what a lifted coordinate lacks. New identity_derivation="caid" keeps a route-bearing row in the snapshot instead of dropping it; unresolvable_identity now means no identifier at all and falls from 204 rows to 59. The pass runs at draft time — a build that fetched would forfeit the offline reproducibility the dated input exists to give — and --offline withholds those rows as caid_unresolved: unplaced, never unplaceable. One-sided indels are anchored VCF/Picard-style. The registry states an insertion with an empty referenceAllele and a deletion with an empty allele, in interbase terms; neither is a row a ref/alts pair can hold. Prefixing both sides with the reference base before the event is the left-aligned form VCF requires, and the interbase start is that anchor for both shapes, so one rule covers them. All 35 rows that previously read no_identity are one-sided indels and every one anchors. Verified twice: the base at chr3:10142013 is G, and ClinGen's own HGVS is g.10142013dup, which is exactly the G>GG produced. An unreadable anchor is withheld under its own reason, never guessed. Refused: liftover. Reopened on request and closed on the number — ceiling 13 evidence rows on 9 variants, honest recovery at most one. Three are gene-level assertions no genotype satisfies on any build; five are imprecise by the source's own HGVS and the registry refuses to parse them at all; and variant 2099 lifts exactly to a different allele than its own name and alias describe, which is RM48's hazard demonstrated rather than argued. pyliftover agrees with Ensembl on all 18 endpoints so buys no accuracy and downloads an unpinned chain; Picard LiftoverVcf was not run because 8 of the 9 carry no REF/ALT to feed it. The registry's terms are unestablished and recorded as such: reg.clinicalgenome.org/site/terms answers 200 with a generic broken-link page, so every axis is None. ClinGen's CC0 grant covers the gene-curation surface, not this one. Nothing is redistributed. Measurements in CIVIC_SURVEY.md and CIVIC_UNRESOLVED.md.

  • RM152 — CIViC adopted on the axis it can answer, and refused on the one it was proposed for. (just-dna-enricher; additive, and there is no schema change of any kind — no column, no table, no vocabulary member. The adoption rides on direction and state, which have existed since 0.3, which is why it lands inside an uncut release without sizing it.) S84 proposed CIViC as a second clin_sig concordance authority; measured, its germline subset carries five ACMG-tier calls and zero benign-class, so discordant is unsayable and the check would read single or concordant by construction. The same measurement found 1,458 germline rows on Predisposition/Protectiveness — the direction axis — where the NA count is 0. So the drafter that was rejected for "rows with an empty significance column" is right on direction and wrong only on clin_sig (812 NA), and that is what shipped. New civic build: a dated release (--release 01-Aug-2026) reduced to one parquet plus release.json, byte-reproducible. It reads the bulk TSVs, not the GraphQL API, because only the download side is dated — and that surfaces a fact neither surface declares: the TSV is accepted-only at 4,903 rows while the API defaults to NON_REJECTED at 11,518, so a count from one is not comparable with a count from the other. Three inputs are required, because MolecularProfileSummaries.tsv is what tells a combination genotype from a dangling reference. New draft-panel --source civic, writing direction and never clin_sig; --clin-sig is reported inert rather than silently ignored. New CIVIC_TERMS (CC0 1.0), distinctive not for permitting redistribution — five sources here do — but for asking nothing back. A direction-axis concordance record was refused, which was RM152's open question: genuine risk-versus-protective opposition is 0 at every scope and status basis, the three contested variants are claim-against-refutation and dissolve under ACCEPTED, and nothing else in the enricher fills direction, so there is no second authority to be concordant with. Coordinates are GRCh37 or absent, never GRCh38, and the snapshot survives that by reading the rsID and GRCh38 accession CIViC publishes beside them rather than lifting anything — RM48's rule applied, and better than it hoped, since a published rs-number is an independent value resolution can cross-examine. The residue is RM153. A refutation is kept with its direction withheld: "does not support predisposition" removes a claim without establishing its opposite. One fix in shared code, found by dogfooding rather than review — append_partial_rows built its covered-set from partials[0].match_on while comparing each row against its own, so a mixed-arity batch re-added rows on every lap: invisible on a first run, a file that grows thereafter. Providers now pass one constant tuple and a mixed batch is refused. Measurements in CIVIC_SURVEY.md; the decision is the RM152 addendum in PROPOSAL_0_7.

  • RM146 — every authored column now says which release it appeared in. (just-dna-format; additive — a marker on each field declaration. No column, no parquet, no signature, no manifest field.) A module authored on 0.6.6 met a registry running 0.6.1 and got studies.csv line 2 [curator]: Extra inputs are not permitted — byte-identical to what a typo ([curatr]) produces, and the two want opposite actions from an author: upgrade the reader, or fix the cell. pydantic's message under extra="forbid" could not carry the distinction, because the information was not in the model.

New: just_dna_format.base.since("0.6.5") on every authored field, read with base.field_first_seen(model) — a plain {field: release} map, so any tool rendering our findings can answer when did this column appear offline. It composes with vocabulary() in one json_schema_extra dict rather than replacing it. stamped_identity_field now takes first_seen as a required argument: a compiler-stamped column is still one an older reader refuses.

The backfill was measured, not recalled — parsed per release tag from the AST rather than imported, since old code need not import under a current Python. 414 fields across 31 models: 115 date to 0.2.0, 81 to 0.4.0, 150 to 0.5.0, 160 to 0.6.0, 3 to 0.6.5, and 78 land in 0.7.0. The answer is per (model, field), never per name: curator is on VariantRow from 0.2.0 and gains its StudyRow twin only in 0.6.5.

  • RM117 — an answered conflict the archive has since resolved says so, instead of reading as a typo. (just-dna-format + just-dna-compiler; one new warning code, no field and no signature moves.) clin_sig_concordance.csv holds contested subjects only and is rewritten whole, so a subject leaving the record means the authorities stopped disagreeing — which is how an author learns the archive caught up with them. An overrides.csv row answering that conflict then reaches nothing, and until now it drew the generic the subject may be mistyped, or the correction may be aimed at a row the compiler drops, put to an author whose judgement had just been confirmed.

New: overlay_answer_vindicated (actionable — the author can retire the overlay row, and nobody else can). It is an observation about the record and not a verdict: the authorities now agree and the row is unnecessary; who was right about the biology is not something a compiler can say. Scoped to that one table, because every other unmatched update is ambiguous and takes RM137's split.

This is RM117's observability half. The severity half stays closed, and the second signal — a justification written about a value the source has since changed — is RM151, below: it needs the archive's value now against its value at record time, so it needed a fetch and an enricher home.

  • RM151 — an answer written about a disagreement the archive has since changed says so. (just-dna-enricher; two new warning stems, no field, no schema and no signature moves.) An overrides.csv row against clin_sig_concordance.csv is a judgement about a particular disagreement, and its reason explains why. When the archive later says something else, that reason describes a disagreement that is no longer the one on record — and nothing said so.

enrich() now compares each authority's fresh call against the previous run's clin_sig_authority_calls.csv, per (variant_key, genotype, authority), for every subject the module's overlay answers. That table is the only place this format keeps what a source said at the time, so the finding names it: an overlay row against frequencies.csv or resolution.csv has no recorded baseline and nothing here promises one. Warns in both modes, escalates in neither.

Three states, and withheld is never unchanged: no_prior_record (first run, unreadable previous record, or the authority was unchecked when the answer was written) and unchecked_now are info notes, not agreement. A difference that is only our own normalizer moving — same verbatim clin_sig_raw, different normalized member — is reported as a separate sentence, because folding it in would accuse a source of a change we made.

A move is observable exactly once, by the run that notices, because the commit rewrites the baseline it read. That is the honest shape for an observation; persisting it needs the overlay row bound to the value it justifies, which is an authored-surface change and the binding RM117's objections all turned on missing. Silent on a module with no overlay answers, which is every module today. New surfaces: clinical.answered_call_shift, concordance.shifted_authority_calls / read_recorded_calls / answered_call_sentences / answered_call_notes, licensing.overlay_answered_subjects, and EnrichmentResult.answered_calls.

  • RM138 — closed with its numbers measured; no code changed. (documentation only.) carried duplicates the text of the warnings it names, growing compilation 1.84× across the reference corpus. Re-measured with the compression a real transport uses: 1.06× gzipped over the corpus and 1.13× on pathogenic_clinvar, the 113-warning module the item was filed about — because carried is a verbatim subset of warnings, which is precisely what DEFLATE's back-references eliminate. The whole with-carried payload gzips to 0.21× the uncompressed warnings-only one.

The encoding stands, and the recommendation is to serve manifests compressed where the size lands — a catalog, an API response, anything shipping manifest.json over a wire. The three cheaper encodings stay rejected for the reasons the item already recorded. Closed inside 0.7 deliberately: a fourth encoding after 1.0 would be a removal, and removals are major-only.

  • RM136 — the enricher reads the author's overlay, so a correction stops coming back forever. (just-dna-compiler + just-dna-enricher; additive — one loader made public, one new counter on an internal result, no manifest field and no signature moves.) The compiler applies overrides.csv before any check reads a row; the enricher re-read the raw derived file, so an author who corrected a resolution.csv cell through the overlay kept being told the same finding on every run, with nothing saying the correction had been recorded and honoured one tier over.

Read-only, at input reads, per field. The enricher never writes through the overlay. Passes that read resolution.csv as an input (frequencies, assertions, identifiers' gene-locus check) now see the post-overlay rows; every merge baseline stays raw, because a pass that reads its own output file writes it back and post-overlay rows would bake the correction into the derived table. A finding is answered only when the overlay updates the very cell it is about.

Answered is not agreed: the pair stays in the denominator and PairCheck.answered counts it, with one INFO line saying the correction is recorded. Dropping it would report a cleaner module than there is. New public API: just_dna_compiler.compiler.load_overlay, plus licensing.overlaid_input_rows / licensing.overlay_answers in the enricher — there is deliberately no second implementation of apply_overrides.

  • RM137 — the unmatched-overlay warning is now a property of the module, not of the lap. (just-dna-format + just-dna-compiler; one warning code added, one reworded — no field, no parquet, no signature moves.) An overlay update naming a row the compiler drops matched on lap 1 and warned on lap 2, so a module and its own compile → reverse → compile disagreed on manifest.compilation.warnings. The stable quantity turned out to be a property of the target — could an artifact of this module carry that row at all — which is computable from data that survives the round trip, so it answers the same on both laps whether or not the row is there to match. The unreachable finding therefore fires matched or not; that asymmetry is the fix.

Two codes where there was one, and neither reading is "a typo" — a mistyped PMID is also an uncited one, so a mistake lands in the unreachable bucket: - overlay_update_target_unreachable (new, actionable) — no artifact of this module can carry the row: mistyped subject, or a correction aimed at a row the compiler drops, which is fine. - overlay_update_unmatched (reworded, so grep it afresh) — the subject is cited or positioned, so the table is short rather than the correction wrong. Re-run the enricher.

Scoped to literature.csv and resolution.csv (LOSSY_OVERLAY_TABLES), the only two a reverse rebuilds from something narrower. The other six rebuild whole, so their warning was lap-stable already and its text is unchanged. apply_overrides gains an additive defer_unmatched keyword; update_targets / classify_update_targets are the split, and cited_pmids / literature_target_survives / resolution_target_survives are the predicates.

  • RM150 — direction gains contested; unknown stops meaning two things. (just-dna-format; additive — a new member on an existing vocabulary. Nothing re-points, so no published module changes meaning and nothing drifts.) unknown had been carrying both nobody assessed the sign and the sources disagree about it — an absence and a finding — and RM148's own field description said so without giving a consumer any way to tell them apart. unknown keeps its original meaning, because re-pointing a shipped member would silently change what every published module already says by it; contested is added beside it, using the word this workspace already uses one table over (clin_sig_concordance.csv).

This is the shade RM148 could not fold into a pair. An unestablished sign is still a sign, so evidence that does not exclude either direction is direction=<sign> + stat_significance=not_significant and needs no member — but no pairing of the two axes says two sources disagree about the sign.

Upgrading a legacy module is unchanged. _STATE_TO_DIRECTION deliberately gains nothing: no legacy state value means contested, so only an author writing direction directly can produce it. stat_significance gains nothing either — a disputed sign is not a disputed strength. contested projects to state=neutral through trimmed_state(), like unknown, and that entry is explicit: the map is read with a default, so a member missing from it projects silently rather than failing, and the guard is now an equality over the walked vocabulary.

  • RM108 — a re-curation is recognised, and currency is derived rather than marked. (all three packages; additive — one new manifest field and two new warning codes. No column changed, so gene_validity.signature does not move and no module recompiles to new bytes.) ClinGen's assertion_id embeds the curation timestamp, so a re-curated assertion arrives under a different id, misses the merge key, and is appended beside the row it replaces — manifest.gene_validity.classifications then published ["definitive", "refuted"] with nothing saying which stood. The newest classification_date is now read as current and nothing is deleted; both rows stay in the file so the drift is visible.

The marker column the item was filed for does not exist, deliberately. The row that must be marked is the one already in the file, and merge-not-clobber forbids the pass editing it — so a stored marker would be correct on every run except the one that created the ambiguity. Currency is derived at every read (just_dna_format.gene_validity.classify_currency, public), grouped on (gene, disease_id, moi, submitter) — the source's grain minus dataset.

Two edges withhold rather than guess: a tie on classification_date, or any member of the group stating none, leaves no row current and none superseded; those groups publish all of their classifications. New: manifest.gene_validity.superseded_count, and the warning codes gene_validity_superseded / gene_validity_currency_undecidable, both carried — an author's only available edit is deleting a row, which falsifies the record. They warn in both modes in both tiers and escalate in neither; validate reports what compile reports.

One behaviour change beyond gene-validity: the compiler's fact-table loop now de-duplicates check warnings on the message, as every both-sides check already did. This is the first fact check the pre-flight also runs, and without it the line — and its warnings_summary count — arrived twice.

  • RM103 — the manifest now records the version that was read, not only the one that was invented. (just-dna-format + just-dna-compiler; additive — a new optional manifest field, no digest moves.) normalize_version strips every non-digit and pads to three zeros, so version: abc becomes "0.0.0" — deliberate since RM17, and unchanged here. What changes is that the artifact recorded only the result: manifest.identity.version read 0.0.0 with nothing beside it saying the author wrote abc. New: manifest.identity.version_coerced_from, the authored string when the model rewrote it and None when it was already canonical SemVer. The compiler's warning naming both values is unchanged and still fires in compile and validate alike; a build log is just not what a consumer holds.

reverse_module now re-emits the pre-coercion string, recovered from the artifact's own manifest.json (version_coerced_from first, version second). Re-emitting the coerced one would leave the next compile nothing to coerce, so the new field would go absent on lap 2 and a module would disagree with its own round trip on a published field. An explicit --version still wins, and a bare parquet directory with no manifest still leaves the key out. Side effect worth knowing: reverse previously dropped module.version entirely unless a caller supplied one, so a reversed spec now carries the authored version where it used to carry none. Nothing hashes on it.

The refusal half of RM103 — should a digitless version be rejected outright — is unchanged and stays on the 1.0 cleanup tracker; it is a tightening, and no sentinel SemVer exists to coerce to.

  • RM110 — constraint_flags had two producers, two encodings, and one of them inside the fact set. (just-dna-format + just-dna-enricher; one existing artifact's digest moves — see the cost line.) The live gnomAD route pipe-joined its flag list; the bulk-TSV snapshot route copied gnomAD's JSON array literal into the cell. Re-probed against the published v4.1 parquet before any code moved: of 18,111 rows none is null or empty — 17,403 carry [] and 708 carry a real literal in 14 distinct shapes — so if row.constraint_flags: was true for 100% of snapshot rows where the flagged fraction is 3.9%, and splitting a two-flag cell on | returned one bogus token. The column is inside GENE_METRICS_FACT_FIELDS, so the same gene fetched two ways minted two gene_metrics.signature values.

Fixed on the model, not in the fetching tier: just_dna_format.gene_metrics.normalize_constraint_flags, bound as a mode="before" field validator (neither producer hands over the declared str | None, and a mode="after" validator cannot rescue a value the type rejects first). The published snapshot is immutable and every gene_metrics.csv already written from it carries [] on disk, so a producer-side fix would have left those tables contradicting the column's own description. Three call sites share the one function — the live route, lookup_snapshot (the published snapshot) and constraint_build (future ones) — and it is idempotent, so a rebuilt snapshot passes unchanged. "Empty → null" was only half the fix: the non-empty cells needed parsing, or the 708 flagged rows stayed unreadable. A bracketed string that does not parse is kept verbatim rather than guessed at.

Cost, measured: exactly one row in the reference corpus — reference_examples/hboc_palb2/ carried constraint_flags=[], so its gene_metrics.signature and artifact.digest move; the checked-in CSV was corrected in the same commit. Nothing else in the sixteen examples changes. If you have compiled modules from the v4.1 constraint snapshot, their gene_metrics.signature will move on the next recompile, and that is the fix arriving rather than a regression. The field description, false on the snapshot leg in both directions, was rewritten.

  • RM147 — a source read by hand that yields no row had nowhere to go, and the home already existed. (just-dna-format; documentation and a test — no behaviour changed.) Filed and answered on 2026-08-31 from a consumer report (S82), which asked for a view rather than proposing a shape.

An agent read five literature services by hand to confirm the papers behind two rows and recorded it as five licensing.csv rows at layer=literature. The reporter removed them on our own two rules — a literature source's terms are per article (RM46), and a pass that put no row in a table records no source (RM142) — and measured that they bought no enforcement. Then asked the real question: after removal, nothing anywhere records that a human went and looked.

The home already exists. A literature.csv row no study, bin or pharm row cites is kept in the CSV and dropped from the artifact with literature_row_uncited, shipped in RM79 for a citation the author deleted — and it is the same shape for the opposite case, a paper read that did not become a row. It is structured and checked, it cannot make a licence claim, and it is about the paper, which is the thing consulted; a service is only how the author reached it.

Consumers: nothing changed. LiteratureRow's docstring now says this is that case's home and why the licensing table is not, and a test authors the reported shape — two articles, one cited and one not, green --strict, the uncited one named, and no licensing.csv at all, because nothing is owed for reading an abstract.

  • RM148 — direction and stat_significance are one pair, and the description did not say so. (just-dna-format.) Filed and built on 2026-08-31 from a consumer report (S83), an hour after S80 was accepted and in the same spirit.

Two runs of a byte-identical prompt, same model and paper, wrote different values for one variant on one body of evidence: risk/suggestive against unknown/not_significant, both green. The evidence was two cohorts trending the same way at p ≈ 0.073, combined OR 3.58 with a CI of 0.96–13.4 at 28.4% power — and both readings were defensible against a description that named the four members and said only orthogonal to state.

Not a vocabulary gap. direction records the sign of the reported estimate; stat_significance records how far to lean on it; and the orthogonality is itself the answer to is a sign you cannot lean on still a sign — yes, because the other column says you cannot lean on it. The state the report wanted a member for is already the pair, direction=risk + stat_significance= not_significant, which authors and validates today.

The description now carries that reading and bounds unknown: no sign to record — not assessed, or the sources conflict — never a sign you may not act on. Writing unknown for a weak trend discards the sign the paper reports and leaves stat_significance speaking about nothing.

Consumers: one description string, so anything rendering model_fields picks it up. No member added — a looked, no sign established member would be a second spelling of the pair, which is Principle 5's overloading arriving as a synonym, and permanent under P3. A test asserts the two vocabularies stay disjoint but for unknown.

  • RM144 — the licence-disagreement warning printed the remainder as though it were the whole set. (just-dna-compiler.) Filed and built on 2026-08-31 from a consumer report (S79).

The check filtered the annotation-layer rows to those whose licence differs from the declaration, then rendered that remainder as the whole set. A two-source module declaring CC-BY-NC-ND-4.0 — an exact match for one row, and the binding constraint on the artifact — printed sources report ['CC-BY-4.0'], with the agreeing row invisible in the sentence complaining about agreement.

Two problems with different repairs read identically: your declaration is unsupported versus your declaration is not universal, the second being the ordinary mixed-licence shape where the most restrictive term binds and the declaration is already correct. It cost two agents a full re-adjudication that found nothing wrong, and it survived RM142's fix — removing the phantom row leaves a real disagreement still rendered as total.

The count now leads and the agreeing rows are named beside the disagreeing ones, with a distinct sentence for the genuinely-unsupported case. The denominator counts rows, not distinct licences — two sources sharing a licence are two obligations — and a licence-less row or a non-annotation layer stays outside it. Suppressing the warning when any row matches was refused, on the reporter's own argument: a module declaring the least restrictive of several is exactly the one worth warning about.

Consumers: the message text changes; declares license still leads it, which is the fragment the existing test keys on, and the non-escalation is unchanged and re-pinned. No reference example moves a digest, signature or warning — the corpus has no mixed-licence module, which is why this survived it.

  • RM145 — state's six members were printed as peers, and two of them are retired in our own code. (just-dna-format.) Filed and built on 2026-08-31 from a consumer report (S80).

VariantRow.state was described as One of: risk, protective, neutral, significant, alt, ref — six values, no standing — while derive.py calls alt/ref the retired descriptors and maps both to direction=unknown. A consumer whose authoring surface passes our descriptions through verbatim (deliberately, so a vocabulary change reaches an author without being restated) therefore offered an agent six equal choices, and it picked alt for a heterozygote. The reporter had to read derive.py inside their own .venv to author one cell honestly. Measured across the sixteen reference examples: 377 risk, 4 neutral, and zero uses of significant, alt or ref.

The description now separates current members from superseded ones and names each group's successor. Three groups rather than the two the report asked for: state is the Principle 5 anti-pattern the charter names by hand, so the split is by which axis a value was really on — significant is a significance claim that stat_significance owns, not a dead value, and grouping it with alt/ref would tell an author it means nothing when it means something this column is the wrong place for.

Consumers: one description string, so anything rendering model_fields — describe, requirements, reference and the reporter's own surface — picks it up with no change of their own. No vocabulary member added or removed and nothing invalidated; removal stays major-only and was not requested.

  • RM143 — the enricher diagnosed a wrong-assembly coordinate and compile --strict built over it anyway. (just-dna-compiler.) Filed and built on 2026-08-31 from a consumer report (S78).

A GRCh37 coordinate pasted into a GRCh38 module — rs61849494, 5.6 Mb and a strand flip away from where the module says it is. enrich --strict refuses with a diagnosis naming the rs-number to author instead; enrich best-effort reports it and writes the table; and compile --strict then succeeded silently, over a module internally consistent and about the wrong locus.

Two of the reporter's three asks were already shipped, which is half the answer. Their option (2), re-running the rsid↔coordinate check in the compiler, is refutable on the data: resolution.csv holds one coordinate, not both — for a coordinate-authored row the enricher records what the author wrote — so there is nothing to compare without a fetch, and P2 forbids the fetch. Their option (3), a compile warning, is verification_findings_recorded, which shipped in this same release and is absent from the 0.6.6 they measured.

What was missing was the last step of their option (1): the diagnosis reached verification.json and no severity attached to it. build_disagreement_error now refuses a strict compile on a recorded genome_build_agreement finding, in both validate_spec and compile_module, with the error equal on both sides and placed ahead of output_dir.mkdir() so a refusal writes nothing.

The strict line does not move. strict still means reproducible, never right — the compiler has no reference and a file shifted by one base still passes. This one check is the exception on internal-consistency grounds: its findings say the rows are on a different assembly than the genome_build the module declares, which is one authored file contradicting another rather than the module disagreeing with an outside archive. Every other recorded finding still only warns, pinned over four checks including reference_allele, which produces this diagnosis's own input.

Consumers: a strict compile refuses a spec whose verification.json records a wrong-build finding — new behaviour for anyone compiling one, and the point. Everything else is unchanged: no attestation is silent (an unverified module is the ordinary case), findings=0 is a clean bill, and a skipped record is unknown, which is what --offline writes. best_effort builds and warns as before. No reference example carries such a finding, so the 0.7.0 release record is unaffected.

  • RM142 — the dosage pass declared a ClinGen obligation for a module ClinGen curates nothing of. (just-dna-enricher.) Filed and built on 2026-08-31 from a consumer report (S77).

A single-variant SIRT6 module. The pass reported dosage: missing: [SIRT6], wrote no gene_metrics.csv row, and wrote a ClinGen row into licensing.csv anyway. That table travels to the registry and is read as this module uses this source, so the module carried a false statement — and it fired declared_license_disagrees against a declared licence that never met ClinGen's, sending an author to adjudicate a conflict that does not exist. Both reproduced.

The compiler cannot catch it. _source_checks exempts the annotation layer from its orphan warning by design (RM46), because that is where an author is told to record a hand-read source and warning about it would make compliance noisy while omission stayed silent. Only the pass knows whether it contributed.

The fix is the rule the rest of the family already follows: gene_metrics, frequencies, assertions and gene_validity all derive the source set from the rows they wrote, so a pass that wrote nothing records nothing. clingen.py alone built a fixed row and wrote it unconditionally. The siblings were checked rather than assumed — run offline over a module they cover nothing of, neither writes a licensing file at all.

Consumers: a module whose dosage pass covered no gene no longer carries a clingen row, and no longer warns about a licence disagreement it had no part in. A module ClinGen actually fed is unchanged — keyed on what this run covered, not on what the table holds and not on the absence of missing genes, with tests for all three. No reference example changes, so the 0.7.0 release record is unaffected.

  • RM141 — validate --strict blessed a module compile --strict refused, whenever the resolution table was partial. (just-dna-compiler.) Filed and built on 2026-08-31 from a consumer report (S76) whose headline mechanism did not reproduce, and whose second finding was ours.

The report: an enrich killed by a quota limit left a resolution.csv covering 201 of 263 subjects, and — the urgent half — merge-not-clobber was said to turn that into a silent wrong answer, since re-running would merge onto the stale rows and never retry the missing 62. That does not reproduce, on this tree or on the v0.6.6 they ran. enrich gap-fills: a run over a table covering one of three subjects asks the source about exactly the other two. The atomic write they asked for is also already shipped, as RM128 in this same release, which is why the interrupted run left the previous table rather than a truncated one.

The reporter then corrected their own account, which sharpens what this closes: the file was never truncated — 203 sorted rows, a clean final newline, and the 62 absent rsIDs scattered across the whole range rather than forming a tail. A complete write of an incomplete resolution set, which matches the code: a subject whose live request could not be made is written as no row at all, so the table never states a negative nobody established. Nothing was interrupted, so the atomic writer would not have prevented it — and the same file comes out of a best_effort run that completes normally over an unreachable source. The closure is this item, by the right route: reading the table against the spec beside it is indifferent to why a row is absent.

What is real is that nothing said so until the compile. compile --strict refuses a module whose variants have no position after resolution; validate --strict did not report it at all. So the pre-flight passed clean and the compile immediately after refused — the third break of the parity rule, hiding behind that rule's own exemption. What stays compile-only is a check reading resolved rows, and whether the injected table can place a row is arithmetic over bytes the pre-flight has already loaded.

resolution.unresolved_subjects is the predicate resolution applies, now called from both sides rather than restated, and under strict the pre-flight appends the compile's error verbatim. Two of the compile's own distinctions are kept: nobody-asked is not asked-and-absent, and --no-resolve silences the check the way it silences the fill. A double-report was found doing it — both passes reach the finding for one subject, measured at 24 warnings for 12 subjects with warnings_summary counting 24 — and is de-duplicated on the message.

Consumers: no schema change, no new field, no new warning code. A spec that was going to be refused by compile --strict is now refused by validate --strict too, with the identical message — which is a behaviour change for anyone treating a green validate as a compile guarantee, and is the point. Measured: no reference example moves its artifact.digest, content_signature or warnings, so the 0.7.0 release record is unaffected; none of the sixteen has a partial table, which is why this survived a corpus that size.

  • RM140 — a study row's p-value and effect size were asserted to belong together, and nothing recorded what either came from. (just-dna-format + just-dna-compiler.) Filed and built on 2026-08-31 from a consumer's reproducibility benchmark (S75), after the proposal round had closed — it is not one of the twelve and lands inside this same uncut number, because a new optional column is what sizes a release.

Two agents, byte-identical prompts, the same three DOIs, overlapping on exactly one row: p_value 0.36 against 0.75, effect_size 1.42 on both. Neither was a misreading. The paper reports two analyses of the same association — an allelic Fisher's exact test (OR 1.4, p 0.36) and a univariate logistic regression (OR 1.42, p 0.75) — and one run's row carried the second's magnitude beside the first's p-value. Everything was green, strict on both validate and compile, audit_module and quotes_found included: the provenance quote is verbatim and correct because it grounds the significance verdict and contains no statistic at all, so quote verification is structurally blind to this. study_design describes the study; nothing described the analysis, so a correct row and a mispaired one were byte-indistinguishable to every consumer and every check — and no check could be written, because the facts it would compare were recorded nowhere.

StudyRow.statistical_test is one optional free-form column shaped like study_design: which test or model produced this row's numbers, and what it was adjusted for. No vocabulary, and deliberately no gate — the reporter argued that against their own ask and is right, since the gate is unwritable before the column exists and shipping both would make every published row retroactively incomplete.

The one behaviour change is the duplicate-citation warning. duplicate_study_citation reads a repeated (variant_key, pmid) as the same claim written twice, which two rows naming two analyses are not; both stated and different now suppresses it, and nothing else does. An absent statistical_test is unknown, and unknown against a stated value cannot establish that two rows describe separate work — Kleene rather than a != b, which would suppress on every absent cell and silently retire the check for every module written before the column existed.

Measured, not asserted. artifact.digest moved on 10 of the 16 reference examples — exactly the ten carrying a studies.parquet — and content_signature on none of the sixteen. The published 0.7.0 release record was re-measured with it rather than left standing: its two parquet axes read 4/15 before this item and read 14/15 after, the studies.parquet column is declared, and the concordance-parquet declaration that called itself the release's most visible consequence was corrected, because it was a measured claim and stopped being true. The gate exits 0 against the amended record.

Consumers: one optional column on studies.csv and studies.parquet, read by nothing that does not want it; unset, it is omitted from content_signature, so no published module's identity moves. The warning's code and message are byte-identical for every case that still reports. _KEY_FIELDS is not widened — the published key.columns for studies.csv is still (variant_key, pmid), and re-keying a shipped authored table is major-only. Note the premise this corrects: two rows sharing a variant and a PMID have always both reached studies.parquet — the duplicate is a warning, never a drop — so the capability was there and only the legibility was missing.

  • The 0.7 round's files were closed out. (Documentation only — no package changed.) Four entries sat in forward-only files with a SHIPPED banner on them, which reads as late rather than done: RM126 and RM71 in ROADMAP_0_7.md, RM133 and RM134 in ROADMAP.md. All four moved to ROADMAP_HISTORY.md, so the round's twelve are now in one place.

ROADMAP_0_7.md closed with them, and the succession is the point rather than the tidying. The minor-deferral file is named for the release that will decide its contents, so a cut closes one and opens the next: the round is in history/ROADMAP_0_7.md with the seven entries that shipped or closed, and the ten still waiting — RM122, RM23, RM16, RM28, RM56, RM65, RM66, RM67, RM68, RM84 — became ROADMAP_0_8.md, unchanged but for the file they sit in. PROPOSAL_0_7.md stays beside the other five concluded threads in docs/proposals/ and is marked a record rather than a plan; it no longer wins over anything, because everything it decided has landed.

Every inbound link was retargeted by item rather than by path — a link to a deferral that moved now resolves in ROADMAP_0_8.md, one to an entry that shipped resolves in ROADMAP_HISTORY.md, and one to the round itself resolves in history/. Verified by test_doc_links.py, which is the authority here rather than a fresh sweep: it exempts a consumer's byte-frozen prose, and three links in an archived report were reverted after being retargeted — including the ROADMAP.md#rm89 in S35 that the exemption test asserts on, which records where RM89 lived on the day it was written and whose current pointer is in the reply above it. Our own replies in that file are ours to correct and were. Three stale counters in ROADMAP.md were fixed in passing, including the Active items enumeration, blind for the third time — it read not one of them is a decision through the three decisions filed beneath it.

  • RM139 — the release gate could not tell a broken compile from a spec that outgrew the old compiler. (just-dna-format + just-dna-compiler.) Filed by running RM126's own gate for real at this cut, where it refused the release over reference_examples/cyp2c9_warfarin_grch37/: RM70 put the optional requires_callable column on pharm_variants.csv, that example uses it, and 0.6.6 refuses the spec under extra="forbid". Nothing had failed — the previous release cannot produce a before state for it, so no like-for-like comparison exists — and this recurs in every minor that adds an authored column and exercises it in the corpus, as it does for any example newer than the last release. The gate read one side only as a compile failed and the tag had to be waved through by hand.

The two directions are facts about different releases, and the gate now says which. A module in the BEFORE tree and not the AFTER one is a regression in the release being cut: still fatal, unconditionally, and now carrying the compiler's own errors, which build_outputs had been logging and throwing away. A module in the AFTER tree and not the BEFORE one fails until the published record's new ReleaseRecord.unmeasured list names it. SweepMeasurement splits unmeasured into only_before/only_after and keeps the union as a derived property; the sweep's JSON keeps its original unmeasured key and adds the two halves beside it, so a release script piping to jq is unaffected. UNMEASURED_MODULE_PHRASE still appears in both messages for the same reason.

unmeasured is a denominator, not the per-module escape hatch the roadmap entry refused, and the difference is that the gate checks an equality against what it could not measure rather than a membership test: a module measured on both sides cannot be excused by listing it (that is reported as a note), a regression cannot be excused by listing it, as_record refuses to mint a record over one, and a movement on a measured module gates however the list reads. as_record fills the field from the measurement, so the exclusion is forced into the record the same way declared is — extending the existing mechanism, not adding a second kind of gate input. Records written before the field read [], which is their correct claim; the 0.7.0 record is backfilled with the module its own evidence sentence already named.

Consumers: ReleaseRecord gains one optional list field, minor-legal under Principle 3 and read by nothing that does not want it. build_outputs now returns (outputs, failures) — a compiler internal, but a release script calling it directly must unpack.

Re-measured end to end rather than asserted: the 0.6.6 → 0.7.0 sweep was re-run over all sixteen reference examples with 0.6.6 installed in an isolated environment — fifteen measured, one refused on requires_callable]: Extra inputs are not permitted, gate exit 0 against the shipped record where the cut had needed a human to read past it, and exit 1 naming the module when one is removed from the spec root instead.

  • RM134 § B — the concordance check takes a second authority, and the record has a producer. (just-dna-enricher, plus one optional module_spec.yaml field in just-dna-format and its one-line passthrough in just-dna-compiler.) PubMind (RM134 § A shipped its snapshot) joins ClinVar as an annotation authority, and clinical.clin_sig_concordance becomes an N-authority check whose seam is RM130's classify_concordance — a pure function that already knew nothing about ClinVar. pubmind.py is the runtime reader over the snapshot, duckdb beside the polars builder; there is deliberately no lookup_loci on it, because PubMind's coordinates are back-mappings of extracted text and nothing it produces may enter resolution.csv.

The three-way check subsumes the two-way rather than running beside it. With no PubMind snapshot that authority's call reads unchecked on every subject, and the degenerate case is exactly the ClinVar-only finding: the same subjects are contested, the same conflicts are logged in the same pinned words, and no author meets one disagreement twice. What changes is what the record withholds — authority_concordance reads unchecked rather than single, because one authority speaking while another was never asked is not corroboration and must not be recorded as any.

The record now has a caller. RM130 shipped the models, the classifier and the writer with nothing producing them; enrich() builds the record from the comparison it already ran — never a second pass over the snapshot — and commits both tables at the gate, after every strict refusal, so a refused run leaves none behind. A module holding a ClinVar snapshot therefore carries the record even when nothing is contested: the empty pair is the claim we asked, and nothing here is contested, which is a different claim from the absent pair a run with no authority to ask leaves.

The tautology is decided per leg. Where a module's clin_sig was drafted out of a snapshot it would be compared against and has not moved since, that authority is not consulted at all and states nothing — a call recorded there would agree with the module by construction. But the check is skipped only when every leg is hollow or unasked: a module drafted from ClinVar still gets a real comparison out of PubMind, and throwing that away to suppress the hollow half is the error enrich_pgx already made once. is_tautological_leg states the conjunction once and reads the checked column out of DRAFT_PROJECTIONS, so § C's drafter needs no edit here.

Several PVIDs over one allele fold to one call, camp guard first. PubMind consolidates on extracted text, so one allele carries several records whose verdicts disagree — 35,742 disagreeing keys in the measured file. Where they straddle the pathogenic/benign line the answer is conflicting, the vocabulary's own word for it, in the camp that opposes nothing; folding by severity there would silently answer pathogenic, a winner picked by an ordering nobody defined. Within one camp the fold is the shared normalizer's own severity rule, the same one that resolves a composite token. Confidence is withheld unless exactly one record stands behind the call, and the multiplicity is counted on the record rather than discarded.

authority_precedence: is recorded and computed with by nothing. One optional ordered list in module_spec.yaml, beside weighting:/authorship:, saying whose call the curator weighted while deciding — machine-readable so a consumer can see the stance, and read by no tier. Two modules differing only in it compile to byte-identical parquets with the same content_signature and the same artifact.digest, which is the property the tests assert rather than a round trip of the value. Nothing resolves a split: no majority, no consensus call, no resolved winner, because choosing between a declared order and a majority needs a weighting model this workspace has declined to invent three times.

Warning-tier in both modes, escalating in neither (@clinsig-never-escalates), with more force at two authorities than at one: a disagreement with a literature miner's aggregate is a statement about that extraction's limits at least as often as about the module, the measured corpus join agreed 62 % of the time, and discordant is a fact about the field rather than a defect in the module. A run that found nothing contested reports no zero, and a run that could not put the question writes nothing at all.

  • RM85 — a recorded release, compared against the one its source publishes now. (just-dna-enricher, plus one VALID_VERIFICATION_CHECKS member in just-dna-format.) SourceRow.dataset has recorded which release a module's rows came from since RM4, and two things read it — the tautology skip, and withdraw_stale_dataset when a module ends up mixing two. Neither answered "ClinVar has published since you drafted this", which is the actual trigger for a source-refresh pass, so an author who had forgotten and a curator who inherited the module were both on their own. currency.check_dataset_currency is that comparison, run at the end of enrich(), attested as dataset_currency, switched by --verify-datasets/--no-verify-datasets.

A comparison, not a column. The column-shaped repair — a field saying what this module was made from and what would age it — was refused one table over in RM71 for restating dataset and then rotting where dataset is maintained. This reads sources.csv and writes nothing at all; repairing a stale label is a re-draft, which is an author's decision and a different command.

Three states, and the third is the item. behind is True (the source has published since, with both labels named), False (still current) or None — nobody could ask. --offline is where that bites: an offline run makes no request, so every recorded release is unchecked, never up to date. subjects counts the legs asked and answered comparably, an unaskable source is named in the record's detail rather than counted, and with no leg settled the pass records a skip rather than ran(0, 0). Comparability is a withhold of its own: clinvar_dataset_label's digest form and its stated-date form name one release space in two spellings, so a digest against a date is uncomparable, not behind.

Severity follows the mode, over the superseded set alone. strict refuses on a release the source has moved past and never on an unchecked one — escalating an unreachable source would make --offline --strict impossible forever over something no author can edit, which is the unreachable_rsids rule.

One probe ships: ClinVar. It is the only source this tier can ask for a release label in the namespace it already records — the live VCF's ##fileDate=, read through the same function clinvar_build uses on a downloaded file, because two spellings of one label would not fail, they would simply never match. The stream is abandoned after 256 kB rather than sending a Range header a server may ignore. Every other source reports unsupported, and PROBE_SOURCES is derived from the registry rather than restated beside it.

It is --rederive's cheap neighbour — the same has the world moved question asked about the release label instead of the rows, at one request per source — so ENRICHER § rederive now says to put it first. And it cannot agree with itself: the rows compared are the ones on disk before this run's commit, and the licence rows enrich() writes are at the resolution layer with no dataset at all.

  • RM130 — a contested clinical call now has a name to act on, and a key that survives five authorities. (just-dna-format + just-dna-compiler + just-dna-enricher; two new optional derived tables, three new closed vocabularies, one new warning code.) The ClinVar cross-check reported twenty findings of 141,616 subjects and kept none of the twenty: they reached a logger and nothing else, so an author could read the number and not act on it. clin_sig_concordance.csv is those subjects, named and joinable, keyed (variant_key, genotype) — genotype because the comparison is of an authored call for a genotype, and a table keyed on the variant alone collapses two authored calls that disagree with the archive differently. Beside it clin_sig_authority_calls.csv carries what each authority actually said, keyed (variant_key, genotype, authority).

The shape is RM134's, not the one RM130 first decided, and the swap is the interesting part. The original record carried "the authored value, the source's value" and named its authority in a field — ClinSigConflict.clinvar — which would have cost a key change or a retype the moment a second archive arrived, one item later in this same release. P3 reserves both for a major. So the parent row carries an agreement state instead: authority_concordance (concordant/discordant/single/none/unchecked) and authored_position (matches_all/matches_some/matches_none/absent/unchecked), two orthogonal closed vocabularies that are five members at two authorities and five at five. A single vocabulary was drafted and failed a stress test at five sources because its members named the authority inside themselves; the root cause was one field carrying two axes, which is the Principle 5 anti-pattern and the combinatorial growth was its symptom.

Nothing resolves a split. With five authorities in a two-against-three disagreement, precedence names one winner and majority names another, and choosing needs a weighting model this workspace has declined to invent three times. There is no majority column, no consensus call and no resolved winner anywhere in the tables or the manifest block. Confidence is likewise not normalized across authorities — a gold-star count and a literature miner's evidence-depth count are different instruments — so the detail row carries the published value with confidence_unit beside it, and the model refuses a magnitude with no instrument named.

A conflict is a question and an overrides.csv row is the answer. The record joins the overlay's covered set, taking it from seven to eight, and it is in for a different reason from the other seven: not because it carries curation to lose — it carries none and is rewritten whole on every run — but because answering a contested subject is what an overlay row is. The paired detail table stays out by name: the author answers the question and does not get to rewrite what an archive published. The table's documentation and the new clin_sig_concordance_contested warning both name overrides.csv and never provenance.json's outranks, steering new authors onto the side of the succession that survives 1.0. The warning is actionable rather than carried, unlike its neighbour verification_findings_recorded, because it counts the post-overlay rows — so writing the answer clears the finding, and overlay_rows_suppressed reports the removal so nothing goes quiet. It never escalates under --strict, for the reason the check beneath it does not: half the time the archive is the stale side, and escalating would have this format arbitrate a clinical dispute.

Two smaller things travelled with it. _CLIN_SIG_CAMP moved out of clinical.py into the new concordance.py, because the record and the two-way check must draw the opposed-versus-differing line in one place — two maps for one distinction is how a drift in our own code comes to read as a disagreement between two archives. And an honest note that came out of building it: at a single authority every reported conflict is opposed by construction, since the check only fires where both sides are opinionated and pathogenic/benign are the only opinionated camps, so _clin_sig_detail's differing-but-not-opposed group has no producer today. It gets one as soon as two authorities can disagree with each other while neither contradicts the module, which is what the record is shaped for.

  • RM131 — the warnings channel says what each finding is, and whether an author can clear it. (just-dna-format + just-dna-compiler, plus the deprecated resolver in just-dna-enricher; a structure change to a surface that already ships.) ValidationResult, ClosureResult and CompilationResult all carried warnings: list[str] with no code, no count and no way to tell a finding an author can clear from one they cannot. A compile of a 190-row module returned roughly 14 kB of prose, and strict=false does not help — it changes what counts as an error, not how much the channel carries — while every document on both sides of this seam tells an author that warnings on a green run are the real output.

The answer was already being computed and spent on severity alone. _BLAME_TIER/_BLAME_ROW is literally whose limit this is, and its own comment says "blame decides severity and nothing else"; _closure_warning reaches the same distinction from the other end and says so in prose.

Two derived fields beside warnings, which is unchanged down to the byte. compilation.carried is the subset no edit to the spec directory can clear — a limit of this tier or a fact of a source — so a consumer subtracts it to get the actionable set; compilation.warnings_summary is {code: count} over vocab.VALID_WARNING_CODES, with the values summing to len(warnings) so the digest accounts for the whole channel rather than the part somebody remembered to key. Both ship on all three result types too, on every path including a failed compile. manifest.json sits outside artifact.digest (a Merkle root over the parquet FileEntry list, and that is the mechanism — not a slot in ARTIFACT_PARQUETS, which name-sorts), so neither moved a hash on any published module.

A code names the finding, never the emission site, because a code derived from where the function lives is renamed by a refactor and a published key is permanent within a major (P3/P6). One code carries one remediation: two sentences cleared by the same edit share a code and the sentence says which cell (the weight-sign pair, the five orphan fact tables), two cleared differently do not. Sixty-eight members, nine of them carried.

The audit was the real cost and was done once. Every emission site across three tiers now names a code — the ~29 named-list appends in compiler.py, the findings/messages collectors a survey of .append cannot see, the two .extend sites that reach into binning.measurement_shape_warnings and deprecation_warnings, validate_bins, overrides.apply_overrides, compiler/resolution.py, and the deprecated DuckDB resolver in enricher/resolver.py, whose warnings land in the same manifest.compilation.warnings as everything else. The three sites that reformat a message from another tier go through findings.restate, which is the one operation that carries a code across a rewrite; every other string operation on a finding returns plain prose by construction.

classify has three answers and no flag: derive, withhold, refuse. A warnings_summary with a catch-all key is the rejected repair wearing a different hat — it silently omits what nobody classified, and the reader takes the digest as complete. But refusing outright was wrong in the other direction: CompilationResult(warnings=["prose"]) has been legal since 0.6, so an unclassified channel now withholds (an empty pair, which reads as not classified rather than as complete and short), and only a part-classified channel — which no legitimate caller can produce — raises. That also keeps a raise out of _build_manifest, which runs after every parquet is on disk, where it would have left an output directory with no manifest.json beside them. The published contract is therefore either empty or accounting for the whole channel, stated in both field descriptions; a supplied pair is checked against the channel, since a wrong claim is worse than a withheld one.

The guards are equalities over walked sets, and one of them was a hole until a mutation found it. test_warning_codes.py asserts the emission sites against the vocabulary in both directions (a code with no emitter and an emitter with no code both fail), and the corpus half runs every reference example through both entry points, which is what catches a message that lost its code on the way rather than at its site. The receiver names are themselves a registry now: a hand-typed list omitted lines, which _carried_vrs_warnings collects into, and stripping that site's wrapper passed the guard — so the declared set is checked for equality against the receivers actually derived from the source, with the channel/refusal split declared beside it because one builder feeds both.

The transport stays list[str]. findings.CodedWarning is a str subclass, so every .extend, every if w not in all_warnings de-duplication and every consumer already grepping a phrase keeps working untouched — including the de-dup-by-message that lets a check run in both validate_spec and compile_module and publish one line. Pydantic flattens the subclass at a model boundary, which is right for the published surface and fatal for a caller that keeps building on it, so compile_module and close_module seed from an internal _validate_spec that hands the classified list back beside the result. That trap is pinned by a test rather than left in a comment.

The suppression record rides the same channel (RM124 × RM131). A row removed by a suppress was invisible in the build product — absent, with no trace of why — so the overlay now says so, one line per reason with a count, which is what reason being a required column buys. Counted over the overlay's rows and never over the rows removed: after reverse_module the derived table is already post-overlay, so an effect-based count would say a number on lap 1 and vanish on lap 2, moving a published field between a module and its own round trip. Proved against the real compile → reverse → compile path, not argued.

RM126's sweep gets the discriminator its own comment promised. sweep.compare_module reports carried_added beside actionable_added, read off the after-manifest's carried rather than re-derived from prose. axes["warnings"] deliberately still fires on any movement — narrowing it would make a published axis mean something other than what every record already written claims — and a pre-0.7 manifest reports every addition as actionable, the safe direction, since calling an unrecorded finding carried would tell a reader that something fixable is not. Both new fields join compilation.warnings in EXCLUDED_MANIFEST_FIELDS, because they are derived from it and move exactly when it does.

Repairs rejected, recorded so they are not re-proposed: a plausible code set shipped unattended (the container is free, the vocabulary is a one-way door); codes derived from the pinned phrase catalogue (partial by construction); codes derived from the emission site (a refactor renames a published key); a cap, a truncation or a verbosity flag (all three hide findings, and the author with the most warnings most needs the hidden ones); a new metainfo artifact (the channel already ships and is outside the digest).

Two consequences recorded rather than left to be discovered, both in COMPILER.md beside the catalogue: the code vocabulary is closed, so a consumer pinned to an older just-dna-format refuses a manifest carrying a code added later — the standing cost of every closed vocabulary on a published field here, and "additive" describes the writer; and carried holds full message text, which grows the channel 1.84× across the reference corpus (1.96× on pathogenic_clinvar). The second is the shape the item decided rather than an oversight, and the three cheaper encodings are weighed as RM138 instead of being taken unilaterally. → COMPILER.md § Warning texts a consumer keys on

  • RM124 — overrides.csv, an authored overlay over the derived tables. A correction to a derived row now lives beside the spec rather than inside the file, so resolution.csv, frequencies.csv, gene_metrics.csv, gene_validity.csv, clinical_assertions.csv, literature.csv and gwas_effects.csv become pure build products — derived = f(source, overlay). Columns are table, subject, member, field, operation, value, reason, decided_by, decided_at, with reason required: that is what makes the table a record rather than a knob. Operations are update / insert / suppress, keyed (table, subject, member, field) with one member column whose meaning the named table fixes rather than a per-table key grammar. An empty member on a grouped table is group-scoped for update and refused for the other two. An insert is written as several rows sharing (table, subject, member), one per field, and lands at the end of its subject's group in overlay order, because parquet bytes depend on row order.

What it costs an existing module: nothing. The table is optional, an absent optional table contributes nothing to content_signature, and a module that carries no overlay carries no overrides.parquet either, so no published module's identity moves on either axis. (Its slot in ARTIFACT_PARQUETS is not what protects them: artifact_digest name-sorts the listing before hashing, so tuple position is invisible to the digest.) The overlay is authored input, so when a module carries one it is inside content_signature and byte-hashed into manifest.inputs — editing a correction un-closes a module, which is correct.

What it costs a second pass: also nothing, and that is the point. Merge-not-clobber's behaviour is unchanged — a re-run still gap-fills rather than re-asking every subject — but a recorded row now carries no authored content, so a full re-derivation (rm plus a re-run) is free. Every writer of a covered table says so in its docstring. licensing.csv / sources.csv is deliberately outside the covered set: it has its own merge path and is the one derived table a human is told to write.

reverse_module emits the post-overlay derived tables plus the overlay, so the overlay applies twice and the fixed point is checked by test rather than assumed — it holds because all three operations are idempotent set operations, which is why there is no previous_value column. One consequence stated rather than hidden: no operation reports its own no-op, because after a reverse all three no-ops are true of a healthy module, so a suppress with a typo'd subject does nothing and cannot warn. Two warnings do exist and are pinned in COMPILER § warning texts — an update reaching no row, and an overlay naming a table the module does not carry.

ProvenanceItem.outranks still stands and the duplication is stated in SCHEMAS rather than hidden; the unification is RM135 on the 1.0 tracker, and 0.7 emits no deprecation warning because an author warned off outranks today has nowhere to go.

Cross-repo: just-dna-registry needs overrides.csv added to SPEC_DATA_FILES / RECOGNIZED_SPEC_FILES, or the file is dropped on the next re-publish — the way licensing.csv was lost before their 0.16.2. Recorded in INTEGRATION_0_6.md.

  • RM83 — closed, not shipped. Dissolved by RM124 rather than argued down: with the corrections in the overlay there is nothing inside a sidecar to preserve, so the --refresh command it asked for has no problem left to solve and the drift detection it wanted falls out of an ordinary re-derivation. Its residue shipped as enrich --rederive, in RM128's entry below. See ROADMAP_HISTORY.

  • RM128 — an enrichment run is a transaction, and it holds a lock, reports progress, and can re-derive. (just-dna-enricher; no schema, no manifest field, no vocabulary — a new module, three keyword arguments and two flags.) enrich() persisted nothing until its tail, so a run killed at minute 29 had written zero bytes. The obvious repair, checkpointing the table as it goes, trades away a property somebody may be relying on — that a refused strict run leaves the module exactly as it was — and the trade turned out to be unnecessary.

Durable staging beside the target, plus an atomic commit at the gate. Each live link's answer is written to .<name>.staging/answers.csv beside resolution.csv as it arrives; the table itself is still written once, at the bottom, by a writer that renames into place. Same-directory staging is the correctness condition rather than a convenience: os.replace is atomic only within one filesystem, so staging beside the target makes a cross-device move structurally impossible instead of merely avoided. What is staged is the raw answer, never the assembled row — the hosting filter, the pseudoautosomal selection, locus_index and the minted ids all recompute — so a resumed run reproduces the table an uninterrupted one produces, which is the item's P7 obligation and has a test.

A refused strict run commits nothing, now as a written promise rather than an accident of statement order, asserted on the bytes on disk of a pre-existing table. --keep-staging leaves the staged answers after a successful commit, for debugging; the default removes them. Not mode-conditional: write gates the persistence machinery whole, so it means one thing everywhere.

flock on the spec directory, non-blocking, no lockfile. Two concurrent runs were last-writer-wins over a merge with neither knowing, and a zombie run once overwrote a restored 330-row table with 162 rows that then validated, closed and compiled green. A lockfile left by exactly the kill this item is about would block every subsequent run, and the staleness rule that would fix it is a clock; flock dies with the process. A second run refuses with a pinned message. The degradation is documented, not silent — no fcntl, or a filesystem answering ENOLCK/EOPNOTSUPP, logs that the run is not excluded from a concurrent one. flock is untested here on the network filesystems a consumer may use.

progress: Callable[[int, int], None] | None, reporting (done, total) over SUBJECTS. The incident is an idle timeout — both reported runs died at 1800 s with essentially every variant resolved — so the caller needs a keepalive with monotonic progress. That rules out phases (a 29-minute phase emits nothing) and links (whose total is unknown until resolution finds them). No protocol: two integers, no object, no event vocabulary to keep working forever.

A staged answer is honoured only if the link that produced it would run this time — the seeding reads the same two booleans that gate the live blocks — so a --no-gnomad or --offline resume does not stamp a row a first run with those flags could never have written.

--rederive is RM83's residue, and it composes rather than adding machinery. It re-asks every recorded subject and names the ones that came back different — MODULE_LIFECYCLE § 5.1's canary, finally performable. None means nobody re-derived and [] means nothing moved; only a real difference prints. A recorded subject the sources could not be asked about keeps its rows, or re-deriving would be a way to shorten the table — the reported incident wearing a new flag. The A re-derivation resumes only another re-derivation, because a gap-filling run's staged answers are, after its commit, exactly what produced the recorded table — seeding them would compare the table against its own provenance and report a clean bill for the subjects being re-checked. The honest limit is stated with it: rm plus a re-run re-derives just as correctly and reports nothing, because it destroys the old values before the fresh ones arrive.

  • RM126 — a release now declares what it changed about compiled output. Principle 3, as amended on 2026-08-21, lets a corrected derivation ship in any release but never silently: each release declares its corrections, readable offline and without recompiling. No such channel existed, so the charter named a surface that was not there. This is the one item of the 0.7 round that was owed rather than offered.

just_dna_format.release_records carries a static table of per-release records and a pure needs_recompile(compiled_under, current) over it — pydantic-only work in the tier every consumer already has. Five axes per record (parquet_schema, parquet_bytes, content_signature, manifest_fields, warnings), each tri-state, plus the declared correction-versus-addition split — the half no diff can compute, since only the person who fixed the bug knows whether the stored value was wrong (stats.genes) or merely absent (curator).

Deliberately not a should_rebuild verdict. The same fact costs a registry an immutable PATCH and a local cache a free rebuild, so the decision is the consumer's and only the fact is ours. The per-axis breakdown stays exposed underneath the declared flag.

Intervals compose as a union over (a, b], walked down each record's previous link — linear storage rather than O(releases²), moved-and-moved-back still reads as moved, and, load-bearing rather than incidental, the interval from a version to itself is empty. That last property is what stops an unattended sweep minting a fresh PATCH every run forever, and a field-keyed or latest-known-defect shape would not have it.

Unknown is a state, never an empty result. An interval the installed table does not cover answers None per axis under Kleene semantics, and so does a downgrade. Without it the surface would be worse than nothing, because a consumer would stop recompiling on the strength of a silence.

compilation.warnings gets its own axis, outside the set that drives a recompile. It is a published manifest field and RM131/RM134 both move that channel in 0.7, so folding it into manifest_fields would report a manifest field changed on essentially every module in a catalogue for a reworded message. It cannot join compiled_at in the excluded set either, because a new warning can be a real signal. RM131's carried split is the discriminator that will make the two decidable; the seam is sweep.compare_module, which keeps added and removed warnings apart for it. SpecRow.needs_upgrade is computed over authored row content and a warning never touches one, so no warning change can flip it — verified, not assumed.

The roster ships beside the table (AUTHORED_ROW_DERIVED_FIELDS): which manifest fields are pure functions of the authored rows, so a consumer can recompute the current answer from stored inputs with spec_tables and module_stats instead of consulting the table at all. Its boundary is conditional and the condition rides on the entry: compile_module re-derives stats over the survivors only when the symbolic-allele drop removed something, so a module that lost the sole row naming a gene legitimately disagrees with a recomputation, permanently. compilation.dropped_rows (2026-08-24) is what makes that checkable rather than merely stated.

just-dna-compiler sweep is the instrument, and it exists so this never becomes the hand-kept per-release map everyone agrees it must not be — the defect wearing a public name. It compares two trees of compiled output built from one spec root, and with --release it runs the gate: a measured movement that no record declares fails, so the measurement forces the declaration. The gate belongs to the bump → uv sync → tag sequence rather than to the test suite, because it needs the previous release installed.

Backfilled by measurement, not by memory. Both shipped records were produced on 2026-08-28 by compiling all sixteen reference_examples/ from one spec root under each published 0.6 release. 0.6.0 → 0.6.1 moved nothing — recorded as a measured zero with its denominator, never silence. 0.6.1 → 0.6.6 moved the parquet schema and bytes on 10 of 16 (RM120's curator column growing studies.parquet), a published manifest field on 9 (stats.genes/stats.gene_count on seven, RM121; literature.quotes_unchecked on three, RM119) and content_signature on none. Six moved a published, indexed manifest field with both hashes byte-identical, which is precisely the case a digest comparison, a signature comparison and a revalidate all correctly call no change. Only 0.6.0, 0.6.1 and 0.6.6 reached PyPI in that line, so those are the only intervals a stored artifact can name; everything older stays honestly unknown.

Scope v1 is compiler-derived outputs, said out loud. Enricher-side outputs are unmeasured, which is not the same claim as unchanged. And this coexists with a consumer's own recomputation probes rather than replacing them: we state what a release did, they check what a specific stored artifact says, and the two fail differently.

  • RM70 — a star-allele module can state CPIC's core assumption. (just-dna-format; no compiler change — the parquet schema and the reverse writer both derive their columns from the model.) requires_callable was a VariantRow column, so haplotypes.csv, pharm_variants.csv and diplotypes.csv had no way to record the one thing a consumer most needs before trusting a *1/*1 result: CPIC assumes a position it did not call is reference — literally requires_callable=false — and the assumption lived only in the upstream's prose. The optional tri-state column now sits on HaplotypeRow and PharmVariantRow, the two PGx tables that name a locus, so the claim is about a position the row itself states.

Not on DiplotypeRow, and callable_from did not travel. A diplotype names a star-allele pair, not a locus; the same column there could only mean "the variants defining these two haplotypes were callable", which is a fact about haplotypes.csv restated one table over, free to drift the moment a definition is edited. callable_from stays VariantRow-only because the two are different axes — one says a proof is required, the other says where the proof lives — and a curator can state the first from the source's prose with no basis for the second. Declaring it once in module_spec.yaml was refused for the reason RM36 and RM32 already paid for: the verdict is per locus, and CYP2D6 holds a SNP-defined allele beside a structurally-defined one inside one gene.

reference_examples/cyp2c9_warfarin_grch37 — the module the gap was found against — now populates the column on both tables and exercises all three states. Its haplotypes.csv records CPIC's assumption verbatim (false on both defining SNPs); its pharm_variants.csv is keyed on genotype and so splits: the reference-homozygote rows require a callability proof, the rows naming an alternate allele do not, and the rows whose reference allele the module never resolved are left blank rather than guessed. That answers the consumer ask for requires_callable "populated somewhere real, to try the round trip against" on its PGx half. The module was re-closed, so its attestation binds the new bytes.

Optional with respect to every module published before it: content_signature normalizes with exclude_none, so a spec that never writes the column hashes exactly as it did. artifact.digest does move for any module carrying either table, which P3 says is not by itself a reason to defer.

  • RM133 — the card subtitle gets a home a registry can amend. (just-dna-format; Principle 4 deliberately untouched — the closure binding is exactly what it was.)

Package: just-dna-format. Additive under Principle 3 — two new constants and one new number, no authored field, no schema change, no parquet column, nothing invalidated. Principle 4 is deliberately untouched: the closure binding is exactly what it was.

  • normalize.PRESENTATION_AUTHORITY_KEYS = frozenset({"short_description"}) — a second registry-owned key family beside IDENTITY_AUTHORITY_KEYS, with PRESENTATION_AUTHORITY_REASONS mirroring the identity map. Separate sets because ownership and presentation are different reasons for a key to be registry-owned and a consumer may stamp one without the other; one stripper — strip_authority_keys takes either family or their union, so there is one path and not two.
  • normalize.SHORT_DESCRIPTION_MAX_CHARS = 120 — the card-subtitle calibration, published here rather than guessed at downstream. It refuses nothing: no model carries it as a max_length, Display.description stays unbounded on purpose, and absent the key everything behaves as before.
  • Why this and not a smaller binding. Rewriting module.description moves no content_signature, no artifact.digest and no fact signature, and still drops the closure, because manifest.inputs covers the raw bytes of module_spec.yaml. The binding stays as it is — an attestation answers is this the same document a named person signed off, and a partition along content_signature's line would make a closure transferable across a rename. A short_description field on ModuleInfo is refused for the same reason: every key in the spec is on the un-amendable side.

For just-dna-marketplace and any other publishing registry — this is the integration note. Store short_description in your own record beside the module, never in module_spec.yaml. Import the two constants from just_dna_format.normalize and strip the union before handing a block to our validator: strip_authority_keys(block, IDENTITY_AUTHORITY_KEYS | PRESENTATION_AUTHORITY_KEYS). The stored bytes then never move, so manifest.inputs still matches, verify_manifest(check_inputs=True) still passes and a closure over those bytes still stands — your amend_display endpoint is not gated on us. Read the ceiling off SHORT_DESCRIPTION_MAX_CHARS rather than hardcoding 120; enforcing it is yours to do, since we bound nothing. Nothing is retroactive: the seven published modules met every requirement that existed and are untouched. On the CLI, --strip-identity still means identity alone; the second family is --authority-key short_description.

  • RM71 — the drafted-genotype worklist is re-requestable from the command that produced it. draft-panel leaves genotype as the placeholder, correctly, and reports the alleles the author must write it from in one uncapped line per open row. That report used to be scoped to the rows a single run added, so a second run added nothing and therefore said nothing, and the alleles could not be asked for again. The worklist now covers every row in variants.csv whose genotype is still the placeholder — refused rows were never written and settled rows no longer carry the stub, so the earlier narrowing survives on a better basis — and draft-panel --dry-run obtains it without appending anything, meaning what it means on draft. The state-placeholder line moves with it, because it reads as "these rows also". An open row this run holds no record for has its alleles withheld and is named by gene in one aggregated line rather than guessed at or silently dropped from the count; the line says only that nothing this run selected covers them, since the usual causes (another gene, a tighter --clin-sig/--min-review-stars) are not ones the pass establishes. A scaffolded template row — the placeholder in rsid as well as genotype — is not drafting work and stays out of both lists. No schema, no cell filled, no digest moved. The 0.7 proposal recorded draft-panel as lacking --dry-run; it has had one since 0.5.1, so that half of the item was already true and is now pinned by a test over the whole drafting family rather than asserted.

  • RM134 § A — the PubMind snapshot, and one significance normalizer. (just-dna-enricher; nothing in the format or compiler tiers moves and the compile path imports none of it.)

Package: just-dna-enricher. Additive: a new module, a new snapshot kind and two new commands. Nothing in the format or compiler tiers moves, no schema field is added, removed, promoted or retyped, and the compile path imports none of it. Decided in PROPOSAL_0_7 § RM134; the evidence is PUBMIND_ASSESSMENT.md.

  • clin_sig.normalize_clin_sig is now the one raw-significance → VALID_CLIN_SIG fold, moved out of clinvar_build._normalize_clin_sig, and two defects were fixed on the way out. Its map keys are underscored because that is ClinVar's spelling; PubMind spells the same concepts with spaces, so Uncertain significance and Conflicting both fell through to other while ClinVar's Uncertain_significance and Conflicting_classifications_of_pathogenicity mapped correctly — two sources that agree, reported as disagreeing. Repaired with a whitespace→underscore step in the tokenizer (an identity on every existing key) and a bare conflicting key. No ClinVar answer changes, which is asserted as an equality over the walked map rather than spot-checked. A second hand-written map was the rejected repair: the concordance check RM134 § B builds compares two normalized calls, so a drift between two maps would report a disagreement between our own tables.
  • A whitespace-only or token-less significance value is now not_provided, not other. "The source states no classification" and "the source stated something we do not model" are different answers and only the second is a disagreement. No ClinVar CLNSIG takes that shape, so nothing built to date moves.
  • New: just-dna-enricher pubmind build ([dev], polars) — the ANNOVAR-redistributed hg38_pubmind_db.txt.gz → data/pubmind.parquet + release.json, byte-reproducible across rebuilds. Columns are chrom, start, ref, alt, pvid, clin_sig, clin_sig_raw, pathogenicity_score, confidence, derivation; the significance names are unprefixed, matching the ClinVar snapshot, because one column vocabulary across every snapshot is what lets a later check read N authorities with no per-source mapping. pathogenicity_score is nullable and null means not computed, never 0.0; confidence is PubMind's own 0–3 and is deliberately not normalized against review_stars.
  • Every dropped row is counted, as an equality over a walked registry. input_rows == record_count + sum(dropped.values()) over PUBMIND_DROP_REASONS — off_target_chrom, non_acgt, ref_equals_alt, multi_substitution, identical_duplicate. A codon block differing at exactly one base is decomposed onto it (derivation=codon); one needing two or three simultaneous changes is dropped, because it asserts a change to the protein rather than to a genotypable position. Length-changing rows are kept and stamped derivation=indel: upstream left-normalization is unverified, and a consumer must be able to exclude them without re-deriving why.
  • A bad cell and a bad row are different outcomes. A row with no PVID or an unreadable Start is dropped and counted (no_pvid, unparsable_position) — the PVID is the record id everything here is keyed on, and a null one would have merged distinct records under identical_duplicate as well as crashing the emit sort. A malformed pathogenicity_score or confidence withholds that value and keeps the row, counted in unparsable_score / unparsable_confidence. NaN and inf count as unparsable rather than being stored (float() accepts both), and a non-integral confidence is withheld rather than truncated to a count the source never stated.
  • download_pubmind_table raises PubMindUnavailable, a subclass of PubMindBuildError, so a moved bulk URL surfaces as this tier's type rather than httpx's and one except arm covers both an outage and a malformed table. The subclass makes a caller's except order load-bearing.
  • A contested coordinate keeps every PVID as its own row. Consolidation into a PVID is keyed on the text the model extracted, never on a coordinate, so one variant fragments into records whose verdicts disagree. Collapsing them was rejected: it needs an ordering nobody defined, which is mode() over an unsorted group. release.json records multi_pvid_keys, max_pvids_per_key and contested_keys.
  • New: just-dna-enricher pubmind publish, which refuses and says why. On the PharmVar precedent, with a different reason: PharmVar's bytes arrive under terms that bar passing them on, PubMind's under no stated data terms at all. The command exists rather than being absent, because a missing command reads as an oversight somebody will helpfully add. The refusal text is pinned by a test.
  • PUBMIND_TERMS records None on every licence axis, and the nulls are load-bearing: null means the terms could not be established. Unknown terms warn and never gate — taints_commercial_use requires commercial_use is False — so a module carrying PubMind values compiles, lands pubmind in manifest.sources.unknown_terms_sources, and drives the module-wide verdict to None: undetermined, never permitted. That is also why pubmind build has no --use flag: check_declared_use returns a skip for every declaration, so the gate would refuse every build and the flag would do nothing.
  • New cache: pubmind/, $JUST_DNA_PUBMIND_CACHE, with resolve_pubmind_reference and default_pubmind_cache_dir and deliberately no ensure_pubmind_snapshot — a refused publish means no repo to provision from. cache status lists it as build-your-own.
  • Fixed a counted-prose assertion that a correct addition broke. test_every_resolver_and_default_dir_takes_the_off_switch asserted len(named) == 12 with "expected six resolvers and six default dirs" in its message, so the seventh snapshot failed it on arithmetic. It now asserts an equality between the two walked families keyed by snapshot name, which says something the count never did: no resolver is missing its default directory and none is orphaned.

Still open, and named rather than implied. The ANNOVAR-distributed table's data terms are unestablished and only CHOP can settle them; the unblock action is to ask WGLab and CHOP's Office of Technology Transfer in writing. Whether the indel rows are left-normalized is likewise unestablished, which is what derivation=indel exists to let a consumer act on.

  • RM134 §§ C and D — drafting from PubMind, and the hint. (just-dna-enricher; nothing in the format or compiler tiers moves and the compile path imports none of it.)

Package: just-dna-enricher. Additive: one flag and two options on an existing command, one option on hint variant, a new provider module and a fourth DRAFT_PROJECTIONS entry. No schema field is added, removed, promoted or retyped, and no published module is invalidated. Decided in PROPOSAL_0_7 § RM134; the evidence is PUBMIND_ASSESSMENT.md.

  • New: just-dna-enricher draft-panel --source pubmind — a flag on the existing command rather than a draft-pubmind beside it. The two write the same rows into the same table from the same gene argument, so a twin command would carry a second copy of the genotype worklist, the placeholder guard, the dedup pass and the refusal summary — the four parts that are hard to get right. pubmind_draft therefore imports the ClinVar provider's machinery, _open_stubs included: scoping the worklist to report.added is the once-only defect RM71 removed in this same release, and a second copy of that rule is the one that goes stale. draft-clinpgx stays separate because it writes different tables.
  • The gene map is ClinVar's own per-record attribution, matched at the exact position. PubMind names no gene, so --gene BRCA1 needs a gene→locus map, and this repo deliberately holds none — the compiler's gene/locus check is chromosome-granular for that reason. Every position ClinVar records for the gene is the universe, with no clinical or review filter, because filtering the map would narrow what PubMind is asked about while looking like a filter on PubMind's calls. A min/max span was rejected: it invents a boundary nobody defined and writes a gene cell that is a false claim wherever two genes overlap. Both snapshots are required and each absence names its own switch. The cost is stated rather than counted — a PubMind verdict at a position ClinVar has no record for cannot be reached by gene at all, and that class is not countable, since attributing it to a gene is exactly what there is no map for.
  • Five withheld classes, each named at draft time, and the accounting is an equality over the walked reason set. PUBMIND_WITHHELD_REASONS covers a contested key, a length-changing row, a call outside --clin-sig, a confidence below --min-confidence, and a confidence the source never stated — that last one its own class, because None is not 0 and reading an unstated confidence as 0 invents a reading. candidates == drafted + Σ withheld holds over the set, and a class that withheld nothing reports no zero.
  • Contestation is decided over every PVID at the key, before either dial runs. A --clin-sig or --min-confidence applied first would remove the dissenting record and pick the winner exactly as the mode() this design already refused would; a test constructs a dissenter that either dial alone would have hidden. Where the records agree, one row is written and the record count, the PVIDs and the best stated confidence stay in the transcription.
  • Identity is the whole coordinate or nothing, since the snapshot has no rsID column at all — and the row still matches on the same five columns a ClinVar-drafted row does, so a coordinate both sources speak about is one row rather than two. No clinvar, pathogenic or benign is folded: all three are ClinVar flags by their own field descriptions, and a position appearing in ClinVar's gene map says nothing about whether this allele is in ClinVar. No studies.csv row is drafted — the ANNOVAR channel carries no PMID — and the run says so, because that table is mandatory. A position two requested genes both claim leaves gene empty, counted and named: a joined BRCA1, BRCA2 is not a symbol check-identifiers resolves, and picking one is the gene model the pass went to ClinVar to avoid inventing.
  • New: pubmind in DRAFT_PROJECTIONS, projected onto clin_sig, so a module drafted from PubMind cannot confirm itself when the concordance check reads the same column (@draft-digest). Its identity is the coordinate and not the provider's match_on: the source states no rs-number, so an rs-number an author adds later is a change to the row's spelling, not to the call.
  • Unknown terms warn and never gate, and the drafter does not call the acquisition gate. check_declared_use decides whether a fetch may proceed and skips on unknown terms, which is right for a pass that would go and get such data. Nothing is fetched here — there is deliberately no ensure_pubmind_snapshot, the operator built the snapshot with our own command, and refusing to read it would make that command's output a file nothing may consume. The reason is reported in the source's own words instead, and the licence row the provider writes carries None on every term, which does not taint: taints_commercial_use requires an explicit False. That is § A's finding, re-asserted here over the row this provider actually writes rather than trusted from a fixture.
  • New: just-dna-enricher hint variant --pubmind-cache (§ D) — PubMind's records beside the cell an author is about to fill, never filling it: clin_sig is what the concordance check cross-examines, so a hint supplying it from one of the authorities being compared would make the check agree with the source it is checking (@hint-redundancy-bearing). Three states rather than two: no snapshot is nobody asked and names $JUST_DNA_PUBMIND_CACHE, a snapshot holding nothing at the allele is an absence in their corpus, and disagreeing records are all reported with none picked. It answers for a coordinate the caller typed as well as one an rsID resolved to, because their channel is coordinate-keyed and most of its rows carry no rs-number.
  • A dial belonging to the other authority is named rather than ignored. --min-review-stars and --max-citations under --source pubmind, and --min-confidence under --source clinvar, warn when set away from their default: a run that honoured neither the flag nor the author's expectation is the failure worth reporting before it happens.

  • RM132 — pharm_variants.csv can cite the evidence for its own claim. (just-dna-format + just-dna-compiler + just-dna-enricher; a new optional authored column, additive under Principles 3 and 8, and no published module is invalidated.)

A ClinPGx-drafted module carried 1,482 drug-response rows and had nowhere to ground any of them: sixteen model fields, thirteen authored, none a PMID or DOI. studies.csv cannot close it, and that is structural rather than an oversight — a study row keys on (variant_key, pmid) and attaches to the variant, while a pharm_variants.csv row keys on (variant_key, drug, genotype, phenotype_category, annotation_id), so one study row would attach the paper to every drug, genotype and phenotype category recorded for that variant at once. Widening studies.csv's key was refused for the reason RM47 already refused it: it would make a study row's subject depend on which table read it. evidence_level is not the handle either — it points at somebody else's grading of the evidence rather than at the evidence — and the licence row's source/dataset state redistribution terms rather than grounding a claim.

PharmVariantRow.pmid, optional and free-form under the one grammar spec.validate_pmid_cell owns, so an author who has met StudyRow.pmid or MeasureBinRow.pmid learns nothing new. That is why a full-cost authored column (P9) is taken here rather than deferred for demand: demand fixes an unfixed shape, and RM47 fixed this one a release ago for a structurally identical table.

provenance_quote does not follow, and the release says so rather than leaving it implied. The row cites; studies.csv and literature.csv describe. That is the line that stops StudyRow's whole provenance column set — population, p_value_num, effect_size, provenance_quote, curator — migrating onto a citing row one column at a time, and a body of clinical claims this size is exactly where the question gets asked next.

Both literature cross-check sites learned the site in this release, which is RM47's recorded lesson in its own words: a column shipped without them would make every citation from it read as a stale orphan in one direction and be invisible in the other. _cross_check_literature (with split_cited_literature and _check_quote_counter_is_current beneath it) and the enricher's enrich_literature both read it, so a pharm-grounded citation is checked for existence and identifiers exactly like a study-grounded one — and a module whose only citations are pharm pointers is now enriched rather than refused.

The roster is derived, so there will not be a fourth round of this. _CITING_TABLE_KINDS is every _TABLE_KINDS model declaring a pmid, and the new public load_citing_rows / table_citations walk it — the pair the enricher reads through, so a second copy of the table roster in that tier (the RM40/RM41 shape) is not repeated, and a test walks the enricher's own source with ast to assert none is kept. load_binning_rows / binning_citations stay and stay narrow: a caller asking for the binning kinds is asking about thresholds, not about the citations a module makes. The internal third parameter of split_cited_literature is renamed bin_rows → kind_rows to match what it always held.

One warning text moved, deliberately: literature_row_uncited now reads "no study, bin or pharm row in this module cites". The code is the stable handle and is unchanged; the phrase is pinned by the suite, which is what makes the rewording a deliberate act rather than a drift.

Optionality is proved by running it rather than by citing exclude_none: a spec carrying no pmid header hashes to the same content_signature as the same spec carrying the header with every cell empty, and a filled cell moves it. The round trip is asserted on the pharm_variants.parquet bytes as well as on both identities.

2026-08-24 — twelve consumer items in one pass (S63–S74)

Packages: just-dna-format, just-dna-compiler, just-dna-enricher — a MINOR, cut inside 0.7.0 (tagged v0.7.0 on 2026-09-12); it was deliberately left uncut when written (2026-08-27). The number is not decided, so nothing here names one: the replies' markers read next-minor and every doc dates a change by its Sn rather than by a version that does not exist yet. Answered is not installable, and here it is not even tagged. Most of what follows is patch-class legibility, but three changes are each independently additive and so size the release under P3: VerificationRecord.producer, compilation.dropped_rows, and the new public load_spec. Legality sizes the release and severity only orders the queue inside it, so a pass that is mostly warnings still cuts as a minor when one field is new. Two reporters, twelve items, answered serially. Eight shipped code, four are filed; the counts below are off the tree rather than remembered.

Filed: RM128 (enrich()'s lost work), RM130 (a conflict sidecar), RM131 (warnings structure), RM132 (pharm_variants.csv citations), RM133 (an amendable card subtitle), plus the stats counter retype queued to the 1.0 tracker.

The two findings that cost real work:

  • A sidecar writer that truncates in place leaves a valid short file, and a merge believes it (S66). Nine writers across the enricher and format tiers used open(path, "w") + csv.DictWriter, so a killed process left a table that parses cleanly and is simply shorter. For resolution.csv that is the worst residue available: the next run reads it back, merges on subject, and believes it — and the three branches of enrich() that deliberately write no row for an unanswerable subject make "fewer rows" a state the table reaches honestly. The reported incident had a client-killed run keep going, reach the write, and replace a restored 330-row table with 162 rows, after which the module validated, closed and compiled green. Fixed with layout.atomic_writer/atomic_write_text. Three writers were reported and nine were routed — the guard walks the set with an AST check and asserts an equality, and was watched failing on pre-fix source first.
  • The better-resolved module was the loud one (S67). _verify_vrs_ids emitted one warning per allele where _vrs_coverage aggregates the same fact, and which path an allele took was decided by whether the enricher happened to mint an id for it. Noise ran inversely to how well-resolved a module was: 80 of a 101-row module's 85 warnings, with the three findings its author could act on at positions 83–85, against a 57,595-row module producing one line. The governing rule was already in the file — a finding no authored edit could clear is not a strict matter — spent on severity alone.

Also shipped: field descriptions on the three ModuleInfo fields that had none, plus a guard asserting every authored field across 28 models carries one (S63). VerificationRecord.producer, so a merge stops restamping records it did not produce, with the document-level field's own description corrected because it was the false claim (S71/RM129). The panel: deprecation gated on its replacement existing and no longer claiming "nothing else is lost" — both halves, since gating alone leaves the false clause standing (S69). A detail on the clin-sig record and a warning when any check reports findings; nothing read VerificationRecord.findings at all (S70). row_count, table_rows documented, and a composite-gene warning (S72). compilation.dropped_rows, closing the residue a consumer's recomputation guard could not see (S65). Public load_spec, ending a private-symbol reach the enricher itself was making (S74).

Two answers are the deliverable and shipped no code. S64 asked us to justify the attestation binding or split it along content_signature's line — the split is refused because that line excludes name, version and namespace, so a binding drawn there makes a closure transferable across a rename; and the route that actually unblocks the registry is registry-owned metadata (IDENTITY_AUTHORITY_KEYS' shape), which means their endpoint was never gated on us. S73 asked which provenance model pharm_variants.csv was meant to use, and the tree had answered it a release earlier under a rule nobody had stated generally: a row cites when its claim is finer-grained than studies.csv' key.

A 26-finding dogfooding note had arrived in ROADMAP.md rather than the inbox, where the ledger cannot see it. Eight of its findings had Sn twins; the other eight are dispositioned in place — one (D17) did not reproduce, one (D11) is our own documented policy, three went to the idea-book with the cheap half separated from the design half, and one is cross-repo.

Suite 2881 → 2916 (+8 skipped), green throughout; ruff check clean.

2026-08-21 — the Constitution says rules only, and gained three (RM127 closed, RM126 → 0.7)

Documentation only; no package changed. S62's finding was that a corrected derivation — an existing published field whose derivation changes, so the same spec yields a different value — is in none of the three sizing rows, and that the safety argument used to clear it (content_signature is unchanged, measured) is incapable of failing, because stats sits outside the signature by design. A check that cannot fail read as a pass (@tautology-zero), one level up from where RM123 caught the same shape the same week.

Principle 3 gains two rules. Release class and artifact staleness are different axes — a corrected derivation is a bug fix, so it may ship in any release, since deferring it to a minor means knowingly serving a wrong value meanwhile; what it may not do is ship silently. And authored identity is not the sizing test, which disarms the clause that sized RM121.

The charter is now rules only, by its own header item, and came out 11.5% smaller (17,239 → 15,263 bytes) while gaining three rules. Reasoning, evidence, open questions, superseded states and rhetoric are banned from it, along with any outward reference beyond a published version a rule turns on — the old phrasing said the file "points to no other document" as a description and the 0.4.1 plan drifted in anyway, so it is now an instruction. The four existing amendment entries moved verbatim to the new CONSTITUTION_AMENDMENTS_HISTORY.md, which may cite RMn and consumer reports freely.

Principle 9 exists because the audit caught a rule about to be deleted. The cost-by-layer pricing — parquet approximately free, derived CSV half, authored schema full — was stated only inside an amendment entry, and is cited by CLAUDE.md's coding standards. Moving the amendments out wholesale would have silently removed it, so it was promoted to a numbered principle instead.

RM127 is closed — filed, rewritten and answered inside one pass. RM126 moves to 0.7 as owed rather than deferred: the amendment obliges a release to declare its corrections, and that channel does not exist yet, so the charter currently names a surface that is not there. Two axes — output_differs measured by a previous-tag sweep, correction-versus-addition declared — with a gate that fails a release whose measured change carries no declaration.

2026-08-21 — a patch changed a parquet schema, and nothing could say so (S62, RM126, RM127)

Nothing shipped; two items filed. A new consumer, just-dna-registry, adopted 0.6.1 → 0.6.6 and ran the catalog sweep whose job is to find published artifacts that should be recompiled. It correctly reported nothing to do, at all three layers, while manifest.stats.genes — which feeds their gene facet — sat stale on every star-allele and copy-number module they publish.

Measured, and wider than the report. All sixteen reference_examples/ compiled under v0.6.1 in a detached worktree and again under 0.6.6, spec inputs byte-identical across the interval, which is entirely patch releases: 16/16 changed a published manifest field, 10/16 moved artifact.digest, 0/16 moved content_signature. The digest movement is studies.parquet +257 bytes on each of the ten — RM120's curator column, first present in v0.6.5. So the parquet schema moved across a patch interval, which the reporter had not seen; they reported changed manifest fields.

Authored identity held on all sixteen, which is the charter working: an unset optional column is omitted from content_signature. The sharpest number is the one that took a recount: six of the sixteen changed a published, indexed manifest field with both hashes byte-identical — apoe_epsilon went genes: [] → ["APOE"] at the same artifact.digest and the same content_signature. (stats.genes moved on seven modules; an earlier draft of this entry said eight.) That is precisely why nothing can see this — a digest comparison, a signature comparison and revalidate are each correct to report no change while an indexed field goes stale. RM126 is the missing third axis, asked for as an interval-keyed declaration with the axes separated, explicitly not a should_rebuild verdict, and with unknown-interval as a state rather than an empty result. The guard it needs is a measurement rather than a hand-kept map, and the sweep above is its prototype.

RM127 was filed and rewritten the same day, and the withdrawal is the useful part. It first read the release table and the practice disagree, indicting curator's patch. Withdrawn: curator is additive, no published module can carry it, no stored value became wrong, so a patch is defensible and the table is merely strict. The defect is RM121 alone, and its change class is not in the taxonomy — an existing published field whose derivation was corrected. Why it read as safe is the keeper: the only test applied was does authored identity move, and stats sits outside content_signature by design, so that test cannot fail there. A tautology read as a pass (@tautology-zero), one level up from where RM123 caught the same shape that week. The structural edge — a field outside identity is one nothing can see move, so the cheapest changes have no detection channel: six of sixteen examples changed a published indexed field with both hashes byte-identical. And since a corrected derivation is a bug fix, deferring it to a minor means serving a wrong value meanwhile — so the release number cannot carry staleness at all, and the two axes separate rather than reconcile.

Not an instance, and separated deliberately: RM106's warning de-duplication. The release table already sizes a warning or a count as patch-level legibility, so compilation.warnings was never promised stability across a patch. It is the argument for the reporter's axis decomposition — warning text is patch-legal and a column is not, so one "did the output change" bit would have been useless.

Scope of the negative: the sweep is an offline compile over sixteen specs. Enricher-side outputs were not compared across the interval, so verification.json and the documents RM123 touched are unmeasured rather than unchanged.

2026-08-21 — the triage loop's threshold counter went blind for the second time

Documentation only; no package changed. The triage loop § 4 counters grep ROADMAP.md for open items to decide when to ask the user whether the next minor should start. The 2026-08-21 decision round rewrote all four open items' status lines to lead with what had just been decided, which pushed the release-class phrase out of the slot the grep reads and wrapped one of them across a line. The count read zero with four grounded items in the file — the direction that matters, because a zero reads as an all-clear.

This is the counter's second failure: the first counted a literal **0.6** and kept counting it after 0.6 shipped, fixed by counting the idiom instead. The idiom then drifted. The durable half of the repair is that the release-class token is now named as a fixed field in ROADMAP § Active items — verbatim, on one line, with everything else the status wants to say after it — so the rule sits where a status line is written rather than only in the file that reads it. The incident is in CONSUMER_TRIAGE_LOOP § 6. A phrase a tool greps is an API whether or not the tool belongs to a consumer (@warning-text-is-api).

Also: the CHANGELOG's (latest) marker had stayed on the second entry when the decision round prepended a new one, against this file's own newest-first rule. Moved — and then it went stale a second time within the same pass, when this entry was prepended above the one that had just been corrected. That is twice in two prepends. The marker is a derived fact — "the topmost entry in a newest-first file" — restated by hand, which is the exact pattern the RM104–RM111 batch named as the thing worth carrying out of it, and the file's preamble already says Newest first. Surfaced rather than acted on, because retiring a convention across every heading is a bigger call than this pass: the next person to trip on it has the evidence to delete the marker outright.

2026-08-21 — just-dna-lite's consumer-side changes from the just-module-creator hand-off

Consumer-side only; nothing in this repo changed. Recorded because the working agreement asks for cross-repo integration changes, and because the agents most affected are the hand-off's own authors — they asked just-dna-lite for ~60 reads and this is what the first tranche did. Their document lives at just-dna-lite/docs/CONSUMER_HANDOFF_from_just-module-creator.md; the reply, with a verdict per item, is at just-dna-lite/docs/reviews/consumer-handoff-triage.md.

A correction to what a consumer believed about the published corpus. Both that repo's CLAUDE.md and the hand-off assumed no module on HuggingFace publishes a manifest.json — CLAUDE.md said "every module on HuggingFace today" and the hand-off called the logo fallback "dead code in production" on the same premise. Measured 2026-08-21 against just-dna-seq/annotators: all ten modules publish one. Attestation (INTEGRATION_0_6 § 2.8) is therefore the normal discovery path there and probing is the exception. Checked for the failure that implies — a file present at the path but absent from artifact.files is now dropped where it used to be probed and found — and the attested set matches what is present on all ten, so no side table was lost.

RM43's status was stale downstream by two releases. just-dna-lite's docs still said the pharm_variants coordinate fill "waits on RM43", and two comments in its annotation engine asserted that the compiler applies resolution.csv to weights.parquet alone. Verified against installed compiler 0.6.1 (compiler.py:499) and corrected. Consequence worth knowing if you maintain a similar consumer: classify a lead table by the values it holds, not by its family name. Both generations are live — that repo's shipped pharmgkb is a 0.5 artifact measuring 0 of 1482 rows placed, while a 0.6 recompile of the same spec qualifies for a position join — so a value probe routes both correctly and no positional_rows gate is needed in the join path.

A phase asymmetry that is a consumer bug, not a format one, and is now reported rather than silent. A VCF reader that sorts a genotype (as that repo's does) cannot match an authored genotype held in homolog order, in either ordering. Sorting the module side is not the fix — it folds A|G and G|A into one key and manufactures a match the module never stated — so the module side still never sorts and the unmatchable rows are now counted and logged before normalization strips the | that reveals them. Nothing here needs to change; noted so another consumer does not "fix" it by sorting.

Two consumer-side reads that now use this repo's own constants rather than restating them. README_CANDIDATES reached that repo's HuggingFace publisher allowlist, which had omitted it — so manifest.readme attested a file the upload never sent, and verify_manifest(check_readme=True) passed anyway because absent is not a failure there. And discovery now keeps identity.version, artifact.digest and the weighting block from a remote manifest it was already fetching and validating; before, every remotely discovered module reported no provenance at all. All three stay tri-state.

Filed separately in ROADMAP.md: manifest.stats.genes is derived from variants.csv alone, so a module whose gene is stated only in a PGx or binning table publishes gene_count: 0 and cannot be found by gene. Originally measured by just-module-creator; relayed because neither consumer owns variant_stats.


Older entries are in two archives, both cut at a release boundary rather than a date. Entries that shipped in a 0.6 number are in CHANGELOG_0_6.md, split out on 2026-09-12 at the v0.6.6 tag — four entries dated 2026-08-21 shipped in 0.7.0 and stay above, which is why the cut was read off git show and not off the dates. Entries dated 2026-08-11 and earlier — the whole 0.5 line and everything before it — are in CHANGELOG_PRE_0_6.md. Neither split is a change of format: each archive is this same document continued downward.