Skip to content

just-dna-format — Roadmap

Forward-only, and now active-only. Every item here is open work. What shipped moved to ROADMAP_HISTORY.md with its rationale intact, so this file answers one question: what is left to do, and how bad is it?

  • RM_TOC.md — the complete index of every RMn, active and shipped, with the document that defines each and every document that mentions it. Start there if you are looking for an item.
  • ROADMAP_HISTORY.md — the shipped items, plus the 0.4.1 and 0.5.0 release narratives.
  • CHANGELOG.md — what shipped in each release, newest first.
  • USE_CASES.md — where most RMn were derived (the what-blocks? lens); PROPOSAL_0_5.md — where their shape was argued.

Split on 2026-08-13, and it changes where to look. This file is now the line being built; a deferral is filed against the release that will decide it:

  • PROPOSAL_0_6.md — the authoritative entry for every active item below. Each was argued to a decision on 2026-08-13, with the facts probed, the repairs rejected and why, and the consequences that follow without being chosen. Where an entry below and the proposal disagree, the entry below is stale — several of these sections were written before the decision and describe a shape that was rejected. The proposal also carries RM53–RM67, from VCF_4_4_AUDIT.md, which are not repeated here.
  • PROPOSAL_0_6_PT2.md — the second 0.6 design round, decided 2026-08-16. Everything filed behind the first round accumulated in the minor-deferral file because that was the next release at the time; PT2 re-asked which release each belonged to now that 0.6 is uncut, and took five back: RM55's fix, RM72, RM82, RM84 and RM87. It is authoritative for those five, and for three of them the probe overturned what the roadmap entry says.
  • ROADMAP_0_9.md — items legal in a minor, each waiting on a design question, a corpus or a caller. Its predecessor ROADMAP_0_8.md, 0.8's scope, closed at the 0.8.0 cut on 2026-11-05 into history/, and its one waiting item, RM28's predicate half, moved here. What follows describes the 0.8 file as it was. 0.8 is a stabilization and competitor-parity release, decided 2026-09-11 with the maintainer; the theme is stated in that file's header and does not itself admit an item. RM10 closed there, folded into RM28. It succeeded ROADMAP_0_7.md at the 0.7 cut — that round is closed in history/, which keeps four of the five records the line above names; only RM84 is still waiting and still here. Read the membership rule in its header, not a list here: these two bullets carried per-item enumerations until 2026-08-27 and both had gone stale, which is what let RM69 sit in the 0.7 file gated on a 1.0 item without anyone noticing. RM_TOC.md is the complete list, and it is the only one.
  • ROADMAP_1_0.md — items that need a major, the one that is release-blocking for it, and (since 2026-08-27) items gated on a major without needing one. Plus the upgrade ledger. The 1.0 cleanup tracker below did not move and stays the home for the unnumbered major-only items.

Code comments citing "ROADMAP item N" / "ROADMAP 0.3 item 5b" are historical breadcrumbs — follow them to CHANGELOG.md / COMPILER.md.

Status: 0.8.0 is cut, tagged and published — all three packages at 0.8.0, on 2026-11-05. The tag is v0.8.0 at e97339a3, and PyPI has the three releases as of 19:56 UTC that day; what a consumer checks and changes is INTEGRATION_0_8.md, and the cut's own entry is in CHANGELOG.md. The previous cut was v0.7.3 (2026-09-27: format 0.7.1, compiler 0.7.2, enricher 0.7.3), the last of the 0.7 patches after v0.7.1 (2026-09-21) and v0.7.2 (2026-09-25).

0.7.0 was cut, tagged and published on 2026-09-12, all three packages. The tag is v0.7.0 at 2001215 and it moved through four positions before settling there, each because a gate found something rather than because anyone changed their mind; the sequence and the reason for each move are in INTEGRATION_0_7 § 5. PyPI has all six artifacts as of 15:49–15:50 UTC that day, and their hashes are the ones that section lists. The previous cut was 0.6.6, v0.6.6 (2026-08-21), carrying nine patch fixes: the 2026-08-19 doc-audit round (RM104–RM107, RM109, RM111), the two shipped items of the S57–S60 batch (RM121, RM123), and S61's lookup fix (RM125). Everything committed after that tag — the 2026-08-21 output-contract, decision and lookup rounds' successors, the 2026-08-24 consumer round, the source-adoption round and the whole 0.7 build round — went out in 0.7.0 rather than in any 0.6 number.

Re-read this paragraph whenever a version moves, because it has lied three times. It read "0.6.4 is the current line" for two releases, then "0.7.0 is bumped and not tagged" for as long as it took somebody to run git rev-parse v0.7.0, and then "0.7.0 is cut, tagged and published" as the current status through three 0.7 patches and the 0.8.0 cut — the same failure the Active items heading had, and the reason each is called out rather than quietly corrected. schema_version stays "1.0" and has since 0.4. Tagged is not published — they are separate steps and the second is the maintainer's, so check CHANGELOG.md before promising a field to anyone; for 0.8.0 both have happened. 0.5.0 went to PyPI on 2026-08-07, with just-dna-enricher 0.5.0 the first release of that package.

0.6.0 was cut and tagged v0.6.0 on 2026-08-17, and 0.6.1 followed on 2026-08-18 with RM93–RM100 (see ROADMAP_HISTORY § 0.6.1). This paragraph read "open on the 0.6 branch, unreleased" for a release and a half after that stopped being true, which is the same failure the Active items heading had — a status line nobody re-reads is a status line that lies. What 0.6.0 landed: manifest.readme (S25), manifest.derived (S26), then RM44, RM51 and RM49 — the last two shipped together, because "the same table in two possible places" is one problem whether the two places differ by name or by directory, and it wants one resolver and one collision rule. Every reference example kept all four of its signatures across that batch.

Then the design round itself was built, in eleven parallel lanes plus a charter amendment that went first and alone: RM4, RM5, RM24, RM25, RM27, RM43, RM45, RM46, RM47, RM48, RM50, RM59 and the VCF 4.4 cluster RM53–RM65 — the whole of PROPOSAL_0_6.md's build list, with the rationale in ROADMAP_HISTORY § 0.6.0. Across the corpus that batch moved content_signature on exactly two modules (mt_heteroplasmy and htt_repeat_expansion, both re-authored deliberately because their VCF pointers named the wrong field), artifact.digest on seven, and it gained a resolution_signature on the four table-only modules that never had one. The suite went 1535 → 2046. schema_version is still "1.0" and cutting the release is the user's call.

Shipped since: just-dna-enricher + just-dna-compiler 0.5.1 and 0.5.2 — 0.5.1 was RM38 (a cache for every licence-gated source, so a hosted enricher never reaches one live per request) plus RM39–RM42 from a consumer field report; 0.5.2 is the panel-scale batch behind S3–S6 in CONSUMER_SUGGESTIONS.md — the quadratic DuckDB probe that stopped a gene panel finishing, a clin_sig cross-check that no longer reports a structurally guaranteed zero, a drafted genotype on the contigs where only one is expressible, and the .env-ordering bug behind three separate "the cache is right there" reports (see CHANGELOG.md). 0.5.3 answers S9 the same way: it does not widen resolution to the 0.4 families (that is RM43, whose prerequisite column is 0.6 work since the charter amendment) but makes the scope legible — per positional table, how many rows a VCF cannot join and how many of those resolution.csv could place — and adds heteroplasmy.csv to the enricher's subject list so that family can be resolved at all. The three packages version independently, so the network tier took a patch while just-dna-format stayed at 0.5.0; RM41 is the one item that also touches the compiler, which is why that package moved too. None of it touches a parquet, a model or a manifest field, which is why none of it is in the 0.6 table below. That table is format/compiler schema work, and enricher-only work sits outside it entirely; do not read "additive work is 0.6" as covering the network tier.

The "digest window" is retired — the charter was amended on 2026-08-11 and the sort below follows the amended rule. What gates a change is what it does to a reader, not to a recompile's bytes:

  • A new optional column is additive and lands in a minor. An unset optional column is omitted from content_signature and the per-input hashes cover authored bytes nothing rewrote, so the authored identity does not move; only a recompile's artifact.digest does, and Principle 4 already scopes that to a fixed compiler_version. Measured, not assumed — see CLAUDE.md for the numbers.
  • A new optional table is additive too, and more cheaply: file_entries skips missing files, so a module that does not carry the new parquet keeps even its artifact.digest byte for byte.
  • Major-only is removal, promotion to required, and retyping — the moves that break an existing reader or invalidate published data.

The practical effect is that several items below were deferred on a rule that no longer holds, and have been re-sorted on their own merits instead; where an item stays deferred, the reason is now stated and is never "it moves the digest".

Spent while the pre-publication window was open, recorded so it is not re-litigated: the VRS identity switch (RM19) and the cofactor columns (RM29) — the two that actually needed it — plus ResolutionRow.authority (provenance, so no signature moved), the continuous-bin semantics (mt_heteroplasmy re-authored), indel reconciliation (shox_par1 gained a resolved coordinate), and pseudoautosomal locus selection (shox_par1 halved, 20 rows to 10) — RM33, RM35, RM31 and RM32 respectively. Of the last four only RM31 and RM32 moved an artifact.digest; none moved a content_signature, which is pre-resolution by definition.

The 2026-08-06 readiness audit spent none of it, which is worth recording because the findings were severe: it fixed a Principle 7 break where compile → reverse → compile relabelled a non-GRCh38 module's assembly and re-minted its identity key, three further build-confusions in the enricher (including a frequency pass that would have fetched a different variant's counts), and a validate that passed modules compile refused (see CHANGELOG.md). Every fix is confined to behaviour that was only reachable off GRCh38 or through the error channel, so all ten pre-existing reference examples keep their exact artifact.digest, content_signature and resolution_signature — verified by comparing before and after, not assumed. The one addition is a new example (reference_examples/grch37_build/), and a new module cannot move an existing digest.

Each entry below is ## RMn — name, a metadata line (severity, status, owner, motivating case), then the detail. Severity is how much it costs to do, not how urgent.

0.6 — what a minor permits

Kept as the record of a settled question. Every ✅ below that named a 0.6 item has since been built — see ROADMAP_HISTORY § 0.6.0 — so this table now answers why each was legal in a minor, which is the part worth not re-deriving. The rows that stay open (RM23, RM16, RM28, RM15) are the ones deferred to a later line.

Sorted by the amended rule: what a change does to a reader, not to a recompile's bytes. Nothing here is gated on the digest any more, so every ❌ or ⚠ below carries a reason of its own — a design question, a corpus question, or a genuine break:

Item Shape Minor-legal now?
RM23 predictions.csv new optional table ✅
RM24 gene_validity.csv new optional table ✅
RM25 ClinVar assertion tier new optional table ✅
RM16 authored PRS weights new optional table ✅
RM28 meta-conclusions new optional table + injected cofactors ✅ — parked on the corpus, not on the window
RM5 symbolic / structural alleles widens a grammar ✅ — P3 bars tightening, not widening
RM27 redistribution gate a gate over a column that already ships ✅ — reads sources.csv, writes no parquet
RM4 gene-panel materialization compiler behaviour, opt-in per spec ✅ — row-set expansion pinned on compiler_version; only a module that declares a panel gains rows
RM10 inheritance expectation a column, its own table, or yaml metadata ✅ — all three placements are minor-legal now; pick on orthogonality (P5), not on cost
RM43 resolve the 0.4 families a stamped-identity column per positional table, then the join ✅ — the column is additive; what is left is the design round, not a version gate
~~RM44 resolution_subjects count~~ one additive integer on Compilation ✅ shipped in 0.6.0 — see ROADMAP_HISTORY
~~RM51 licensing.csv alias~~ a second accepted spelling of an input filename ✅ shipped in 0.6.0, old spelling deprecated, removal queued for 1.0
RM50 PMID↔PMCID a diagnosis (no schema change) + one optional id column ✅ for the guard, which is an enricher patch; ⚠ for the column — additive, but it wants deciding beside the 1.0 requiredness demotion
RM15 multi-build identity changes the semantics of variant_key and of every coordinate ❌ — 1.0, and not for digest reasons: re-keying published identity is the identity-change class
~~RM38 gated-source cache~~ enricher-only: new builders + cache resolvers, no parquet touched ✅ shipped in just-dna-enricher 0.5.1 — never a 0.6 item; kept here so the reason an enricher change bypasses this table stays visible

Two consequences worth stating outright:

  • RM10's gate dissolved. It was parked partly because "where it lands" decided whether it was a minor or a major. Every placement — a column on an existing table, its own optional table, or module_spec.yaml metadata reaching only the manifest — is minor-legal under the amended rule, so the placement is now a pure design question: which one keeps the axes orthogonal (P5). Decide it on merit.
  • The 1.0 pile split when the rule changed, and the two items that used to sit together show why. weights.parquet's end is an additive optional column, so it is 0.6 work now, gated on the one thing that was always its real blocker: the coordinate convention a second coordinate needs (interbase-half-open vs inclusive), which RM15 must settle. Removing the dead likely_pathogenic/likely_benign pair stays major, because removal is exactly what the amended rule reserves for a major. Same tracker, opposite answers, and the reason is now legible.

Active items

No count here: the open items are the ## RMn sections below, and test_triage_tools.py walks them. This line stopped stating a count on 2026-09-30, the day it was caught stale again (it read "four as of 2026-08-21" for two rounds after it stopped being four, then not one of them is a decision through the three that are, then three for the hour it took a fourth to be filed, then two until RM151 shipped, then one naming RM152, then one naming RM153, then none, then seven for the 2026-09-01 source-adoption round, then one, none again on 2026-09-11 when RM164 moved to the 0.8 file, one again on 2026-09-12 when RM232 was filed, none again the same day when it shipped, and one again that evening when RM235 was filed, then none on 2026-09-20 once RM235 and RM244-RM246 had all shipped, then one again the same day when RM247 was filed, none once it shipped in 0.7.1 — this line went on naming it until 2026-09-24 — and one again on 2026-09-24 when RM258 was filed, then two on 2026-09-25 when RM263 was filed, and three the same day when RM264 split from RM256, then two when RM264 shipped on the rebuilt patch line, and it went on saying two, RM258 and RM263 while the 2026-09-27 postmortem sweep and the rounds after it filed some thirty more, and after those two had moved to the 0.8 file. That is why the paragraph under it says to count off the sections rather than off this sentence, and why the sentence no longer offers one).

RM232 was filed open and shipped in the same session, and the filing is the part worth keeping. It was written into this file the moment it was confirmed rather than when its fix was approved, because until then it existed only as two named exemptions in a passing test — a state this file cannot see, and a release reading no open RMs, all green therefore cannot either. Whether to fix a confirmed defect now is a question; whether to file it is not.

The last one was RM164, and it is filed against 0.8 now. The 2026-09-01 source-adoption round (RM163–RM168) was asked for as a batch — what else should we adopt as enrichment sources, besides CIViC and PubMind — and filed together because its items shared a shape and a discipline rather than a mechanism. Five of the six shipped inside the uncut 0.7.0, and RM164 parked on a measured negative: no source publishes a heteroplasmy level per tissue, so the table kind stays without one. It stayed open rather than closed because a source would reopen it, and its own entry names in advance the finding that would close it instead. It sat in this file for ten days after the decision that sent it to 0.8, which is the failure mode the split exists to prevent — a decided deferral that no release file can see — so on 2026-09-11 it moved to ROADMAP_0_8.md. A decision to park is not filed until the section has moved.

The spin-off RM164 filed shipped on 2026-09-03. RM171 adopted MITOMAP's curated mtDNA tables as the increment they carry over the ClinVar cache — the variants half of the source RM164 probed for its heteroplasmy half and did not find. ROADMAP_HISTORY has it. The two are deliberately separate items: widening one by changing what it is about is how an item stops meaning anything, and RM164's negative would have been kept artificially alive by a positive that has nothing to do with it.

The discipline the round is worth remembering for, since it is the reason these entries read the way they do: not one of the six had been probed when it was filed, every entry said so in its own words, and each separated what was measured from what was merely candidate while asserting no licence. The failure it was written against is a remembered licence or a remembered API hardening into a permanent false constraint in these files (@probe-names-the-table), so each item's first step was the probe — and five of the six entries turned out to say something their own probe contradicted.

The candidates deliberately not filed are not restated here: the predictor tier is RM23, authored PRS weights are RM16, and Google Scholar, OpenAlex/Unpaywall fulltext and an offline gnomAD frequency snapshot are dispositioned in the 0.5 idea-book below. Academic-only sources (OMIM, dbNSFP) and subscription-gated ones (HGMD) are out on terms; callers (PharmCAT, Cyrius, PyPGx) are out on scope, which is Annotating core, not format scope further down.

RM152 and RM153 both arrived and both shipped on 2026-08-31, which is the sequence worth keeping. RM152 stood here carrying no release class, on the stated grounds that both its candidate adoptions had been refuted and an item with no repair has none to state — the correct state for it. What changed that was a measurement, not an argument: the probe it had named was run, both refutations held, and a third route nobody had proposed turned out to be buildable with no schema change. RM153 was its residue, and it answered its own two questions in opposite directions — the ClinGen CAID pass taken, liftover refused with its ceiling measured at 13 evidence rows. Both entries are in ROADMAP_HISTORY; the measurements are in CIVIC_SURVEY and CIVIC_UNRESOLVED. RM151, filed on 2026-08-31 the same day RM117's other half shipped, was built the same day. Both items the RM124 wave-1 audit filed — RM136 and RM137 — shipped on 2026-08-31, and so did RM117's observability half and RM146; RM138 was closed the same day with its numbers measured. RM110, RM103's manifest half and RM108 were three more settled ones and all three shipped on 2026-08-31.

Count them off the sections, not off the sentence: this line said three for as long as it took to notice that a narrowed item is still an item, not one of them is a decision for as long as it took three decisions to be filed beneath it, and two while one of the two was already built — the same arithmetic failure recorded two paragraphs down, three times more.

The release-class token in a status line is a fixed field, not prose — a tool reads it. An open item's **Status** reads open — and then, immediately and in bold, a minor, release undecided (or a patch, …), because the triage loop's threshold counters grep for exactly that to decide when to ask whether the next minor should start. An item saying the same thing in other words is invisible to them, and a counter seeing nothing reads as an all-clear. So write the token verbatim, keep it on one line so a line-based grep sees it whole, and put whatever else the status needs — narrowed, shape decided, the manifest half only — after its closing **. The 2026-08-21 decision round rewrote all four lines below and took the count from six to zero while four items sat here; that is the second failure of this counter, the first having been that it counted a version number that then shipped. The third was an item filed with no such token at all — RM235, on 2026-09-12 — so the field is now asserted by a test rather than only described: schema/tests/test_triage_tools.py walks every ## RMn between this heading and # Not format scope and refuses the one the counters cannot see. Write the token, or the suite says which item is invisible and to whom.

The 2026-08-21 decision round is what emptied the other half. Six items stood here, every one of them a decision rather than a missing line of code, and one pass answered all six: RM102 closed outright, RM122 parked on demand and moved to the minor-deferral file (ROADMAP_0_8.md since the 0.7 cut), RM117 narrowed to the half that costs nobody a decision, and RM103 split — the manifest half stays here, the refusal is a tightening and moved to § The 1.0 cleanup. The reasoning is in ROADMAP_HISTORY § the 2026-08-21 decision round.

Two of the six turned out not to be design-blocked at all, and that is the finding worth keeping. RM110's encoding was already pinned by a test on one producer — test_gnomad.py asserts an empty flag list is None, not "" — so nothing was undecided; it was parked because normalizing the other producer moves a fact signature and the round that found it was a patch round. That is a release-class reason wearing a design label, and it held the item for two days of looking harder than it was. RM102's was the mirror image: the item read as a security question and the record held one lost hour and no incident, which is a disposition nobody had asked for out loud. Ask what a filed decision would actually change before treating it as one — a status line saying undecided is not evidence that anything is.

The six patches filed beside them shipped on 2026-08-20 (RM104–RM107, RM109, RM111) and moved to ROADMAP_HISTORY § the 2026-08-19 doc-audit patch round.

RM88 and RM93–RM100 all shipped in 0.6.1 and moved to ROADMAP_HISTORY § 0.6.1, with their rationale and with the five places the eight filings turned out to understate what was there. An empty list here is not an all-clear — it means nothing is filed, and this file has read empty twice before while carrying real work one heading down. The live consumer inbox (CONSUMER_SUGGESTIONS.md) is the other half of that question, and ROADMAP_0_9.md / ROADMAP_1_0.md hold the deferred items. Every one of them broke a rule this repo had already written down — which is the finding worth keeping out of the item entries and stating once: the gotcha book is not the thing that catches a regression. In four of the eight the file carrying the violation also carried the rule, sometimes in an adjacent comment. So the durable half of each is a test, and in six of the nine that test walks a registry rather than a list.

RM88 and RM89 were both filed 2026-08-17 out of the 0.6 PT2 batch's lane D, both found by building RM84 rather than by planning it — neither a defect in what shipped, and neither a blocker for it. Both are closed now, and RM88's close is worth one line here because the shape recurs: the code half was always cheap and the entry had mispriced it, so what actually held the item for a release was an undecided policy wearing a technical objection. RM89 closed the same week: the consumer's answer arrived as S35 the day after it was filed, the open question it was waiting on was the only thing holding it, and building the answer found the defect underneath it — see ROADMAP_HISTORY. RM74–RM79, the whole 0.6 dogfooding fix round, shipped on 2026-08-15 and moved to ROADMAP_HISTORY.md — but read that as the sprouts are repaired, and the ground with them: RM76's narrow repair is what shipped in that round, and the question it asks from underneath — RM73, the root several of these grew from — closed on 2026-08-16, both halves. What remains of it is the promotion: making a closure a precondition of compiling is major-only and is filed, with its own blocker, in ROADMAP_1_0.md. Everything that was open on the 0.6 branch before that round was built in the 0.6 batch and moved to ROADMAP_HISTORY.md with its rationale; what was deferred moved to the roadmap of the release that will decide it — ROADMAP_0_8.md (RM16, RM23, RM28, the deferred halves of RM55, RM56, RM65, plus RM66 and RM67, and the dogfooding items RM68–RM72) and ROADMAP_1_0.md (RM15, RM52, RM55's removal half). That sentence records the 2026-08-13 split and is not a current inventory — RM69 has since moved to the 1.0 file, and five of the 0.7 entries shipped in the PT2 round. Follow RM_TOC.md for where an item lives today.

This section read "None in this file" for a day after the six were filed, because they were appended below the Not format scope heading and nothing moved the boundary — so the roadmap's own summary line said it had no open work while carrying two high-severity items. Recorded rather than quietly fixed: it is the RM_TOC.md failure mode (an item nobody can find) arriving in the file the index points at, and it is why a new item starts here as a ## RMn section and gets its RM_TOC row in the same commit.

It happened a second time, to RM160, and the repair is the same one line. Filed 2026-09-01 as open — worth doing with owner enricher, it was appended below the same heading and sat under Not format scope — a section whose own intro says it lists things "so they are not mistaken for format scope" — until the 2026-09-01 source-adoption round moved the boundary back down to RM7. Twice is a pattern rather than a slip: appending an item to this file puts it wherever the last heading left you, so check which # heading you are under before writing the section, not after.

The trackers further down are the other live part of this file: the reserved-namespace tracker and the 1.0-cleanup candidate tracker, which the Constitution deliberately keeps out of itself.

RM285 — the MITOMAP lane's two open remainders: VUS* withheld, and the unrated-miss identity rows

Severity low · Status open — a patch (either answer is drafter behaviour), two decisions first; the anchoring pass shipped as RM293 · Owner enricher (mitomap_miss_build, mitomap_draft) · Motivating case RM171's "Still open, and none of it blocking" paragraph and PROPOSAL_0_7_PT3 § What a first cut still owes, never filed, found by the 2026-09-27 postmortem sweep (P12) · related RM171, RM273, RM293

RM171 listed what it left, and nothing carried the list. Checked on 2026-09-27 against the local mitomap_miss snapshot (47 unmintable rows, 388 unrated misses).

  1. Closed by RM293 (2026-09-27, v0.7.3). The : deletions (refna="TA" against regna=":") are anchored on the vendored rCRS base at position - 1 in mitomap_miss_build, checked against rCRS and joined through RM273's event key: 39 of the 47 unmintable rows left that bucket, and the 8 that stay are ref_equals_alt and non_nucleotide. RM293 was filed as RM273's residual and did not name this entry; the two now point at each other.
  2. VUS* is withheld rather than understood. A MITOMAP legend, or McCormick 2020 read in full, decides whether it maps to a clin_sig member. Until then the withhold is right, and this is a reading task, not code.
  3. The unrated misses are an identity increment with no mappable class. They are counted and never drafted. Whether an identity with no classification earns a drafted row at all is the smaller call RM171 deferred.

Indel normalization, RM171's fourth remainder, is RM273 (1). The MITOMAP snapshot has been on HuggingFace since 2026-09-03 (just-dna-seq/mitomap), so RM171's publish remainder is already closed.

RM311 — a gene_validity.csv correction the source has since adopted, or superseded, says nothing

Severity low · Status open — a patch if both arms reuse existing codes (a new one would make it a minor), a decision first · Owner schema (gene_validity, overrides) + compiler · Motivating case RM290's refuted premise, 2026-09-29 · related RM290, RM108, RM117

When ClinGen adopts an author's corrected classification, it publishes a newer curation in the same currency group, and the corrected row stays, marked SUPERSEDED (classify_currency deletes nothing). The overlay update still matches that old row, so no finding fires. The author is never told the source caught up, and never told their correction now targets a superseded curation.

Two arms, both a value comparison against the group's CURRENT row. If the overlay's value equals the current curation's, the author was vindicated (overlay_answer_vindicated's meaning). If it differs, the correction patches history while a newer curation says something else (gene_validity_superseded is the nearest existing code). A group whose currency is undecidable withholds, as classify_currency already does. What to decide: whether the second arm is the existing superseded finding or a new code, which sizes this item.

RM309 — a risk/protective state whose own conclusion negates it goes unflagged

Severity low · Status open — a patch (a hint), parked on a precision measurement · Owner compiler (conclusion.py) · Motivating case RM279's second rule, 2026-09-28, never measured · related RM279, RM308

D14 proposed two rules, and RM279 built the first. The second is a row whose state is risk or protective under a conclusion that negates it: thrombophilia rs1799889 G/G carries state: risk under "…is not increased". A negation reader is noisier than a genotype token, since "not associated" on a neutral row is correct, and "does not reduce risk" can support either state. Parked on the same gate RM279 had, a hand-judged precision over the reporter's six curated modules plus the registry corpus. The gate is satisfiable today, because both the corpus and the rule exist. Like RM279 it is patch-class only as a hint, and its compile half would join RM308.

RM312 — test_a_wrapped_diagnostic_is_matchable_through_cli_text passes or fails on the length of tmp_path

Severity low · Status open — a patch (test only) · Owner enricher tests (test_cli_rendering.py) · Motivating case the 0.8 post-merge suite, 2026-09-29 · related RM233

Reproduced on main and on 0.8 alike. The test's last assertion says the raw CLI output does not contain "nothing will fetch it", because Typer's error box wraps the sentence. Whether it wraps depends on how long the missing file's path is, and that path is tmp_path. Under /tmp/pytest-of-mau/pytest-9/… the phrase fits on one line and the test fails. Under a longer basetemp it wraps and the test passes. Its docstring says this half "was never environment-dependent"; it is, through the path length. The cli_text half, which is the behaviour, holds either way.

The patch. Pin the path length. Build the missing path from a fixed long name, or feed a path whose length makes the box wrap inside the phrase by construction. Drop the negative assertion if it cannot be made deterministic, since the helper's reason is recorded in the docstring.

RM321 — VariantRow.flags calls its reserved tags "the tags the tooling acts on", and no tier acts on any of them

Severity low · Status open — a patch for the description; deciding what a reserved tag should do is a separate question · Owner format (spec.RESERVED_FLAGS, the flags description) · Motivating case the longevitymap v2 requirements (just-dna-lite docs/longevitymap_v2/longevitymap_v2_requirements.md §4.2, §6.2, §10), read against the schema 2026-11-03 · related RM28, @field-description-is-a-claim

For longevitymap v2 (2026-11-03): v2's plan sets pleiotropic and conditional expecting behaviour. See the group.

What was confirmed. The flags description says "Reserved tags the tooling acts on: conditional|phased|pleiotropic", and the comment above RESERVED_FLAGS says the same. A grep over schema/src, compiler/src and enricher/src finds exactly one reader. It is the compiler's INFO line, and all it does is leave reserved tags out of the "non-reserved flags in use" list. The phased boolean in weights.parquet comes from a | in genotype, never from the tag. Nothing reads conditional or pleiotropic.

Why it matters now. The longevitymap v2 plan sets pleiotropic on every pleiotropy row and uses conditional as its interim for KL-VS-in-APOE4 buffering. Both choices assume the tooling does something with the tag. It does nothing, so for a reader the tag is a label and the meaning stays in the conclusion prose. An author who trusts the description believes some behaviour is attached, and none is.

The patch. Make the description and the comment say what is true: the tags are reserved names, held so a later release can give them behaviour, and today they only stay out of the INFO list. Pin that with a test that walks the three tiers' sources for readers of each reserved tag. Then a reader added later forces someone to edit the description.

Not in scope. Giving conditional a meaning. That is RM28's predicate, and RM28 is parked on a corpus. The longevitymap KL-VS×APOE4 case is a candidate corpus entry there, not a reason to build it here.

Residuals none yet.

RM344 — a .duckdb Ensembl cache written by duckdb 2.0 crashes a 1.x enricher instead of falling back to the parquet beside it

Severity low · Status open — a patch, unattended (about 15 minutes) · Owner enricher (resolver.py) · Motivating case the liaison's adoption map for duckdb 2, item 2, docs/UPSTREAM_DUCKDB.md (2026-11-04) · related RM343

Measured. duckdb 1.5.5 refuses a database file written by 2.0: ... with version number 999, but we can only read versions between 64 and 68. resolver.py opens the deprecated prebuilt ensembl_variations.duckdb with duckdb.connect(..., read_only=True) at both of its sites and has no handler. An empty or stale db already falls back to the parquet data/ beside it, but a db written by a newer duckdb raises from connect. A cache built on a machine with duckdb 2 therefore stops a 1.x enricher that reads it, although the parquet next to it would answer.

The patch. In the cache-directory case, a db this duckdb cannot open falls back to the parquet data/, the same way an empty one does, and says so in a warning that names the file and the storage version. A .duckdb file passed directly has no parquet to fall back to, so it raises EnsemblReferenceError naming the version and suggesting the parquet cache. The path is deprecated and retired at 1.0. The test writes a db the running duckdb cannot read rather than mocking connect.

Keyed on the type, not the number (the pm, 2026-11-17, from the lead). duckdb 1.5.6 (locked at 830962e9) refuses the same 2.0 file as "version number 69" where 1.5.5 said 999, with the same IOException. So the handler catches the IOException that connect raises and never matches the version text; the warning and the EnsemblReferenceError quote duckdb's message whole instead of parsing a number out of it, so "names the storage version" above means as duckdb words it. The test asserts the fallback and that the message is carried, never a version number.

Residuals none yet.

RM388 — placement readers bypass usable_loci

Parked, the pm (2026-11-05), under the maintainer's rule for audit findings: "for once per known universe scenario defer to 'when roadmap is empty/someone steps on this defect'". Unparks when a real resolution.csv that reaches compile carries a row off the module's build or a not_found row with coordinates (the enricher writes neither; only a hand-written or multi-build table does), or when the roadmap is empty, or when a scope the maintainer highlights touches it.

Severity low · Status open — a minor, parked · Owner compiler (allele_checks.py, positional.py, resolution.read_resolution) · Motivating case audit round 2, class 3 (the lead's scratchpad audit2/report.md (audit round 2 of 0.8, 2026-11-05)) · related RM362

The gap. RM362 made resolution.usable_loci the one rule for which loci an identity sits at: the module's build, and placements only. allele_checks.py (~:60-65, :102-105, :405-409) and positional.py (~:239-251) read loci without it and filter neither build nor not_found, and read_resolution does not filter by build.

What it would be. Every placement reader goes through usable_loci, with an AST guard that no reader of resolution loci bypasses it.

Residuals none yet.

RM390 — the enricher never states a table-level finding the compile acts on

Parked, the pm (2026-11-05), under the maintainer's rule for audit findings: "for once per known universe scenario defer to 'when roadmap is empty/someone steps on this defect'". Unparks when an author reports an enrich that passed and a compile --strict that then refused on the same table. It is not a defect: compile still refuses correctly, and the enricher would only say so earlier. If it unparks, design it after RM387, which settles what RM374 compares, or when the roadmap is empty, or when a scope the maintainer highlights touches it.

Severity low · Status open — a minor, parked · Owner enricher (enrich/orchestration.py, verification_records.py) · Motivating case audit round 2, class 5 (the lead's scratchpad audit2/report.md (audit round 2 of 0.8, 2026-11-05)) · related RM374, RM373, RM387

The gap. RM374's rsid_placements_disagree runs only in the compiler (resolution_step.py ~:167), so an enrich run that writes a mixed or disagreeing resolution.csv says nothing, and the author learns of it only when compile refuses. RM373's mixed-release reason likewise appears only in the ran arm of verification_records.py (~:433-438).

What it would be. The enricher runs the same predicate the compiler does on the table it is about to commit, and warns with the same code and text. RM373's withheld-release reason appears in every arm. Design it after RM387 settles what RM374 compares.

Residuals none yet.

RM396 — six paced clients keep their default interval or their gate private, so a host sharing one gate reads a private name

Severity low · Status open — a minor, parked (the pm, 2026-11-05; triage proposed 0.9). The reporter's interim, reading pgs._REQUEST_INTERVAL, works, and only one host shares one client's pace today. Unparks when a host needs a second client's pace or clingen_allele's, when someone renames a private name the registry reads, or when the roadmap is empty. Until then, keep pgs._REQUEST_INTERVAL and _gate as they are · Owner enricher (pgs.py, litvar.py, civic_api.py, clingen_allele.py, pharmvar.py, atlas_client.py, ENRICHER.md) · Motivating case S131 in CONSUMER_SUGGESTIONS_HISTORY.md (just-dna-registry's liaison seat, 2026-11-05) · related RM203, RM156, RM384

The gap, read on origin/0.8 at 64c835b9 (the same on main). A host that answers many checks holds one PacingGate per upstream and builds a fresh client around it per check, the shape gate= and the S15 sharing contract invite. Building that gate needs the client's interval, and reading the gate back (gate.spent, RM203) needs the gate. The dataclass clients answer both publicly: a gate field, and an interval as a public constant or a settings field. Six do not:

Client Default interval Gate kept as Takes gate=
pgs.PgsCatalogClient _REQUEST_INTERVAL (private) _gate yes
litvar _REQUEST_INTERVAL (private) _gate yes
civic_api _REQUEST_INTERVAL (private) _gate yes
clingen_allele _REQUEST_INTERVAL (private) _gate no: no host can share its pace at all
pharmvar PHARMVAR_MIN_INTERVAL _gate yes
atlas_client.AtlasClient ATLAS_REQUEST_INTERVAL _gate yes

ENRICHER.md's RM203 paragraph sends a meter to client._gate.spent and names five of the six; atlas_client is missing, which is the hand-kept-list failure (@registry-completeness).

What it would be. Every paced client takes gate=, keeps it as a public gate, and names its default interval publicly (PGS_REQUEST_INTERVAL and the like, beside the existing public ones). The private names stay as aliases, since a downstream reads pgs._REQUEST_INTERVAL today (P3 does not bind a private name, but nothing is gained by breaking it). The reporter's (a) and (b) together, over the six rather than PGS alone. A test walks every PacingGate( construction in the enricher (AST) and asserts its client exposes gate and its interval under a public name, so a seventh client cannot reopen it; the ENRICHER.md sentence about _gate goes.

Sizing. New public names in the enricher are new surface: a minor. P9 prices it near free (no authored or published field moves). About an hour with the walking test. Not asked, and not proposed: a PGS field in LookupClients, whose bundle is shared while the PGS score cache is per run.

RM261 — strip comments from the enricher wheel (measured, deferred)

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when the wheel's size becomes a reported problem (an install that hits a size limit), or the packaging is reworked for another reason, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · before parking: open — deferred by the maintainer on 2026-09-25 as risky ("dangerous, defer") · Owner enricher (packaging) · Motivating case the 0.7.2 enricher wheel is 893 KB zipped, all of it Python source

Measured on just_dna_enricher-0.7.2. Of 2.58 MB of .py, docstrings are 31%, comments 19%, and code plus whitespace 51%. Removing the comments alone takes the zipped .py from 847 KB to 664 KB (−22%). Docstrings stay in any version of this item: they are the help() text and the API pages.

Why it is not simply done. hatch_build.py would have to leave the original .py out of the wheel and force-include stripped copies. artifacts / force_include collide on the same path (see that file's own note). The copies must keep line numbers, with a comment-only line becoming an empty line, or a traceback from an installed wheel points at the wrong source line. Editable installs must not be touched. It reaches the enricher only: format and compiler are on uv_build, which has no hooks, and covering them means changing their build backend.

Considered and refused in the same round: taking the 15 *_build.py snapshot builders out of the wheel (112 KB zipped). scripts/rebuild-caches.sh documents pip install 'just-dna-enricher[dev]' then cache rebuild as a supported deployment, and caches.py imports every builder at module level. The maintainer kept them in.

RM248 — a declarative report schema: research, not a design

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when a consumer (just-dna-lite, the registry) asks for a report a module cannot declare today, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity — (unsized, deliberately) · Status open — a minor, parked — a research item, asked by the maintainer 2026-09-20. Nothing here is a proposal yet; the exit is a measurement and a charter ruling, not a draft schema. The class is the most additive outcome this could have, which is a new optional table or spec block; two of its three exits ship nothing at all · Owner unassigned — the placement question is half the research · Pairs with RM28, RM188, and RM7 below (which is not the same item — see What this is not)

For longevitymap v2 (2026-11-03): v2's §5 is a report specification: a context card, grouping by mechanism, a research view. See the group.

The ask, verbatim in intent: some kind of report schema — Jinja templates or something — where the contents are static and declarative and all the fancy CSS lives downstream.

Why it is filed as research rather than as a design

Because the charter has something to say first, and it does not obviously say yes. The Non-goals are explicit: "No UI and no gene–disease inference. The format catalogs curated annotations that consumers join against variant data; interpretation and presentation belong to those consumers." A report schema that describes a rendered page is that non-goal by another name.

The maintainer's own framing is also the resolution worth testing: a report is a declaration of CONTENTS, not of appearance. What a report says, in what order, keyed to which annotation rows is arguably still annotation; how it looks is unambiguously downstream. If that line holds, this is legal and the non-goal is untouched. If it does not — if a contents schema cannot be written without smuggling in ordering, emphasis or layout — then the honest answer is that this belongs to a consumer and the entry closes. That ruling is the first deliverable.

Principle 1 decides the other half before any syntax is chosen. OakVar's reporter modules are Turing-complete, and that is exactly the shape P1 rejects: code in the compile path, no byte-reproducibility, a runtime every consumer must embed. Jinja is not obviously safer — it has loops, filters and arbitrary attribute access, so "a template language" is a spectrum and the sanctioned escapes are named: a non-Turing-complete boolean predicate and declarative pattern grammars. Any candidate has to be placed on that spectrum explicitly rather than adopted because it is familiar.

What this is not

Not RM7. That entry is a per-sample evaluation output — a measurement, and therefore a consumer contract by the data-agnostic north star. This one is about the structure of a report over annotation content, with no sample in it. The two are adjacent enough to be confused and are listed together on purpose.

Not a fourth library, on today's evidence. The maintainer's read is that it looks like overkill, and nothing here contradicts that yet: a declarative contents schema is a schema, which is just-dna-format's job, and a renderer is a consumer's. A fourth tier would only be justified if the research finds a real body of shared rendering logic that is neither schema nor consumer — and finding that out is part of the item.

The competitor evidence, which is the reason this is worth research at all

This pairs with RM188 because the surveys already ran and the reports are where every competitor's declarative layer visibly ends:

  • ClawBio — pharmgx-reporter is a 2,327-line Python file holding its variant tables inline (PGX_SNPS, GENE_DEFS, GUIDELINES: 32 variants, 13 genes, 59 drugs), and clinical-variant-reporter and cnv-acmg-classifier are planned. Their reports are good and their data is a module in a shape nothing validates. See probes/CLAWBIO_SURVEY.md.
  • genomi — a runtime rather than a format, with a curated thirteen-record catalogue. See probes/GENOMI_SURVEY.md.
  • OakVar — reporter modules, Turing-complete. The counter-example, and the one this repo's P1 already has an answer to.
  • SelfDecode and the consumer-genomics vendors — the reports are the product. Not surveyed here, and a survey of what their reports contain (as opposed to how they look) is cheap and is probably the highest-value first measurement this item can take.
  • just-dna-lite already has a built-in reporting system. It is the reference consumer, so what it does today is the baseline any schema has to beat — and if it needs nothing from us, that is an answer rather than a gap.

Why it pairs with RM28

RM28's surviving half is the predicate — a claim keyed on more than one subject, which no brick holds. Its second corpus entry is exactly a report finding: ClawBio's pharmgx-reporter renders one AVOID across 59 drugs, warfarin, and reaches it through "special": "warfarin" — a hardcoded branch calling get_warfarin_rec(profiles). The report is where the missing predicate surfaces: a section that has to say "CYP2C9 and VKORC1 together" is RM28's gap wearing a consumer's clothes, and a report schema that cannot express it would be shipping the same hardcoded branch one layer up. So the two items constrain each other: RM28's answer bounds what a report section can be keyed on, and the report is the place where an unexpressible pairing is most visible.

There is already a presentation surface, and it is small and deliberately bounded

Worth knowing before anything is designed, because it sets the precedent this item either follows or breaks. module_spec.yaml carries report_title; normalize.py splits identity keys from presentation keys with a stated reason per key (PRESENTATION_AUTHORITY_KEYS, today just short_description); and RECOMMENDED_* vocabularies exist for colours and icons. So the format already admits some presentation, as named keys with per-key justifications and a storing authority. The question this item should answer is whether a report schema is the same thing at a larger scale, or a different kind of thing that happens to rhyme.

What would close this

One of three, and the research is choosing which:

  1. Dissolved — just-dna-lite's existing reporting needs nothing from the format, and the contents/appearance line cannot be drawn cleanly. The entry closes with the reasoning recorded.
  2. Closed additively — a bounded, declarative contents schema (a table kind, or a spec block) that passes P1 and the human-authorable ⇔ machine-precise gate, with a motivating report from a real consumer. Then it is an ordinary minor-legal addition.
  3. Parked with a sharper reason — the shape is real but blocked on RM28's predicate, which is the most likely outcome if the warfarin case turns out to be representative rather than singular.

The cheapest first measurement, and the one to take before any syntax is discussed: read what just-dna-lite actually renders today, and list every field it needs that a module does not carry. That is a list, it is finite, and it decides between (1) and (2) without a design round.

RM181 — a byte digest that moves beside intact signatures says something changed and not what, and provenance has no shift tracker

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when a consumer reports an artifact.digest that moved and that the signatures could not explain, or the 1.0 review of the hash family opens, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · before parking: open — filed 2026-09-03 for the 0.8 review of what the hash family covers, and a candidate for the 1.0 one if it turns out to want a manifest field per domain · Owner format · Motivating case the maintainer's decision on S87 (RM180), in CONSUMER_SUGGESTIONS_HISTORY.md

The observation. With RM180, an author rewording an overlay reason produces: content_signature unchanged, every fact signature unchanged, resolution_signature unchanged, artifact.digest moved. Read from outside, that says something changed and nothing more — the byte digest is a canary, not a locator. Before RM180 the same edit moved content_signature too, which was wrong for the opposite reason: a provenance edit read as a content one. Either way there is no hash whose movement means the provenance moved, and the maintainer's words for the gap were that metadata has no dedicated shift tracker.

The shape named: digest by domain. One identity per concern — content (have), facts per sidecar (have), bytes (have), and a provenance or metadata one (do not have) — so a consumer holding two manifests can say which domain moved by diffing the hash family rather than diffing parquets. That is a separation of concerns, not a new axis on an existing hash, and it is why this is not a repair to RM180.

Why it waits. Three questions before it is an item. What the provenance domain contains — the overlay's three cells only, or also sources.csv's fetched_at, verification.json's checked_at, the README bytes and module_spec.yaml's display half, each of which is outside some hash today for its own reason. Whether it is a manifest field (additive, minor-legal) or a member of the *_signature family, whose roster rule in SCHEMAS is one per derived sidecar — and this is not a sidecar. And whether manifest.inputs already answers it: the raw-bytes entry for overrides.csv moves on a reason edit, so per file the answer exists, and what may be missing is the reading rather than a hash. RM126's release record answers the neighbouring question — what a release changed about compiled output — not this one, what an edit changed about a module.

What would close it. A consumer asking what moved and getting the wrong answer from the family as it stands; or the 0.8 review deciding the family is complete and this becomes a FAQ entry.

RM188 — the competitor survey — run Calwbio's and genomi's pipelines, read their reports, and re-fold the logic into module mechanics

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when the maintainer names competitor parity as a minor's zone, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked · before parking: open — round 1's two surveys filed 2026-09-13 and their findings routed; round 2's roster is below, searched for rather than named · Owner format (the survey), then whichever tier the findings land in · Motivating case the maintainer's direction, not a consumer report

For longevitymap v2 (2026-11-03): its axes 2, 3 and 5 are now numbered, as RM328, RM330 and RM324. See the group.

Progress. Both named competitors are now probed, one document each, and neither turned out to be a competing format: probes/GENOMI_SURVEY.md (genomi is a runtime; its thirteen-record curated catalogue is a variants.csv in Python, and §8 ranks seven annotation gaps, all priced as derived sidecars) and probes/CLAWBIO_SURVEY.md (ClawBio is 97 agent skills — 68 of them planned — whose genotype-interpreting half carries four hand-curated variant tables inlined as Python dicts and JSON). The two surveys overlap in almost nothing, which is itself the result: one reaches for pathway, target–disease and regulatory content we do not carry, the other for the per-variant clinical axes an ACMG engine reads.

The ClawBio half's findings are filed, 2026-09-13 — a probe records and the tracker allocates, so the numbers came from .claude/rm-next.py rather than the probe: RM236 (consequence/impact, and RM23's grain question with them), RM237 (region-keyed ClinGen dosage), RM238 (per-tissue eQTL, open but parked on a consumer), RM239 (fine-mapping posterior), RM240 (land the translated nutrition panel in the corpus) and RM241 (weighting does not survive reverse). The sixth finding was not filed as a new item because it is not one: warfarin's multi-gene call is RM28's surviving pairing across subjects clause, and it went in there as that entry's second corpus entry. The genomi half's seven ranked candidates are recorded in its own § 8 and are not filed here.

What it is. A survey of the consumer-genomics competitors, with two named first — Calwbio and genomi — done the way PUBMIND_ASSESSMENT was done and not the way a feature comparison is: run their pipelines end to end on real input, obtain the reports they produce, and read the logic back out of the reports — which annotations they join, at what grain, which rules turn a genotype into a sentence, and where a number comes from. Then re-fold what survives into the module mechanics here: a table kind, a bounded rule, a vocabulary member, or a use case in USE_CASES.md that names the gap.

What it is not. Not a marketing comparison and not a licence to copy a rule: a competitor's inference is evidence about what a report needs, and the charter's non-goals still hold — no gene–disease inference in the format, annotation tables only. Where a competitor's logic is an inference, the outcome here is the table that would let a consumer make it, never the inference.

Method, so it can be repeated. One document per competitor under docs/probes/, the shape of PUBMIND_ASSESSMENT: what was run, on what input, what came back, what it competes with, what it complements, and the adoption design if any. Real data, the tool turned on its own output (@adversarial-role, @probe-uniform-corpus), and every claim about a competitor pinned to an artifact obtained rather than a page read — the ClinPGx rounds showed a documented surface and the real one disagreeing.

Exit. Each competitor's probe filed, and its findings either dissolved (already enabled), closed additively (an RM with a motivating report in hand), or parked with the reason — the USE_CASES → PROPOSAL loop, entered at the top.

Round 2 — the candidate roster, built 2026-09-13. Round 1's two competitors were named by the maintainer; round 2's were searched for, and the search is the first half of the result. Method: GitHub repository search over eight query shapes (personal genome interpretation, 23andMe raw data, annotation modules, MCP + bioinformatics, nutrigenomics, pharmacogenomics, ACMG, PGS), sorted both by stars and by recency, plus awesome-genetics (last pushed 2024-03) for the pre-agent generation. Twenty-two repositories were then skimmed — not surveyed — against one question: does it carry hand-curated per-variant or per-gene rows a human wrote and committed, or does it fetch public sources at query time? That is the discriminator round 1 established: both genomi and ClawBio turned out to be the second kind, and the curated fraction of each was tiny.

Re-derive the search with:

gh api -X GET search/repositories -f q='topic:personal-genomics' -f sort=updated -f per_page=15 \
  --jq '.items[] | "\(.full_name)\t\(.stargazers_count)\t\(.pushed_at[:7])\t\(.description)"'

varying q over the shapes above. Nothing found in the 2026 crop has more than ten stars, which is itself a finding: the local-first personal-genomics field is a long tail of weekend projects, and the established tools are older and narrower. Do not read the roster below as a competitive landscape — it is a list of places to go looking for annotation shapes.

The four that get a full survey

# Target Pinned Class What its survey asks
1 OakVar d4e8090df8, 48★, licence NOASSERTION a competing module format with a store A fresh peer read, not a migration post-mortem — the maintainer's framing: an OakVar module is in essence a plugin, arbitrary code, so it is a module plus half an annotator. Its manifest (<name>.yml) declares type (annotator / postaggregator / reporter), level, requires (other modules), input_columns, and typed output_columns — a column contract plus a dependency graph, where ours is a row schema with neither. Ask what the column contract buys, what requires expresses that no module here can, and what is left of a module once the arbitrary code is removed.
2 SNPedia, via snappy 2d5255f, 52★, BSD-2-Clause; SNPedia content CC BY-NC-SA 3.0 US the corpus round 1 never had The largest curated variant-trait corpus in existence: 106,603 SNP entries frozen into data/snps.json from a MediaWiki XML dump, plus genotypes.json (per-genotype magnitude / good-bad / summary) and genosets.json. Survey the corpus, not the SPA — snappy is the extraction route that proves a static dump works, where OSGenome (146★) only crawls live. Its terms are CC BY-NC-SA 3.0 US, established off three primary pages 2026-09-13, which makes it adoptable rather than not — as an unsellable, share-alike, attributed module of its own. See USE_CASES § 2e, now the maintainer's named use case for this format.
3 BioMCP bb3a1d4ad7, 630★, MIT the agent-era tier done at scale Thirteen times genomi's stars and the same architectural class. Ask the one question genomi could not answer at its size: when an agent tool surface is the product, what does it end up needing to say about a variant that a table does not? If the answer is "nothing", that closes the whole class and round 3 can skip it.
4 Exomiser 98f4e0b6f2, 265★, AGPL-3.0 phenotype-driven prioritization The only established tool in the roster that ships its annotation as a versioned data bundle rather than fetching it — the closest existing thing to a compiled artifact. Ask what its bundle contains, how it is versioned, and how HPO term sets sit in it, given that HPO ships no route here for licence reasons.

The three small ones worth a read after those

Target Pinned Why
dosedna 530cfe7, 2★, MIT Hand-curated PGx over six genes, and the closest small analogue of our haplotypes/diplotypes/pharm_variants trio — plus a committed, provenance-stamped CPIC snapshot (allele_definition 39, diplotype_phenotype 666, recommendations 1,180).
dna-engine 5791bb3, 0★, Apache-2.0 403 curated markers with the richest per-marker schema found anywhere: tier, effect, transferability, chips, per-genotype label/impact/summary/detail/actions, and citations as (pmid, note) pairs. Also a deliberate non-interpretive posture — it refuses to translate genotype into phenotype, and says why.
allelix 4b56bbe, 30★, AGPL-3.0 Carries no corpus, and is here anyway: 38 ADRs documenting source-precedence and suppression rules (PharmGKB non-finding suppression, somatic-on-germline suppression, a GWAS odds-ratio magnitude modifier). Competing-source arbitration is authority_precedence's problem, and this is the only project found that wrote its reasoning down.
Custom-Personal-Genome-Interpretation 15fe28a, 0★, MIT Two authored tables: 98 rsIDs carrying GRCh37 and GRCh38 coordinates side by side on one row plus population-split MAF, and 540 PRS weights transcribed out of a paper's supplementary table — see RM16.

What the skim already found — five axes, each sighted independently more than once

This is the part that did not need a survey. Convergence across unrelated projects is the @probe-uniform-corpus heuristic firing: when three people who have never met each add the same field, the field is answering a real question.

  1. Which genotyping arrays can call this variant — three sightings: dosedna's array_callable + free-text coverage_note per gene, dna-engine's chips per marker, and MrOrtiz/dna-annotator's measured-vs-imputed tiering per source array (it refuses to let an imputed call outvote a measured one). We have nothing. requires_callable / callable_from / min_quality ask whether the consumer's own VCF saw the position; this asks which platforms in the world interrogate it, which is a fact about the variant and therefore annotation. Note the per-gene / per-variant scope split, which is the shape that already cost 39 variants once.

A fourth sighting, and the first with a source and a measurement: S111 (just-module-creator, 2026-09-24). just-prs already holds the Illumina GSA v3 backbone that 23andMe v5, AncestryDNA v2, MyHeritage and FTDNA v2 share, as typed positions on both builds (648,379), plus a 1000G LD-proxy table. Measured over 16 real modules' resolution.csv: 336 of 965 positions typed (35%). The GWAS modules came out near 22% and the PGx panels at 51%, and lifting to GRCh37 lost one position. The report asks for a module-level expected match rate. Recorded here rather than as an RMn because it is this axis with a candidate source, and three things a design would have to settle came up in the reply (CONSUMER_SUGGESTIONS_HISTORY § S111): (a) the fact is per variant (which platforms interrogate this position), and the module figure is a manifest summary derived from it, the way the literature block summarizes literature.csv. The per-variant form is also what the reporter's other use, picking the typed one of two lead SNPs in LD, actually needs. (b) Counts, never a rate, with typed, LD-proxyable and not position-matchable kept apart. The proxy count describes a substitution the consumer's engine may or may not make, so it stays beside typed and is never added to it. Position-only goes in the field's name or description, since a position match says nothing about allele or strand. (c) The source is a snapshot lane, never an import of just-prs: the per-chip position sets published on the HF org with a release.json, walked through CACHE_SURFACE.md's checklist, and the GSA manifest's licence probed before it is republished (@probe-the-real-file, @no-named-licence). 2. A per-variant claim's ancestry transferability — three sightings: dna-engine's transferability, Bluefinee/kaiseki's Japanese-cohort reconciliation, drhudsonandrade/OmniGenis recording discovery-cohort ancestry per GWAS association. We carry PgsRow.training_ancestry and GwasEffectRow.ancestry, both scoped to their own table; there is no way to say this variants.csv row was established in one population. Numbered 2026-11-03 as RM328, for longevitymap v2. 3. Effect modified by a non-genetic factor — two sightings, and genomi is the third: sinhaankur/open-genome-atlas splits each marker's evidence by kind (diet / lifestyle / geo), each axis independently cited; drdaviddelorenzo/nutrigenomics scopes its weight to a nutrient_domain; genomi's own caveat "folate fortification status of the population modifies effect size" is the same fact in prose. CopyNumberRow's modifier_gene / modifier_cn is the precedent shape for a genetic modifier and there is no environmental one. Note that the second sighting also lands on @weight-has-no-unit: their weight at least names the domain it is a weight in. Numbered 2026-11-03 as RM330, for longevitymap v2. 4. Two sources disagreeing, as a recorded verdict — three sightings: kaiseki's DISCORDANCE_RATIO, which refuses to publish a consensus frequency when cohorts disagree; Gunshipz/genomine's cross-tool confidence/disagreement layer; allelix's ADRs. clin_sig_concordance.csv does exactly this for clinical significance and only for clinical significance — kaiseki does it for allele frequency, where frequencies.csv has per-population rows and no verdict. 5. A hand-assigned salience separate from clinical severity — two sightings: alexlaverty/dna-health-report's magnitude 0–6 per genotype, and SNPedia's own m field carried through snappy. VariantRow.priority is the candidate analogue; whether "priority level override" means the same thing is a question for the survey, not an assumption. Numbered 2026-11-03 as RM324, for longevitymap v2.

A sixth shape was seen and is already decided against, recorded so it is not re-raised. dianguan0105/Custom-Personal-Genome-Interpretation stores GRCh37 and GRCh38 coordinates on the same row, so no liftover is needed at read time. That is the opposite of the rule here — genome_build lives in the manifest and in no parquet column (@build-in-manifest-only), the build is injected at load and never authored on a row (@build-injected), and a module is single-build by design with reference_examples/grch37_build as the worked case. Their design buys convenience and pays for it with two coordinates that can disagree and nothing able to notice. Not a gap. Worth one paragraph in whichever survey reads them, and no more.

And one that is not an axis but a corpus: RM28 has a third entry. snappy's genosets.json carries SNPedia's boolean-combinator DSL — and(rs4988235(C;C), rs182549(C;C)), with genoset-of-genoset nesting for haplogroup trees. That is a third independent grammar for the same thing after CIViC's molecular profiles (RM174) and ClawBio's guideline logic, and unlike those two it is a general-purpose one written for consumer genetics rather than falling out of one domain. Counting its operators and its nesting depth is worth doing whether or not the rest of snappy is read.

Checked and skipped — recorded so round 3 does not re-search them

Same class as genomi (fetch public sources at query time, no curated content): lagodinm/dna-health-report, MrOrtiz/dna-annotator, Gunshipz/genomine, itsrudaynah/Modrik, techninja/asili, drhudsonandrade/OmniGenis (58k LOC and its own scripts say "Nothing here is authored"), mentatpsi/OSGenome (146★, crawls SNPedia live and freezes nothing). Pipeline or converter, no interpretation: GeiserX/Personal-Genome-Pipeline (10★), captainzonks/GeneGnome. Curated but too thin or too stale to teach anything: Michael-Sebero/Genetic-Trait-Detector (245 rows, one 432-line file, last pushed 2025-03), dev-kvt/GenomeUpload (27 uncited rows inline in JS, with an LLM writing the actual report). Curated only as a panel — a list of rsIDs with no fields on them: brandonsaldan/codex (37★, 225 JSON files of bare rsID membership, all interpretation scraped live from SNPedia, dead since 2023). Documentation, not a tool: matbanik/agentic-genomics (a setup guide wiring up three MCP servers that live elsewhere). Not competitors, one line each: the single-source MCP servers (berntpopp/clinvar-link, cyanheads/gnomad-genetics-mcp-server, and our own dna-seq/ensembl-mcp) wrap one API and carry no annotation; the general AI-science workbenches (ScienceClaw, aipoch/open-science, wisp-science, all 600–4,000★) are not genome tools; WGLab/InterVar (213★) is ACMG classification, which ClawBio's half already covered, and its last commit is 2021. Commercial, no source: Promethease (now MyHeritage), Genomelink, SelfDecode, Codegen, Nebula. RM188's method does not reach them — a report can be bought and read, but no pipeline can be run and nothing can be pinned, so a survey of one would be a marketing comparison, which § What it is not forbids. Promethease is reachable through SNPedia instead, which is why row 2 is the corpus and not the product.

The agent-skill class is closed, and it cost one read

BioTender-max/awesome-bio-agent-skills (178★) is a directory of AI agent skills for biomedical work. Enumerating the entries that interpret a human personal genome — as opposed to answering "what does the literature say about gene X" — gives 21, and they cluster under four collections: clawbio/ (six: pharmgx-reporter, nutrigx-advisor, gwas-prs, genome-compare, claw-ancestry-pca, clinical-variant-reporter), openclaw/, omicsclaw/ and bioskills/. The rest are ACMG classification, PRS, and PGx under different names.

Round 1 already surveyed the representative of this class. ClawBio's six skills are in this directory, and CLAWBIO_SURVEY.md read them end to end. Nothing in the remaining fifteen is a different shape — they are the same fetch-and-render procedure over the same public sources, and the directory itself is the evidence for that rather than an argument. So the agent-skill class is closed for round 2, at the cost of one directory read rather than fifteen surveys, and round 3 should not reopen it without a reason that is not "there are more of them now".

Two names in it are domains nothing here touches and neither survey raised: chip-clonal-hematopoiesis-agent (somatic clonal haematopoiesis) and prs-net-deep-learning-agent. Neither is an annotation table, and both are recorded only so the next reader knows they were seen.

Security note, since the search surfaced it. Barrelsravennagrass984/Personal-Genome-Pipeline is a near-verbatim clone of GeiserX/Personal-Genome-Pipeline whose README is replaced with download bait for a .zip committed inside the module tree. Do not fetch or run it. Recorded here because the next person to run this search will find it in the same result set.

RM237 — ClinGen dosage is gene-keyed, so a recurrent-CNV region has nowhere to be written

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when an author wants to write a recurrent-CNV region and cannot, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked · before parking: open — filed 2026-09-13 from RM188's ClawBio half · Owner enricher (the lane) + format (the model) · Motivating case probes/CLAWBIO_SURVEY.md § dosage keyed on a region, not a gene

GeneMetricsRow carries haploinsufficiency and triplosensitivity, keyed on a gene symbol, and enricher/clingen.py fills them. ClinGen publishes four dosage lists — gene curation, region curation, recurrent CNVs and ISCA regions — and enricher/acmg.py's own probe note from 2026-08-03 records seeing all four on the FTP tree while reaching for one. The three unread ones are the half a CNV classifier actually needs: ClawBio's skills/cnv-acmg-classifier/data/curated_dosage_map.csv carries an element_type column precisely because two of its four demo rows are regions (22q11.2, chr22:18,900,000–21,500,000, hi=3 ts=3), and its ClinGen/ACMG 2019 Section 1 scores off them.

A region is genuinely the subject, which is why this is not a widening of the existing table. A 22q11.2 deletion is not a claim about any one gene in the interval; filing it under a gene symbol would be a false attribution of the kind @gene-map-is-another-sources-attribution forbids. So the shape is a sibling sidecar — region_metrics.csv, one row per (chrom, start, end, name) with haploinsufficiency, triplosensitivity and ClinGen's curation id — rather than an element_type discriminator on gene_metrics.csv. "One CSV = one concern" and @sidecar-name-and-place both push apart here: a region row carries coordinates and a gene row carries a symbol, and a table holding both would have half its key null on every row.

Two things to settle before writing the pass, and the second is the one a review will catch:

  • Terms. ClinGen's dosage lists are believed unrestricted and that is recalled, not probed (@no-named-licence). Read the file and its SPDX id first.
  • Which (source, layer) row this claims. @write-the-sourcerow's second-surface clause says a second surface of an already-declared source may not claim the lane's existing row. Whether region curation is a second surface of clingen or the same lane is undecided, and it decides the SourceRow key.

Related RM188, @write-the-sourcerow, @gene-map-is-another-sources-attribution.

RM240 — the ClawBio nutrition panel compiles, and the corpus should carry it

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when the corpus is asked for a nutrigenomics example, or RM188 unparks, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · before parking: open — filed 2026-09-13 from RM188's ClawBio half; the translation is run, the landing is not · Owner format (the corpus) · Motivating case probes/CLAWBIO_SURVEY.md § The translation, run

The survey's headline claim — a competitor's inlined panel is a module written in the wrong language — was run rather than asserted. skills/nutrigx/data/snp_panel.json at 1fcecb7e (28 SNPs) translated field-for-field into a spec, validate green, compile producing 84 weight rows over 28 variants, 24 genes and 12 categories, and compile → reverse → compile reproducing content_signature exactly. Land it as reference_examples/nutrigenomics_panel/, which is @probe-becomes-example and is the only form in which the claim stays true as the code moves.

What it broke on the way, and the README has to say all of it:

  • state is required and direction is not. 84 rows refused for omitting the superseded column while carrying the modern one. Already in the 1.0-cleanup tracker with the right reason (P8 forbids demoting a required field inside a major, so the demotion and the deprecation land together at 1.0). What the run adds is that this is the first error a first-time author meets, before anything about their data.
  • Their 28 ref_allele cells carry no coordinate, and the schema refuses a bare ref — correctly, since an unanchored reference allele cannot be checked against a reference genome. The honest translation drops ref and keeps effect_allele.

What the landing still needs, none of it optional because the corpus walkers assert it: a resolution.csv so validate --strict and compile --strict pass (test_reference_example_compiles_under_strict), a sources.csv whose terms are read rather than recalled — their data_license is blank and the repo is MIT, which is a licence for the code (@a-hosts-terms-are-not-its-contents-terms) — a verification.json closure (test_every_reference_example_is_closed_and_its_closure_still_describes_it), a README, and a section in REFERENCE_EXAMPLES.md. Two enricher passes are worth running first because they are what makes it the worked answer rather than a copy: check-identifiers (their panel names BCMO1, a retired HGNC symbol) and literature over the 28 PMIDs.

One observation for the README, not a repair. Their panel weights rs429358 and rs7412 as two independent additive rows in two different nutrient domains. Those two variants are the APOE ε pair, which reference_examples/apoe_epsilon/ models as a haplotype because the isoform is the joint state. That is a claim about their curation, and it belongs in the README as a sentence.

Related RM188, RM92 (see RM241 — the weighting block this example is forced to write).

RM164 — heteroplasmy.csv is a shipped table kind with no source behind it

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when its own gate below is met: a source of heteroplasmy calls, PacBio's pipeline output included if it carries one, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked · before parking: open — PARKED to 0.8, decided 2026-09-01, moved into this file 2026-09-11 · Owner enricher · Motivating case the 2026-09-01 source-adoption round

Decided 2026-09-01 with the maintainer: parks, on the measured negative below. The candidate field anyone has named is MITOMAP and the population callsets it re-hosts, and none of them publishes the axis the kind binds; that is a fact about what exists, not about how hard anyone looked, which is what makes the deferral honest rather than indefinite. It stays open and visible rather than closed, because a kind with a one-module corpus is exactly what @probe-uniform-corpus says to keep in view — and if a source that bands heteroplasmy by tissue appears, this entry is where it is checked against. Reopen it with a source, never with an argument. The spin-off it noticed is now RM171.

Probed and drafted in PROPOSAL_0_7_PT2 on 2026-09-01 — proposed PARKS to 0.8; the maintainer pass took it as proposed. Answered by reading the source, after the maintainer supplied the 2026-08-24 pg_dump (61 MB, 95 tables). MITOMAP is reachable — plain curl gets the dump at mitomap.org/downloads/, HTTP 206 with ranges; the Cloudflare challenge is on the web surface only, and two earlier readings of this entry (a "refusal", then "unreachable by the machinery") were both a 403 from a path that was not the data path. The axis answer is a measured no. The schema has exactly one tissue column, on mitomap.unpublished — per-patient submissions beside sample_id and ethnicity, i.e. sample data this format does not carry. mitomap.mmutation is 602 rows whose homo/hetero are presence flags (+ 286/270, - 216/238, nr 90/89, plus ./na/NULL), with no threshold, no band and no tissue — the only levels in the table are two rows where a percentage was typed into a flag column. The only heteroplasmy numbers anywhere are re-hosted blood-cohort data (mitomap.gnomad 18,164 rows, mitomap.helix 14,104), where max_observed_heteroplasmy is a cohort observation, not a clinical threshold. So HeteroplasmyRow's binding columns have no source-side value in MITOMAP. Terms are unread, not unestablishable — the dump carries no licence text in 6.7 M lines and the page a browser reaches was not opened. Parks, not closed. Separately noticed and not part of this entry: mmutation is a plausible mtDNA variants.csv source, blocked on status being 29 free-text strings rather than a vocabulary — its own item when taken.

The measurement, taken over _TABLE_KINDS and the enricher's providers. Every table kind is in DRAFTABLE by construction, so structurally all nine are draftable. A provider exists for four: haplotypes/allele_function/diplotypes (pgx_draft ← CPIC), pharm_variants (clinpgx_draft ← ClinPGx), and variants (clinvar_draft, civic_draft, pubmind_draft). heteroplasmy.csv, repeat_alleles.csv, copynumbers.csv, pgs.csv and activity_phenotype.csv have none, and no enrichment pass reads them for a cross-check either. enrich() does resolve heteroplasmy rows — it is the third table that can ask, and the one that keys with alts — but resolution is not a source.

The corpus behind the kind is one module: reference_examples/mt_heteroplasmy, two MT-TL1 variants of one gene, hand-authored from the literature. That is the @probe-uniform-corpus shape exactly — the schema generalized from a single case, and nothing since has taken a second one.

MITOMAP is the canonical mtDNA variant table and the obvious candidate. Two things must be established before that is a plan, and neither is:

  1. The terms. MITOMAP is not CC0, and this entry states nothing further about its licence. Whether it is expressible as a SourceTerms at all, and whether it lands as a draft source or only as a check, is decided by reading its published terms. RM153 is the standing reminder that a source's terms page can answer HTTP 200 with something that is not terms, and that the honest record of an unestablished axis is None (@no-named-licence).
  2. Whether it carries the axis at all. HeteroplasmyRow binds a level band per (gene, reference_sequence, tissue, variant_key). A per-variant pathogenicity table with no tissue and no threshold fills the identity columns and none of the binding ones — it would draft rows that say nothing the kind exists to say. Probe the real file and name the table probed (@probe-the-real-file, @probe-names-the-table); a negative here is as useful as a positive and closes the item cleanly rather than leaving it open forever.

Related RM165 (the same shape on the other uncovered binning kind), RM171 (the spin-off), @probe-uniform-corpus.

RM122 — the measure lookup is specified and nothing anywhere implements it

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when its own gate below is met: a caller of the measure lookup (ask just-dna-lite first, as the gate audit says), when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked · before parking: parked on demand, moved here 2026-08-21 — additive and minor-legal whenever it is wanted; what it waits on is a caller, not a decision · Owner format · Motivating case S58 (just-module-creator, in CONSUMER_SUGGESTIONS_HISTORY.md)

The specification half shipped; this is the part that was not asked for and might still be right. S58 reported that the four binning kinds annotate nothing downstream and asked for either a normative paragraph or an admission that the family is specified ahead of its consumers. Both are now in SCHEMAS.md § The measure lookup a conforming consumer implements, and that closes the item they filed. What is filed here is the next question, which they did not ask: whether the rule should also exist as a public function so that the first two consumers to implement it cannot disagree.

The argument for. It is the shape S51/RM115 settled one layer down — a rule kept as prose is a rule every reader re-derives, and the two derivations differ on exactly the cases that matter. Here those cases are enumerable and sharp: the continuous shared endpoint, the float32 comparison, unresolved versus no-match, and pleiotropy returning several rows rather than one. A consumer will get at least one of the four wrong, and the failure is silent — a wrong bin renders as a confident phenotype. alleles.split_genotype is the precedent: the reader half of RM81 shipped as one public leaf every tier calls, while the retype waits for a major. It costs the format tier nothing — pure arithmetic over loaded rows, pydantic-only, no dependency moves.

The argument against, and it is why this is open rather than done. There is no consumer to check the shape against, which is the same reason measure_step is not a column: a signature fixed against a hypothesis fixes the wrong thing, and this one has real shape questions. Does it take rows or a table? Does it return one row, or one per trait_efo_id (the honest answer, and the inconvenient one)? Does it answer None for no-match, or a three-state result distinguishing no match from unresolved selected — which is what the house algebra would demand and what a None return would collapse. Getting that wrong ships a leaf whose first real user has to work around it, and P3 keeps it working forever.

What would settle it: one consumer implementing the lookup against the paragraph. Their questions are the signature. Until then the paragraph is the contract and this stays filed — the same wait-for-the-demand rule that governs measure_step, applied to a function instead of a column.

Decided 2026-08-21: wait for demand, and the wait is the answer rather than a way of postponing one. The four shape questions are the reason — does the lookup take rows or a table, does it return one row or one per trait_efo_id (the honest answer, and the inconvenient one), does it answer None for no-match or a three-state result distinguishing no match from unresolved selected. There is nobody to check any of those against, and P3 keeps a wrong leaf working forever. This is the same wait-for-the-demand rule that keeps measure_step out of the schema, applied to a function.

It moved out of the active roadmap because "undecided release" was the wrong bucket for it. Nothing about this is undecided; it is parked, which is what this file is for, and it sat under a heading that made it read like an unmade call. The settling event is specific: one consumer implementing the lookup against the paragraph in SCHEMAS.md. Their questions are the signature — file it back in the active roadmap when they arrive, not before.

RM23 — Computational predictor scores as a table

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when its own gate below is met: the acquisition measurement and the per-transcript decision, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked · before parking: deferred — considered for 0.6 on 2026-08-13 and held, on the two blockers unchanged · Owner format (schema + compiler) + enricher · Motivating case pathogenicity triage; splice-impact panels

predictions.csv — the groundwork every predictor source needs, built once: one row per (variant, predictor, score_kind) with score, dataset, source, and an optional transcript. Long-form, not wide, is the load-bearing choice — SpliceAI is four deltas plus positions, CADD is one number, AlphaMissense is one plus a class, so wide columns would make every new predictor a schema bump while long form makes it data. A predictor score is the same class of object as an allele frequency or a LOEUF (a per-variant number from a named dataset, no measurement), so the 0.5 sidecar precedent covers it.

Grain settled by RM236 (0.8, shipped 2026-11-04 on 0.8). consequences.csv is keyed (variant_key, transcript) with Open Targets' transcript set (one canonical transcript per nearby gene); predictions.csv's optional transcript takes the same grain when this is built, so the two never give two answers to "which transcript" (RM236's recommendation, one question asked once).

Why it is still deferred, after the 0.6 review. Neither blocker is code, and neither has moved:

  • Grain. SpliceAI scores are per-transcript, and there is no settled way to name the four splice deltas without inventing a predictor-specific column set — which is the exact thing long form exists to avoid. Whether a per-transcript score is one row each or a picked representative is undecided.
  • Acquisition. Precomputed splice scores need the masked vs raw file sizes measured and the Broad lookup API's terms read. This is the same measure-first question that correctly parked the frequency snapshot, and skipping it is how a shape gets fixed against a guess.

Unlike the two derived tables 0.6 does build (RM24, RM25), this one is a full-cost authored table under the 2026-08-13 charter amendment, so the bar is higher rather than lower.

Licensing is solved, not blocking — do not re-raise it as the reason. SpliceAI/Pangolin, dbNSFP, AlphaMissense, REVEL, CADD and PrimateAI are all non-commercial or academic-only, and licensing.csv plus the compile gate already confine that to the modules that use them, while phyloP/phastCons/GERP (UCSC, free, queryable per-range rather than a bulk download) keep a module sellable.

What would unpark it: the acquisition measurement done, and a decision on per-transcript grain. That is a research task, not a schema task, and it commits nothing.

RM16 — Authored PRS weights (a scoring file, not a manifest)

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when the shape question below is answered: how the weights combine, what reference distribution a score is stated against, and whether a percentile travels with it. That needs nothing built; the PGS Catalog's scoring files and the transcribed table below are enough to try it. The old gate, "a real consumer", was circular (the gate audit), and this restates it. Or when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium-large (on demand) · Status open — a minor, parked · before parking: deferred — considered for 0.6 on 2026-08-13 and held; parked on shape rather than on demand since 2026-09-13, and its unpark condition needs restating · Owner format (schema + compiler) · Motivating case authored-weight PRS modules

0.4 shipped pgs.csv as a manifest of PGS Catalog IDs with the ancestry-validity fields — not authored per-variant weights. just-prs resolves a PGSxxxxxx id to a harmonized scoring file and scores each id itself, so inlined weights would be dead data; and a PRS is a Z/percentile-in-reference, a shape the format does not bin.

What is deferred is a distinct, digest-bearing effect_allele + effect_weight scoring table, for the case where a module must ship weights the PGS Catalog does not host (a score published only in a paper's supplementary table).

Why it is still deferred. It is not derivable — nobody can fetch weights that exist only in an appendix — so it is a full-cost authored table. And the one thing that would validate its shape, a real consumer combining authored weights into a score, does not exist. A score's shape (how weights combine, what the reference distribution is, whether a percentile travels with it) is exactly what a first real case would dictate, so fixing it now spends a one-way door on a guess.

What would unpark it: a real consumer. See PROPOSAL_0_5.md D1.

This entry and RM28 are one axis at two operators (maintainer, 2026-09-13), and they should be decided together rather than merged. Both ask the same question — how does a conclusion get derived from more than one variant row, and where does the combining rule live? — and the format already answers it once:

Combiner Where the rule lives Status
Enumerative — write out each combination in the table: a diplotypes.csv row is the combination shipped, reference_examples/apoe_epsilon
Boolean — a predicate over genotypes nowhere: the table would hold terms, something else the operator RM28, parked on a corpus
Weighted sum — terms plus coefficients nowhere: the table would hold weights, the consumer the arithmetic this entry, parked on shape

Three facts follow from the table and none of them from either entry alone. The enumerative answer is why both are parked and neither is urgent — it covers the small-arity cases, which is most of them, so any argument for a new combiner has to beat it rather than merely want one. Both non-shipped rows split the same way against the data-agnostic goal: the module carries terms, the consumer does the combining, so neither is a licence to put an evaluator in the compiler. And one corpus supplies evidence for both — SNPedia has genosets and per-genotype magnitudes, which is why one survey touched both entries on one day.

They stay separate items. The outputs differ in kind — RM28's is a categorical conclusion, this one's is a number that means nothing without a reference distribution — and RM28 was already halved once on 2026-08-13 for being too broad, so merging would make the smaller hostage to the larger. What they share is a precondition, not a design: whoever opens either should read both, and the where does the combining rule live decision should be taken once rather than twice.

The hypothesised case now has an instance in the wild (2026-09-13), and it does not unpark this. From RM188's round-2 skim: dianguan0105/Custom-Personal-Genome-Interpretation (15fe28a) ships metaprs_weights.tsv — 540 variants transcribed by hand out of a European Heart Journal 2022 supplementary Table S3, with rsid, effect_allele, other_allele, weight, maf and traits. That is this entry's motivating sentence — "a score published only in a paper's supplementary table" — with a real file behind it instead of a hypothesis, and the column set they arrived at independently is close to the one deferred here.

The first reading of that was wrong, and the maintainer corrected it the same day. It was written here as the wrong half of the evidence — the unpark condition is a real consumer, a competitor transcribing a table is a producer, so nothing moves. That dichotomy does not hold, and the entry's own gate is what has to give.

An author with a corpus in hand is demand, and for a format it is the demand that matters. People reach for familiar shapes: somebody who already has a scoring table, or a set of genosets, adopts a format when the format can hold what they already have, and declines it when it cannot. That makes "can I bring this in" a consumer question in the sense that decides whether a format gets used at all — a narrower reading, where a consumer is a downstream tool that reads authored weights and scores them, gates this entry on the one party who cannot appear until after the table exists. Written as it stands, the unpark condition cannot be met by anything except a tool built against a table this entry is refusing to build. That is a gate on its own output, and it should be restated when this entry is next opened for a decision — not silently, and not here, since restating a parking condition is a change to the item rather than a note on it.

What survives the correction, and it is the half that was always load-bearing. The demand question is answered — somebody did the typing, and SNPedia's magnitude-and-genoset corpus is the same argument at a hundred thousand rows. The shape question is not: how the weights combine, what reference distribution the score is stated against, and whether a percentile travels with it are still unanswered, and a transcribed table shows the columns without showing the semantics. The one-way door is the shape, never the demand, so what this instance does is move the entry from parked-on-demand to parked-on-shape — a smaller and much more answerable place to be parked.

Register cross-reference (2026-09-27). RM302 gives conclusion optional lay and concise wordings, two columns by default. If this item lands a separate meta-conclusion table, the maintainer's steer is that the registers may live there instead, so the two are decided together.

RM68 — a drafting provider on a non-GRCh38 module: refuse, or strip to the rsID

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when its own gate below is met: an author with a non-GRCh38 module says which outcome they wanted, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium (high before the warning shipped) · Status open — a minor, parked · before parking: the warning shipped in 0.6 (enrich.source_build_mismatch); what the providers should do instead is deferred here · Owner enricher (the three drafting providers) · Found by dogfooding on 2026-08-13, reference_examples/cyp2c9_warfarin_grch37/

What was observed

draft, draft-panel and draft-clinpgx all take a spec_dir, and until the 0.6 dogfooding round none of them read genome_build. enrich.spec_genome_build — written one release earlier for the bug where the guard existed and the value never arrived — had exactly one caller. Every source these providers read serves GRCh38: CPIC's allele_definitions, the ClinVar snapshot, the ClinPGx annotations. So drafting CYP2C9 into a genome_build: GRCh37 module writes 10,94942290 for rs1799853, whose GRCh37 position is 96702047 — a different base 1.76 Mb away — and nothing anywhere said a word.

Nothing downstream catches it. A coordinate is legal on either assembly; it is simply a different place. What the online diagnosis reaches is now measured rather than asserted: of the module's two GRCh37 rows declared as GRCh38, grch37.diagnose_wrong_build caught one and the other minted ga4gh:VA.pgprki8YgzfOSV9Dpe1ccPX4uNdlyAvB and recorded resolved, because GRCh38 happens to carry the authored ref at that position too. That is the documented ~3-in-4 sensitivity, so "the compiler catches wrong-build coordinates" is not a reading anyone should take. Two of the three providers write coordinates and can do this; draft-clinpgx writes none.

It hid because test_pgx_draft.py's fixture declares GRCh38, and it is the only drafting test that mentions a build at all.

What shipped in the same session. enrich.source_build_mismatch: each drafting command asks before it writes and warns naming both builds, what a drafted coordinate will actually mean, and the two remedies. The provider still writes the row, which is the enricher's standing shape for a disagreement — report, never repair.

The question

Report-and-still-write is the right default. What is undecided is whether a provider meeting a non-GRCh38 module should go further: refuse, or strip to the rsID — writing the identity the source does state without stating an assembly, which derive_variant_key prefers anyway.

Refuse — wrong three ways. It makes a GRCh37 module undraftable, so the author hand-authors instead, and hand-authoring against a printed contract is where this format's most expensive documented failure came from (the 0-vs-1-based start description shifted four whole modules by one base and passed every offline gate, --strict included). It refuses a provider that cannot do the harm: draft-clinpgx writes no coordinate, so a refusal keyed on the module's build stops a command whose entire output is build-free. And it is the wrong granularity by the repo's own rule — scaffolding refuses per file, drafting refuses per row — while a build mismatch is a property of the module, so a build-keyed refusal is per run, which is the granularity that self-defuses into "this module cannot be drafted at all".

Strip to the rsID — wrong, and worst on the rows that need it most. It is tempting because the row's identity does not move (derive_variant_key returns the rsID first) and an rsID names a variant without naming an assembly. But CPIC's sequence_location publishes defining variants with a position and no rsID — 18 in CYP2C9, 14 in TPMT, 4 in NUDT15 — and HaplotypeRow requires an rsID or chrom+start. Stripping the coordinate there is not "write less", it is "drop the row", and it drops exactly the rows the 0.5.1 gene.chr repair recovered after a year of being skipped. It also produces a row indistinguishable from one the source had no coordinate for, so the author cannot tell a stripped row from a bare one. Any partial strip is barred outright: a drafting provider fills identity whole or not at all, and a lone alts on a position-only row makes derive_variant_key mint a ga4gh:VA.… instead of chrom:start:ref, silently changing which variant the row is.

Lift the coordinate over — refused, and RM48 already argues it. RM48 is deliberately one-way and reporting-only: a GRCh37 coordinate recovers an rs-number, and the rs-number is reported, never filled, because filling it would make resolution verify a value against the service that produced it. A liftover inside a drafting provider is that move with an extra assembly on it, and chrom/start are both in hints.REDUNDANCY_BEARING for the same reason.

A --build flag on the drafting commands — wrong twice. A flag saying "write GRCh37" asks the provider to convert, which is the liftover above. A flag saying "yes, I know" is a warning suppressor, and the tier's standing rule is that --offline is the switch and a pass adds no second CLI flag.

What would unblock it. Either a real author with a non-GRCh38 module saying which of the two outcomes they wanted, or RM15, which dissolves the premise: once identity is build-agnostic a provider can write the coordinate under the build it came from, and there is nothing left to refuse or strip. A behaviour fixed before RM15 lands is one RM15 would have to undo, which is the strongest single argument for leaving this at a warning.

RM266 — the phase-confusable diplotype-pair set is computed at compile and dropped: it reaches the artifact only as a capped warning string

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when S116's reporter or another consumer asks for the phase-confusable pairs as data rather than as a warning, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · Owner compiler (_cross_validate_diplotypes) · Motivating case S116, note 2

What was confirmed. Reproduced by compiling reference_examples/hfe_compound_het: the diplotype_phase_ambiguous warning fires and names the exact set S116 says the artifact cannot carry — HFE: 1 group(s) … e.g. C282Y/H63D, C282Y-H63D/wt. So the compiler already computes the phase-confusable pair set. _cross_validate_diplotypes (compiler.py:3457) groups a gene's diplotype pairs by their phase-preserving definition signature and separates indistinguishable at all (identical definitions, diplotype_definitions_identical) from phase would decide (diplotype_phase_ambiguous) — which is precisely the distinction a caller reconstructs at query time. It is then flattened into a warning string and capped at three examples (_examples), so a reader of the artifact who is not running a caller cannot recover the set, and beyond three groups cannot see it at all.

Why this is not RM28. S116 frames note 2 as wanting an RM28 axis — an author stating "this pair is distinguishable from that one only by phase." But the author does not need to state it and should not: it is derived from haplotypes + diplotypes, and the compiler derives it today. This is @dont-discard-computed, not a new authored surface — the fix is to stop discarding the structure, not to add an axis. The general shape is the idea-book's § D3/§ D4 (warnings are list[str] with no code, count or payload); RM266 is the specific case where the payload — the confusable groups, by name — is the thing worth carrying, over and above D3's per-code count.

The shape, filed not built. A structured surface for the computed groups — a manifest.json field keyed by gene, or a small optional parquet — so a downstream reader sees the confusable set without re-deriving it. Additive (a new optional derived surface, P9 ≈ free-to-half), so minor; the open question is manifest field versus parquet, and whether it subsumes or sits beside a general warnings_summary (§ D3). Corpus, not code, decides urgency: this is just-dna-lite's first caller and it already re-derives the set successfully, so nothing is blocked.

RM281 — the PGx lane reads one of ClinPGx's archives, and clinicalVariants.zip bears on a shipped table kind

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when an author drafting pharm_variants.csv needs a row clinicalVariants.zip carries and summaryAnnotations.zip does not, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked (a source adoption) · Owner enricher (clinpgx_build, clinpgx_draft) · Motivating case RM166's "It wants its own number", unfiled until the 2026-09-27 postmortem sweep (P9) · related RM166, RM173, RM175

RM166 noticed that ClinPGx publishes at least twelve archives while the lane builds from one (summaryAnnotations.zip today, per RM173). clinicalVariants.zip is the one that bears on a shipped kind: about 5,190 rows of pharm_variants.csv territory. Its type is a six-member base vocabulary that comma-combines, so an adoption has to normalize the combination, not the token (@one-normalizer-two-spellings).

Confirmed on 2026-09-27. No module in enricher/ names clinicalVariants. The lane's builder reads summaryAnnotations.zip and refuses the retired clinicalAnnotations.zip (RM173), and the drug_labels lane is separate.

Why a minor. Adopting it is a new drafting route, or new columns in a published snapshot. It is also the same terms as the rest of the lane (CC BY-SA, no sale), so it widens nothing on the licence axis. That half is RM282.

RM282 — the PGx lane has one licence class, CC BY-SA with no sale, and nothing looks for a source outside it

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when someone runs the survey its own gate below describes, which needs nothing built, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · parked on a survey chosen for terms first, and the class is settled by what it finds · Owner enricher (source adoption) · Motivating case RM166's "it wants its own entry" (also PROPOSAL_0_7_PT2 § the decision), unfiled until the 2026-09-27 postmortem sweep (P9) · related RM166, RM281, @pgx-research-only

RM166 closed its FDA half measured: "the lane does not gain a member outside its licence class by this route, and the direct route cannot supply one either". It then said that if licence diversification for the PGx lane still matters ("it plausibly does, since it is a single point of failure on the axis the format gates on"), it wants its own entry, with candidates chosen for their terms first. No entry was written.

What is true today. ClinPGx, CPIC and PharmVar are all CC BY-SA with no sale (@pgx-research-only), and the compile gate (@gate-is-data-driven) refuses a commercial declared use on every module drafted from them. A commercial author has no PGx source at all.

The gate, and why it is satisfiable now. A survey of PGx sources ranked by terms before content. Anyone can do it today and it needs nothing built. If it finds a candidate, adopting it is a normal source adoption (a minor). If it finds none, this closes as measured, with the survey as the record.

RM284 — a skip is a fact about the run, and verification.json has no run-level place to keep one

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when a reader reports needing to know, from verification.json alone, that a run could not reach a source, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked · Owner format (the verification document) + enricher (the writer) · Motivating case RM72's "a run-level fact needs a run-level place, which is a separate question and was not opened" (decided in 0.6 PT2), found unfiled by the 2026-09-27 postmortem sweep (P11) · related RM72

RM72 decided the merge rule for verification.json: newest wins between two records of the same disposition, and a skip does not displace an answer while the earlier record still binds the authored bytes (existing_still_binds). The design half it answered and set aside is this one. A reader may legitimately want to know that today's enrichment could not reach a source, and under the shipped rule that fact is dropped: the older ran record is what stays.

Confirmed on 2026-09-27. No open item names a run-level skip record, and verification.json remains a per-check document.

Why a minor. A run-level record is a new field or block in a published document, whatever shape it takes. The shape question is where it lives: in verification.json beside the per-check records, or in the run's own output (EnrichmentResult), which RM272 is already widening on the 0.8 branch.

RM287 — ClinGen's dosage pass is a live download with no snapshot lane, so an offline deployment never runs it

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when an offline deployment reports a dosage pass that never ran, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity low · Status open — a minor, parked (a new cache lane) · Owner enricher (clingen, CACHE_LANES) · Motivating case RM39's "Not done, and it was asked for explicitly: a ClinGen snapshot", never filed, found by the 2026-09-27 postmortem sweep (P14) · related RM39, RM38, @gated-source-caches

RM39 gave the ClinGen dosage pass an offline flag and a skipped_offline state, so an offline run says it did not ask. It left the snapshot the reporter asked for explicitly, calling it "RM38's family and a much bigger question". Nothing filed it.

Confirmed on 2026-09-27. CACHE_LANES has fifteen lanes and none is ClinGen. clingen.py still fetches the gene-curation list live and skips under --offline. Every other pass in the family reads a snapshot, so an offline deployment gets every answer but this one.

What building it owes. A lane, so docs/CACHE_SURFACE.md's checklist applies line by line (three stages, each absent one with its reason as a field, repro_out default, parents). ClinGen's terms have to be recorded as a SourceRow before the lane can be published. A new lane is new public surface, so this is a minor.

RM296 — a PRS has no condition → conclusion table, so pgs.csv only restates what the PGS Catalog already publishes

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when an author writes a PRS module that needs a score → conclusion band, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked (a new optional binning kind) · Owner format (schema) + compiler · Motivating case the maintainer, 2026-09-27: "modules are condition → conclusion by design; just listing PGS is a bit tautological, they have own EFOs" · related RM16, RM47, RM289, RM163

What is missing. Every other kind a module carries maps a measured condition to an authored conclusion: a genotype to a conclusion, an activity score to a metabolizer phenotype, a repeat count or a copy number or a heteroplasmy fraction to a band. pgs.csv maps nothing. Its row is an accession, the trait's EFO ids, ancestry and a research tier, all of which the PGS Catalog already publishes per score. A module cannot say "above the 95th percentile of PGS000018 in a matched reference: elevated CAD risk relative to that reference", which is the sentence a PRS module exists to carry.

Why the refusal it replaces was a choice, not a constraint. pgs.py's docstring says a PRS yields a Z or percentile within a matched reference distribution, "which the format does not bin". Binning one is the activity_phenotype.csv shape: the consumer computes the number and the module bins it, so the data-agnostic line (no measured value in a module) holds exactly as it does there. Nothing in the charter forbids it. A new optional table is minor-legal (P3, P8) and costs the full authored price (P9).

The shape, as a starting point. A MeasureBinRow subclass keyed on pgs_id (joining pgs.csv, which becomes the declaration the bins hang on rather than the whole module). It inherits bounds, measure_tiling, the unresolved no-call sentinel and bin-level pmid (RM47).

What has to be designed, and it is the part that stops this being a tautology too.

  • The measure's unit and frame are part of the claim. A percentile and a Z are different axes, and either is meaningful only against a named reference distribution. A bin has to say which (measure_unit, and a reference identifier), or a band means nothing.
  • Ancestry mismatch withholds. A sample outside the bin's reference or training_ancestry gets unknown, never the band (the house tri-state). The same applies below match_rate_floor, which is what unresolved is for.
  • research_tier constrains the conclusion. A research_only score supports only relative statements ("higher than N% of the reference"). An absolute-risk band needs calibrated. Whether that is a check or a documented rule is open.
  • Continuous tiling is the natural default for a percentile or a Z (@measure-tiling).

RM16 (authored weights) is a different question and this does not depend on it. When this ships, pgs.py's docstring, TABLES.md § pgs.csv and ROADMAP_0_8 § RM16's "a shape the format does not bin" all change with it.

RM297 — the ClinPGx currency check compares a row against an annotation it never cited

✅ Shipped 2026-09-28 on main, uncut — a patch. The entry is in ROADMAP_HISTORY.md. This heading stays only because S122's reply links here, and replies are the triage seat's to retarget; delete it in the same commit that moves that link. It is a ### on purpose: a ## RMn here is an open item to the §4 counter (test_triage_tools.py), and this one is not open.

RM298 — nothing records per row that a pharm_variants.csv row came from ClinPGx

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when the ClinPGx currency check withholds on a real mixed module whose operator needed the answer (RM297's warning is the interim), when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked (a new optional authored column) · Owner format (PharmVariantRow) + enricher (clinpgx_draft, clinpgx) · Motivating case S122 via RM297, decided with the maintainer 2026-09-27: "the mitigation is warning, not blocking. Solution comes in minor" · related RM297, @source-vs-authority, @write-the-sourcerow

What is missing. Provenance for this table is recorded per (source, layer) in sources.csv. A module that mixes ClinPGx-drafted rows with rows a curator took from an article, CPIC or DPWG cannot say which is which. So the ClinPGx currency check cannot tell a withdrawn ClinPGx annotation from a row that never cited one, and RM297 has to withhold on both.

The shape, as a starting point. An optional per-row column naming the source whose accession annotation_id is. draft_pharm_variants fills it, and a curator may set it or leave it empty. The check then refuses under strict on an absent id only for rows that name ClinPGx, and withholds on the rest. Empty stays the unknown arm, never "not ClinPGx".

To design before building.

  • The name and the vocabulary. The value set is the sources.csv source column, so an undeclared name is a finding. Audit the name against the reserved namespace (P5).
  • Back-fill for published modules. An existing drafted row has none. Whether draft-clinpgx fills it on a row it recognises as already_present is a merge-not-clobber question (@draft-appends: drafting appends and never mutates).
  • Whether other tables want it too. The same gap exists wherever one table takes rows from several sources. Build it here first, on the table with the incident.

Price and legality. A new optional column is minor-legal (P3, P8) and costs the full authored price (P9), because a curator has to learn it. That cost is why it waits for a minor and a case, rather than shipping with RM297.

RM306 — no exon structure anywhere in the tier, and AlphaGenome's gene attribution wants one

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the 2026-11-03 sweep that moved everything outside 0.8's scope to 0.9. Unparks when a real module's AlphaGenome rows need an exon-level gene attribution, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked (a second table in the MANE lane) · Owner enricher (mane_build, gene_spans) · Motivating case GKS_SURVEY.md §5, 2026-09-28; the maintainer's ask for gene/exon coordinate mapping behind the AlphaGenome lane

What exists. gene_spans.py reads the MANE summary for one span per gene, and its docstring fixes the rule: a span is a query hint, never an attribution. Nothing in the tier knows an exon. grep -ri exon over the enricher finds only civic_identities.py.

The measured options, lightest first.

Option Size Covers Misses Licence
MANE Ensembl GTF, a second file in the existing MANE lane 8.6 MB gz; 19,437 transcripts, 204,815 exons, CDS and UTR rows, exon_number, versioned ENSG/ENST, RefSeq xref exon boundaries and numbers for MANE Select + Plus Clinical non-coding genes; non-MANE isoforms NCBI policy, already MANE_TERMS
GENCODE v46 basic GTF 29 MB gz the exact annotation AlphaGenome used, all biotypes MANE selection, unless joined by ENST "open access", no named licence (@no-named-licence)
cool-seq-tool + UTA + SeqRepo 309 MB install, a UTA Postgres service, 13.5 GB SeqRepo any transcript version, c.↔g. with exon offsets nothing AlphaGenome attribution needs UTA is CC BY-SA 4.0

The third is refused on weight, not on merit. It needs a database service and a 13.5 GB store, which conflicts with the tier's --offline provisioning and with "builder in polars, runtime in duckdb".

Release skew to state, not hide. MANE 1.5 corresponds to Ensembl 116, while AlphaGenome was annotated with GENCODE v46 (Ensembl 112). An exon boundary that moved between the two releases attributes differently in each. The lane records both release labels, and a lookup answers in MANE's frame by default.

Open questions. Is MANE alone enough (1–1.5 days), or does non-coding coverage justify a GENCODE v46 lane too (about one more day)? And the answering surface: gene_exons(symbol) next to gene_span, or a position → (gene, transcript, exon) lookup. The attribution rule in gene_spans.py constrains both: an exon lookup answers where, and AlphaGenome's own per-record gene remains the attribution.

RM238 — a per-tissue eQTL has no row, and expression_effects.csv is the wrong table to widen

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the pm's placement of 2026-11-04 (below). Unparks when its own gate below is met: someone who needs a per-tissue eQTL answer, when a scope the maintainer highlights touches it, or when the roadmap is empty.

0.9, decided 2026-11-04 by the pm seat, on the lead's build plan (PROPOSAL_0_8_PT2 §4 question 2). The coloc input v2's effector confidence needs comes from Open Targets' colocalisation dataset, in RM333's lane under the same CC0 marking. So this per-tissue GTEx table goes back to waiting on a consumer: one lane and one licence instead of three sources, the maintainer's reduction rule.

Severity low · Status open — a minor, parked · before parking: open — filed 2026-09-13 from RM188's ClawBio half; PARKED on a consumer, deliberately · Owner enricher · Motivating case probes/CLAWBIO_SURVEY.md § nine databases, and the two axes that come back with no home

For longevitymap v2 (2026-11-03): the cis-QTL colocalisation behind effector confidence E1/E2. See the group.

clawbio.py run gwas --demo on rs3798220 returns five cis-eQTLs by named tissue from GTEx and the EBI eQTL Catalogue — LPA in Liver at β −0.82, in Adipose at −0.45, SLC22A3 in Liver at +0.31. Nothing here holds that. expression_effects.csv exists and is the wrong place: it is AlphaGenome-shaped, one row per (variant, gene) aggregating 371 tissue tracks into tracks_agreeing/tracks_total, and expression.py argues at length that one row per track "is lossless and unreadable". A GTEx row is not a track — it is a measured cis-eQTL in one named tissue with its own β and p, ~50 tissues rather than 371, and the tissue is the fact rather than something to consensus over. Different grain, different source, different terms. So: eqtl_effects.csv, one row per (variant, gene, tissue, dataset), tissue as an ontology term where the source gives one.

This is filed open and parked, and the parking reason is a rule rather than a shortage of time. USE_CASES § 7.2 closes with "reopen this with a consumer, never with an argument", and a competitor rendering the table is an argument. Reopen it with somebody who needs the answer. When that happens, the first step is the acquisition measurement that correctly parked the frequency snapshot — what the GTEx and eQTL Catalogue bulk artifacts weigh, and what their terms say (@probe-the-real-file).

Related RM188, RM194/RM200 (the AlphaGenome table this must not be folded into), USE_CASES § 7.2.

RM378 — errors carry no code, so a consumer grading a spec tells one refusal from another by its wording

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the pm's placement of 2026-11-05 (below). Unparks when a consumer needs to tell apart refusals that RM377's two phrases do not cover, when a scope the maintainer highlights touches it, or when the roadmap is empty.

0.9, placed 2026-11-05 by the pm. 0.8 is the work already built, the fixes already chosen and longevitymap v2, and "the rest goes to 0.9" (the maintainer, 2026-11-03). RM378 is new, its code vocabulary is permanent within the major once shipped, and it has two open design questions, so it is not added to a release at its cut. RM377 already gives the reporter the two phrases it needs today. The maintainer can still take it into 0.8.

Severity low · Status open — a minor, parked (placed by the pm on 2026-11-05; triage proposed 0.8) · Owner compiler (the result types, every errors append site) + format (the vocabulary) · Motivating case S129 in CONSUMER_SUGGESTIONS_HISTORY.md (just-dna-registry's liaison seat, 2026-11-04) · related RM377, RM131, RM218, RM359

The gap. RM131 gave the warnings channel a code per finding and warnings_summary; errors stayed list[str] on ValidationResult, CompilationResult and ClosureResult. A consumer that grades a spec must read an error's words to know which refusal it is. RM377 published the two phrases the reporter needs today; this item is the general answer, the reporter's candidate (a).

What it would be. A code from a closed vocabulary on each error and an errors_summary beside warnings_summary, by RM131's shape: derived in the same mode="before" validator, summing to len(errors) or empty, never partial. A mode-laddered finding raised to an error under strict already has a warning code, which is dropped on the error path today, so part of this is keeping a code that exists.

Open questions, for the interview.

  1. Pydantic's own errors. A rejected row's message comes from pydantic and has no closed vocabulary. RM131 refuses both a partial summary and a catch-all key. So either every such error gets one code that names the finding (a row the model rejected), or the summary is defined over the coded errors only and says so. The first keeps RM131's rule; the second breaks it.
  2. The vocabulary's cost. Codes are permanent within the major (P3, P5, P6), so the first set shipped is the one consumers key on. RM131 measured that the audit of emission sites is the cost; the container is near free (P9: a result-model field, nothing authored).
  3. The reporter's candidate (b), a field naming the strict-only module-level refusals, is narrower and cheaper. It is not taken here: it builds one consumer's split (module-level against per-row) into the model, a classification axis under P5, which the reporter named as the reason it might be wrong. With codes, each consumer draws its own line.

Residuals none yet.

RM28 — Meta-conclusions (the predicate half)

Parked, the pm (2026-11-05), out of ROADMAP_0_9.md, under the maintainer's rule that a minor takes the zones they name and nothing by default ("'Auto-next minor' was a policy when we didn't have a dedicated decision maker for this"). It reached 0.9 by the move of the items parked on a corpus. Unparks when its own gate below is met: the corpus count, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Moved here from ROADMAP_0_8.md at the 0.8.0 cut (2026-11-05), unchanged but for the file it sits in: its combination half (the meta-table) shipped in 0.8 (ROADMAP_HISTORY § RM28), and the decisions it was built to are in history/ROADMAP_0_8.md. What waits is the predicate half's remainder (open-world negation, OR, trans phase), parked on the corpus below.

Severity medium (after the corpus) · Status open — a minor, parked · before parking: parked on a corpus — and halved on 2026-08-13: the injected-cofactor half closed, the predicate half stays here · Owner format (schema + compiler) · Motivating case combination annotations; disclosure policy

For longevitymap v2 (2026-11-03): KL-VS buffering in APOE4 carriers joins two loci on two chromosomes, so no haplotype holds it; added as a corpus entry there. See the group.

Read this entry knowing how much of the original item has already dissolved. Probing in 0.5 removed most of it, and the 0.6 review removed the rest of the cofactor axis. What is left is genuinely small and genuinely unsolved.

The corpus this was parked waiting for has its first real entry (2026-09-02, RM174). Not an argument — an adopted source whose grammar was counted. CIViC publishes molecular profiles as boolean expressions over variants: of 1,964 profiles, 209 are multi-variant — AND 141, OR 72, NOT 1 — and they nest (BRAF Amplification AND ( BRAF V600E OR BRAF V600K )). Both halves of this entry's surviving case appear there, with instances:

  • Economy, and it is this entry's own phrase. Evidence item 8721 is a claim about VHL S183L AND VHL D126N whose description reads "heterozygous compound mutation" — two variants in trans, observed, cited. HaplotypeRow cannot hold it: a haplotype is same-strand co-location, so drafting it there asserts cis where the source says trans, silently, because no column carries phase for anything to contradict.
  • Open-world negation. MET Amplification AND NOT KRAS Mutation quantifies over a set no module can close, on a class term rather than an enumerable allele.

The 72 disjunctions are the half that is already expressible — rows are a disjunction — which is where this entry drew the line and it holds. This is evidence for the corpus, not a reason to unpark: one source, one observed trans instance, one negation. RM174 carries the measurement; the decision stays here.

Second corpus entry, 2026-09-13 — and it is a routine clinical guideline rather than an oncology molecular profile. From RM188's ClawBio half (probes/CLAWBIO_SURVEY.md § pharmgx-reporter). Their pharmacogenomic report renders exactly one AVOID across 59 drugs, and it is warfarin, keyed "genes": ["CYP2C9", "VKORC1"]. Our own reference_examples/cyp2c9_warfarin_grch37/ carries the CYP2C9 diplotype phenotypes in diplotypes.csv and the VKORC1 per-genotype claims in pharm_variants.csv, side by side, with nothing keying the pair — which is this entry's surviving "pairing across subjects" clause with a shipped module standing on it.

Three things make it sharper than the CIViC entry rather than a repeat of it:

  • It is the drafter's recorded gap, not an inference. pgx_draft's own comment says it: "the guideline exists, it is a dosing algorithm over several genes rather than a per-phenotype recommendation, so nothing lands here and the author was told CPIC has nothing." What shipped for it was a better warning, deliberately, not a table.
  • The competitor has no declarative form either. ClawBio's answer is "special": "warfarin" — a hardcoded branch calling get_warfarin_rec(profiles), for one drug out of 59. That is evidence the shape is genuinely hard, not evidence that everyone else solved it.
  • It is CPIC, so the population is countable. The measurement this entry still wants, and the one cheap enough to do before any design: how many CPIC guidelines are keyed on a gene pair rather than a single gene phenotype, counted off the CPIC snapshot already provisioned. If the answer is three, this parks again with a sharper reason; if it is thirty, the economy argument changes.

Third corpus entry, 2026-09-13 — and this one is a general-purpose grammar rather than a domain's by-product. From the same item's round-2 roster: SNPedia publishes genosets, boolean expressions over genotypes, and zhaofengli/snappy (2d5255f) has them extracted from a MediaWiki dump into data/genosets.json beside 106,603 SNP entries. The grammar is and(rs4988235(C;C), rs182549(C;C)), it nests, and a genoset may reference another genoset — the haplogroup trees are built that way.

What makes it the sharpest of the three is that CIViC's profiles and ClawBio's warfarin branch each fell out of one domain solving one problem, where this grammar was written for consumer genetics in general and has been in use for over a decade. It is also the first entry where the condition side and the conclusion side are both published, at scale, by one source.

This entry and RM16 are one axis at two operators — boolean here, weighted-sum there, with the enumerative combiner already shipped in diplotypes.csv and covering the small-arity cases both would serve. RM16 carries the table setting that out; read it before opening this one, and take the where does the combining rule live decision once rather than twice. They stay separate items, because a categorical conclusion and a number needing a reference distribution are different outputs.

One argument from it does reach beyond the count, and it is RM16's rather than this entry's (maintainer, 2026-09-13): genosets are a familiar shape, in use for over a decade, and an author's ability to bring one across is an adoption signal in its own right. It bears on this entry only as a reason the count below is worth running rather than as a reason to unpark — this item is parked on a corpus, and a corpus is measured, not argued.

It changes no decision yet, and the reason is the same one that has held twice. The measurement this entry wants from it is cheap and has not been done: how many genosets are there, what is the operator distribution, and how deep does the nesting go — counted off genosets.json, which is a committed file needing no network. A corpus of forty flat conjunctions argues differently from four thousand nested ones. The licence is not the obstacle a first reading here made it: CC BY-NC-SA 3.0 US is established and adoptable as an unsellable share-alike module (USE_CASES § 2e). What still has to be beaten is reference_examples/apoe_epsilon, which answers the two-variant case with no predicate at all.

Fourth corpus entry, 2026-09-13 — and it is the first where enumeration succeeds and the schema around it complains anyway. Raised by the maintainer from their own genotype: Kulminski's subdivision of APOE ε4 by linked TOMM40/APOC1 markers, where a carrier is not simply ε4 but 1-0-0, 1-1-0, 1-0-1, 1-1-1 — a binary vector over the ε4 call plus the linked sites, each pattern carrying a different reported risk.

This one is expressible today, and that is the finding rather than a reason to close. Three markers is eight patterns and at most thirty-six pairs, which is haplotypes.csv (one row per pattern × defining marker) plus diplotypes.csv (pair → phenotype) — the shape reference_examples/apoe_epsilon already ships for ε2/ε3/ε4 across rs429358 and rs7412, with one more marker. The enumerative combiner is not beaten here, so this entry argues for no predicate at all.

What it does show is the shape being outgrown at a different seam — the subject, not the combiner (the maintainer's reading, and it is the right one). Expressing it fires two shipped warnings, and neither is a bug:

  • composite_gene_cell. The haplotype spans TOMM40, APOE and APOC1; DiplotypeRow.gene is required and single-valued (HaplotypeRow.gene is optional, so only the diplotype half forces the choice). Labelling the locus TOMM40-APOE-APOC1 publishes a gene name no registry index will match, beside its parts — which is exactly what the warning says. So the author picks between one findable symbol and an honest locus label, and there is no third option. apoe_epsilon never met this because ε alleles sit in one gene.
  • diplotype_phase_ambiguous, conditionally. HaplotypeRow is same-strand co-location, so this translation is only honest if Kulminski's patterns are phased haplotypes. If they are unphased multi-locus genotype vectors, writing them as haplotypes asserts cis the source never claimed — RM174's trap exactly — and the pairs that collapse to one unphased genotype with different conclusions are what this warning already catches, telling a consumer with unphased calls to withhold.

So the honest verdict is: yes, with caveats, and the caveats are the entry. The format holds the claim; what it cannot do is hold it cleanly. A use case that compiles only by choosing between findability and honesty, and only under a phase assumption, has grown past the gene-keyed subject the PGx tables were cut for — and that pressure lands on DiplotypeRow's key rather than on any predicate language. Read against the other three entries, this is the one that argues RM28 may be the wrong item for its own class of case: CIViC's contribution was that enumeration cannot reach trans and open-world negation, ClawBio's was that nobody has a declarative form for a multi-gene dosing algorithm, SNPedia's was a general grammar at scale — and this one enumerates fine and still does not fit. A locus-keyed subject would settle it and a predicate would not.

The count this entry asked for, 2026-09-20 — and it is between the two numbers the question named. Measured off the provisioned CPIC snapshot (cpic_snapshot_b0ffd4c6f010, built 2026-09-02) while answering S102, whose reporter had counted the same table from the other side: of 103 drugs with a recommendation row, 18 are keyed only on a gene pair and none is keyed both ways — every drug is single-gene or pair-keyed, never a mix — and gene_count never exceeds 2. By rows it is 2,656 of 3,411 pair-keyed, which is what made the drafter's "the snapshot has no row for it" false for a drug with 35 of them. The 18 drugs fall into six pairs:

pair drugs
CACNA1S + RYR1 desflurane, enflurane, halothane, isoflurane, methoxyflurane, sevoflurane, succinylcholine
CYP2C19 + CYP2D6 amitriptyline, clomipramine, doxepin, imipramine, trimipramine
NUDT15 + TPMT azathioprine, mercaptopurine, thioguanine
ABCG2 + SLCO1B1 rosuvastatin
CYP2B6 + CYP2C19 sertraline
CYP2C9 + SLCO1B1 fluvastatin

Warfarin is not among them: it has no recommendation row at any arity, because its guideline is a dosing algorithm rather than a recommendation table, so the ClawBio entry's case is the one CPIC drug this count cannot reach. Read against the question — three parks again, thirty changes the economy argument — eighteen drugs across six pairs is the middle: more than a warfarin-shaped exception, less than a pattern the enumerative combiner is failing at scale, and every one of them is arity 2. The drafter now names the partner gene instead of claiming absence (RM249), which is a message and not a table; whether six pairs earn a pair-keyed subject stays the maintainer's decision, and this entry stays parked on it. The reporter's reading — ClawBio keys each of these on a single gene and never mentions the partner, a lossy answer, not a model to copy — is the right one and is recorded beside the survey's plan item 1, which this measurement answers.

Unmeasured, and flagged as such. This is a reported case: the paper has not been read here, so the marker count, the exact patterns, and above all whether Kulminski phases them are unverified. That last one decides which of the two warnings above applies, and it is the first thing to establish if anyone builds it. Separately, just-module-creator's assets/longevity_2026 was checked and does not attempt this — it takes the opposite reading, collapsing the cluster to rs429358 as the shared causal signal on colocalization evidence, with TOMM40/APOC1 named only in prose and no haplotypes.csv at all.

Still parked, on the same rule: the corpus grows, the decision does not move until the count is in.

2026-09-20 — the count is in, and it does not support the parking reason. Measured directly off the provisioned CPIC snapshot (data/interim/cpic/, dataset: cpic_snapshot_3d2123598711, built 2026-08-07, recommendations.parquet 3,411 rows), prompted by a peer session drafting ClawBio's PGx corpus that hit it from the other side:

  • 2,656 of 3,411 recommendation rows (78%) are gene_count = 2. Multi-gene is not the exception in CPIC's recommendation table; single-gene is.
  • 36 of 121 (gene, drug) pairs have no single-gene row at all. For those the drafter writes nothing, because CpicSnapshot.recommendations filters gene_count = 1 (deliberately — a row about CYP2C19 and CYP2D6 is not a statement about CYP2C19 alone) and the live client applies the same len(phenotypes) != 1 filter.
  • They are recognisable clinical pairs, not edge cases: TPMT+NUDT15 over the three thiopurines (azathioprine, mercaptopurine, thioguanine — 35 rows each); CYP2D6+CYP2C19 over five tricyclics (amitriptyline, clomipramine, doxepin, imipramine, trimipramine — 206 rows each); SLCO1B1 with CYP2C9 and ABCG2 over fluvastatin and rosuvastatin; CYP2B6+CYP2C19 over sertraline.
  • RYR1+CACNA1S is the one that breaks the fallback argument. Seven drugs — the six volatile anaesthetics plus succinylcholine — and neither gene appears in a single-gene recommendation anywhere in the table. Elsewhere a module can at least carry the single-gene rule and lose the refinement; for malignant-hyperthermia susceptibility there is no single-gene rule to fall back on, so a module carries nothing for those seven drugs.

Independently reproduced on a second snapshot, and the two units reconcile exactly. The peer session measured cpic_snapshot_b0ffd4c6f010 and counted drugs where the above counts (gene, drug) pairs: 103 drugs carry a recommendation, 85 single-gene, 18 pair-keyed across six pairs — NUDT15+TPMT (3 drugs), CYP2C19+CYP2D6 (5), CACNA1S+RYR1 (7), ABCG2+SLCO1B1 (rosuvastatin), CYP2B6+CYP2C19 (sertraline), CYP2C9+SLCO1B1 (fluvastatin). Those 18 drugs times two genes each is 36 (gene, drug) pairs, which is the number measured here from a different snapshot built six weeks apart. Two denominators, stated apart deliberately (@two-surfaces-two-denominators): 18 of 103 drugs and 36 of 121 pairs are the same fact at two grains, and the drug grain is the more legible one to quote. Warfarin is in neither: CPIC's warfarin guideline is a dosing algorithm and carries no recommendation row at all, so the shape this entry was reopened on is not even the marginal case — it is off the end of the table.

What this changes and what it does not. It does not by itself unpark the entry: the economy and open-world-negation objections in What remains are untouched, and a table keyed on two subjects is still a design nobody has drawn. What it removes is the premise the parking rested on — that pairing across subjects is a rare shape worth deferring until a corpus appears. Three sources now carry it (CIViC 209 profiles, CPIC 36 pairs / 78% of rows, ClawBio's one hardcoded branch), and in CPIC's case the pairs are the routine half of a guideline set this project already ingests and drafts from.

The smaller finding beside it is filed separately as S102 (peer report, 2026-09-20): the drafter does not just skip these, it misreports them — the snapshot path's knows_drug returns None unconditionally, so the known is None arm always fires and tells the author the snapshot has no row, when it has 35 or 206 of them and filtered every one. That is @answered-is-not-absent (answered-and-rejected is a fourth state; the row stays and the reason moves), and the arm's remedy — consult the live API — reaches the identical filter. The fourth arm, the one whose text names warfarin, is unreachable whenever a snapshot is present.

A blood-group instance, S119 (2026-09-27): Lewis is keyed on two genes' phenotypes. The Lewis phenotype (Le(a−b+), Le(a+b−), Le(a−b−)) is a function of the FUT3 phenotype and the FUT2 secretor phenotype, not of either gene's diplotype. just-dna-lite shipped FUT2 secretor status as its own module and left Lewis out for exactly this reason. It is the same subject-pairing shape as CPIC's six gene pairs, outside pharmacogenomics, and like them it enumerates (a small table over two per-gene phenotype vocabularies), so it argues for a two-subject key and not for a predicate. Unbuilt, so it is corpus evidence and not a measured case.

A longevity instance, 2026-11-03: KL-VS buffering in APOE ε4 carriers. longevitymap v2's plan (§4.2, protection type BUFFERING) holds that the KL-VS heterozygote matters in ε4 carriers. KL is on chromosome 13 and APOE on 19, so no haplotype holds the pair, and both halves are already enumerable on their own (KL-VS as two-SNP haplotype rows, APOE as apoe_epsilon). Like CPIC's gene pairs and Lewis, it argues for a two-subject key over two enumerated per-gene states (3 × 6 rows), not for a predicate. The plan's interim is flags: conditional plus prose, and that tag does nothing (RM321). Unbuilt; corpus evidence.

What dissolved, so it is not re-proposed

  • No operator is missing. Rows are a disjunction and columns are a conjunction, so the existing tables already span any finite boolean function over an enumerable set of genotypes: OR is two rows, XOR and bounded NOT are enumeration, and haplotypes.csv is same-strand AND.
  • APOE — whose ε4 condition (rs429358==C AND rs7412==C) is the Constitution's own example of a predicate — was built with 0.4 bricks and no predicate at all (reference_examples/apoe_epsilon/). HaplotypeRow is a junction table, so same-strand co-location is what it already expresses.
  • The cis/trans motivation closed as a check, not a table. reference_examples/hfe_compound_het/ showed that a diplotype is already a statement about two homologs, so cis and trans are two rows — the relational notion the grammar was going to add is what a diplotype pair is. What building it surfaced instead was that the two rows are indistinguishable without phase, which shipped as _cross_validate_phase_ambiguity (a warning, never a block). A requires_phase column was rejected: it would make an author restate what the data determines and go stale the moment a haplotype is edited. Confirmed from the caller side, S116 (2026-09-26): a shipped just-dna-lite caller ran against this example and the enumeration held — because the module defines the cis allele (C282Y-H63D) as its own haplotype with its own diplotype row, the caller could name the ambiguity (ambiguous, both candidates) rather than guessing or reporting no_match. So this is a case where enumeration succeeds and no predicate is wanted, which is where the entry drew its line. The one thing the caller had to re-derive — which diplotype pairs are the phase-confusable set — is not a predicate the author states but a value the compiler already computes and discards, filed as RM266, not here.

What remains

  • Pairing across subjects — no table keys on more than one.
  • Economy and intent — "any two pathogenic variants in trans" over 300 of them is ~45,000 pairs: expressible, unwritable, unreadable.
  • Open-world negation — "no pathogenic variant in this gene" quantifies over a set the module does not close, and absence is only assertable where the region was callable (requires_callable). No operator fixes this, and a negation feature ignoring it would manufacture reassurance, the worst failure mode this format has.

It also blocks the "shy module" signal.

Why it stays parked

It waits on a corpus to generalize from — roughly 70% built; nutrigenomics and supplements do not exist yet — because fixing a shape against four table kinds and then meeting the fifth is how a one-way door gets spent badly (P3/P5).

The design thread, the starter shape and what is deliberately left open are in PROPOSAL_0_5.md § G3: a new optional table, a predicate that never blocks, a grammar kept to the smallest thing covering the motivating case, and a three-valued algebra (true/false/unknown, Kleene operators) — with unknown withheld, never reported and never negated. Kleene matters concretely: a conclusion gated on "ε4 present AND QUAL ≥ 60" is decidably false at ref/ref whatever the quality was, so a blanket withhold-on-any-unknown would be strictly worse than the tables it replaces.

The cofactor half — CLOSED in the 0.6 review, do not re-open as a general mechanism

The original item proposed injected cofactors: values the consumer supplies at query time that a module must never hold, in three classes. Two of the three were resolved in 0.5 by making them plain columns — clinical context became DiplotypeRow.clinical_context, call quality became quality_from / min_quality (RM29) — and neither needed a general mechanism.

Decided 2026-08-13: the general "injected cofactor" mechanism is dropped as never-earned. Each remaining class gets the same treatment — a plain column, on demand, one at a time. Two classes wait:

  • Ancestry — a panel-scale inference, not derivable from a curated module's own gnomAD frequencies, since real models do not rely on single SNPs.
  • Family structure — this is RM10, folded in here on 2026-08-13. A declarative trio / de-novo / Mendelian-consistency expectation is only meaningful once the consumer supplies family structure at query time, which makes it a cofactor class rather than its own item. Designing it separately would fix a shape for one class before the axis exists. It was on-demand-only and shapeless for its whole life; it stays that way, here.

Neither is built until a real module needs it.

RM402 — a snapshot answer rejected under RM382 is asked again of live Ensembl, the same authority, which answers the same locus

Placed, the pm (2026-11-07): after RM401 ships, and only if a re-measure still shows it. The 600 s here is 38 cold one-at-a-time GETs, which is RM401's cost; batched, the same 38 ids are one POST of tens of seconds. What RM402 would then save is that remainder, bought with a new release-equality rule over two spellings of one label, which is its own risk. So it waits: once RM401 ships, a re-run of the registry's module (or any online re-derivation with reverse-complement genotypes) says whether the repeat asks still cost enough to build it. If not, it closes as superseded by RM401.

Severity medium · Status open — a patch, waits until RM401 ships and a re-measure (fewer requests; the written table is unchanged wherever the two releases agree) · Owner enricher (enrich/orchestration.py, the live link's missing ~:948) · Motivating case S133 in CONSUMER_SUGGESTIONS_HISTORY.md · related RM382, RM373, RM292, RM401

The gap. RM382 sets a rejected answer aside so the next link is asked (@answered-is-not-absent). The live link's missing skips an rsID live Ensembl already rejected (rejected_from(..., _LIVE_ENSEMBL_LINKS)) but not one the Ensembl snapshot rejected. Live REST and the snapshot are one authority (RM292 counts REST self-consistent with it), so at one release REST returns the locus the snapshot did and the rejection repeats. On the registry's module that was 38 asks, none placed, about 600 s: every one an rsID whose authored genotype is the reverse complement. enrich already tells the author so ("check the strand your alleles are written on"), and flipping those genotypes is the author's remedy; this item is the cost of the repeat ask, not that.

What it would be. An rsID the snapshot answered and the hosting rule rejected is not asked of live Ensembl when live REST's data release equals the snapshot's (RM373 records both: /info/data and the lane's release label). The release is asked before the loop rather than after it. ClinVar and gnomAD are still asked, being other authorities, so the rule the gotcha states holds for every other link. When the releases differ, or REST's cannot be read, the ask stays. Open question: the two labels are spelled differently (ensembl_116 against …vep116.0, RM373's won't-fix), so equality needs a stated mapping, tested both ways (@one-normalizer-two-spellings).

RM405 — an offline enrichment's floor is the Ensembl snapshot's lookups, and the snapshot has no rsID index

Parked, the pm (2026-11-07), as filed from S133. An offline enrichment takes about 10 s a module, which nobody reported as a problem: S133's complaint was the online path. The cheap half (measure the four queries, fewer scans) buys seconds; the real fix is an rsID-sorted layout, which is the snapshot builder's. Unparks when a consumer reports offline enrichment time as a problem (a sweep or a deadline it misses), the snapshot builder offers an rsID-sorted layout under RM303's contract, a scope the maintainer highlights touches it, or the roadmap is empty.

Severity low · Status open — a minor or a patch, parked (class by design) · Owner enricher (resolver._lookup_rsid_candidates, _lookup_positions_by_rsid, _lookup_rsid_sets_by_position, _lookup_rsids_by_position) · Motivating case S133 in CONSUMER_SUGGESTIONS_HISTORY.md · related RM372, RM303

The gap. The registry measured 9 of an offline enrichment's 10 s in four duckdb queries over the snapshot's parquet view. Two of the four are rsID lookups and two are by position (the registry's note calls all four by rsID); each scans the 1.17-billion-record snapshot with no index. Which query costs what was not measured here.

What it would be. Measure the four first. Then either fewer scans in one run (one rsID pass, one positional pass), or a lookup the snapshot's sort order serves. Residuals, to be stated at closure: an rsID-sorted sidecar or an index is the snapshot builder's, prepare-annotations, and its layout is RM303's contract.

RM406 — dbSNP witnesses only the indels REST was asked about, though the run already holds its placement for every rsID

Parked, the pm (2026-11-07), split from RM398 on the maintainer's ask: "I was suggesting potentially adding one. As a separate item, for future." RM398, the next patch, puts dbSNP where REST was and no further. Unparks when RM398's build has counted dbSNP-versus-snapshot disagreements over the registry's replay and the count says what a disagreement should do, a consumer reports a misplacement outside REST's subjects, a scope the maintainer highlights touches it, or the roadmap is empty.

Severity low · Status open — a minor or a patch, parked (class by design: report-only adds counts, withholding changes which loci are written) · Owner enricher (placement.py, identifiers.classify_rsid) · Motivating case the maintainer's ask, 2026-11-07; S117 and S132 for the class of error · related RM398, RM267, RM292, RM303

The cost is not requests. The maintainer's worry: "On any placed = very high cost, imagine 1M rest calls. Should be some flag. Some bayesian sampling maybe?" dbSNP's placement comes from the merge check's esummary db=snp. Every online run already sends that for every rsID it carries, on by default (verify_rsids=True), 200 ids a request (EutilsSettings.batch_size). So widening asks nothing new and asks REST nothing: a million-rsID module costs 5,000 esummary requests today, whether their SPDI is read or not. What widening costs is:

  • which loci a run withholds: every substitution or ClinVar-held indel dbSNP disagrees with would be withheld, and strict refuses a subject with every locus withheld;
  • comparison work: an indel's left-alignment reads the reference; a substitution's comparison reads none.

What it would be, cheapest first.

  1. Report only. Compare every placed rsID with dbSNP and count the disagreements in placement_agreement's detail, withholding nothing beyond RM398's subjects. It needs no flag, since it costs no request and changes no written row; it says how often the snapshot and dbSNP part.
  2. Withhold, if what (1) counts shows dbSNP is right where they part. A flag only if two deployments want different answers (@retry-attempt-floor).

Sampling answers a different question: whether a snapshot as a whole is sound, asked once per snapshot rather than per module. A sample of the snapshot's indels checked against dbSNP at build or provision time would estimate its misplacement rate with an interval. What a sample can and cannot say, worked on 2026-11-07:

  • A systematic error is settled by a few samples. The maintainer: "bayessian samples answers alignmend with absurd confidence on 5000 samples." If the whole snapshot, or one event kind in it, is off by one, nearly every sample disagrees. Even allowing 1% chance disagreement per sample, 10 disagreeing samples in a row is about 10⁻²⁰ (near 9 sigma), and 5,000 is beyond any table.
  • A class error is measured, if the sample is drawn from that class. The known misplacements are insertions (S117; S132's 38% of insertions on two chromosomes), so the sample is drawn from indels, not from all records, where substitutions would swamp it. 5,000 indels put a 38% rate at ±1.4%.
  • A rare per-record error is not found. With no disagreement in 5,000, the 95% bound on the rate is about 3/5,000, 0.06%; at the snapshot's 1.17 billion records that still allows hundreds of thousands. S117's single ABO insertion is of that kind. That is why the per-module comparison stays: it costs no request and checks the very records a module uses.

Residuals, to be stated at closure: that snapshot-wide audit is the snapshot builder's (prepare-annotations, under RM303's contract), or a provision-time check filed on its own.

RM419 — a client that builds its own pacing gate paces per instance, not per service, beyond NCBI too

Filed by the lead (2026-11-08) while building RM418, which fixed this for NCBI E-utilities alone, as the pm sized it. The same error class has a second instance: Ensembl REST limits requests per address (15/s), and two of our clients ask it with a gate each, the resolver (ensembl.EnsemblResolver, EnsemblSettings.interval 0.1 s) and the LD pass (ld_pairs, LD_MIN_INTERVAL 0.1 s). A run that resolves and fetches LD pairs at once, or a host running two enrichments in one process, paces each to the full rate.

Severity low · Status open — a patch, parked (the pm, 2026-11-08) · Owner enricher (net._service_gate and every client that builds a PacingGate by default) · Motivating case RM418's build · related RM418, RM396, RM404

The gap, by client. Each builds its own gate when not handed one: ensembl, grch37 (grch37.rest.ensembl.org, a second host with its own limit), ld_pairs, gnomad, civic_api, clingen_allele, atlas_client, gwas, clinvar_gkm, currency, and the private-gate clients RM396 names (pgs, litvar, pharmvar). Whether each is a per-address budget, and so the same defect, is a question per service; litvar is on ncbi.nlm.nih.gov but is not E-utilities, so whether it shares NCBI's address budget is unmeasured.

What it would be. RM418's mechanism, taken by each client whose service limits per address or per key: net._service_gate(<service>, interval) instead of PacingGate(interval), keyed by service so two unrelated budgets at one interval stay apart, and the two Ensembl clients on one key. The enricher conftest already resets the registry per test. No public name moves, so it fits a patch; whether it is one, and which clients, is the pm's.

Sized, the pm (2026-11-08): parked, not built in 0.8.x now.

  • No run of ours meets it. The enricher runs its passes one after another; the only thread pools are in two bulk builders (alphagenome_avi_build, opentargets_build), and neither reaches Ensembl REST. So the resolver and the LD pass never ask Ensembl at the same moment, and no other pair on the list does either. The defect needs a host that runs two enrichments as threads in one process, which is the question put to the registry for RM418 and still unanswered.
  • No incident. Nobody has reported a 429 from Ensembl, or from any service on the list, in a run where two of our clients shared a process. RM418 had the registry's ask behind it; this has the mechanism only.
  • The fix's own risk is the keying, not the code. Each change is one line, but each key is a claim that two clients spend one budget. A wrong key slows a run (two budgets paced as one); a missing one is today's state. Half the list is unmeasured on exactly that point: litvar, PmcIdConverterClient and PmcBiocClient sit on NCBI hosts but are not E-utilities, and grch37 is its own host.
  • Unparks when either (a) a run reports a rate-limit refusal from a service on the list while two of our clients shared a process, or (b) the registry says it runs enrichments as threads in one process. Either can happen while this is parked, so the gate is honest (the gate audit has its row).
  • When it unparks, start with the Ensembl pair (resolver and ld_pairs on one key, grch37 on its own): the one pair whose shared per-address budget is documented. Each further client joins only with its service's limit measured or quoted from its documentation, and a key per service, never one per interval.

RM424 — a verification record carries no binding of its own, so a reader cannot tell which checks are current

Severity low · Status open — a minor, parked (the pm, 2026-11-09, at RM423's sizing; filed at its closure) · Owner format (VerificationRecord, merge_records) + enricher (record_verification) + compiler (the verification block's staleness read) · Motivating case S139's candidate 2 · related RM423, RM45, RM72, RM284

The gap. verification.json is bound once, at the document (VerificationDoc.module_hash). RM423 made a run over moved authored bytes keep only its own records, so the document never signs an earlier record under a binding it was not computed over; the price is the history. An earlier answer about other bytes is now dropped rather than kept and marked, so a reader cannot see that a check was put once, over which bytes, and is not current. Per-record checked_at survives on the records that stay, and nothing compares it.

What it would be. S139's candidate 2: a module_hash on each VerificationRecord, the binding that record was computed over. merge_records could then carry every record across an edit, and the compiler reads each one as stale on its own, as it reads the document today. A new optional field, so a minor; whether it joins VERIFICATION_FACT_FIELDS (and so moves published signatures) is the first design question, answered against RM129's reasoning for producer and checked_at.

Parked, the pm (2026-11-09). A new field is a minor, and a minor takes the zones the maintainer names, nothing by default. Unparks when a reader asks to see which checks are current without re-running them, or wants the history RM423 now drops. Either can happen while it is parked: S139's reporter already proposes, on their side, reading a record whose checked_at predates their newest authoring edit as not current, which is the first.

Residuals none yet.

RM429 — GwasResult carries no count of the page-size refusals RM428 counts, so a caller over MCP never sees the Catalog's ceiling move

Severity low · Status open — a minor, parked · Owner enricher (gwas.py) · Motivating case S144 (CONSUMER_SUGGESTIONS_HISTORY.md) · related RM428, RM354, RM100

The gap. RM428 counts each first request the Catalog refused for its size on the client (GwasCatalogClient.size_refusals). It adds them to requests_made and logs one warning per run when any occurred. GwasResult carries no count of its own. A caller that runs the pass in process and hands the result on, as the reporter's plugin does to an agent over MCP, never reads the server's log. On that path the only trace is requests_made rising by one per refused rsID, which reads as ordinary cost. The pass still succeeds at the Catalog's default page, so nothing else prompts a look. This is the shape @dont-discard-computed names: the pass computes the number and then leaves it out of the place where it reports every other cost (p_value_underflows, unparsed_effects, unusable). Today the Catalog accepts 500. The reporter's lead measured 0.8.4 live (rs1421085, 223 associations, no refusal), so this concerns the next drop, not a failure now.

The ask. size_refusals: int = 0 on GwasResult, set from the count the pass already keeps as size_refusals. The CLI's summary line prints it beside the other counts. It is a new field on a public type, so it is a minor under RELEASE_CYCLE's rule for main. With no minor branch open, it is filed and waits.

Two facts for the sizing. First, 0.8.4 as published already gives a caller the number. enrich_gwas takes client=, and GwasCatalogClient.size_refusals is a public field. A caller that injects its own client and reads the field after the run has the count today, with no log parsing. Second, that field is itself a new public name, and it landed on main in RM428's patch. Under the same rule it would have taken a leading underscore, as RM401's and RM403's names did in b6ea746a. It is published now, so taking it back would be a removal. RM429's field is the place where a result reader looks, and the client's field stays as it is.

Sizing (pm, 2026-11-11): parked. No refusal has happened since RM428 (the reporter's live run saw none), a caller can read the count today through the client it injects, and the field is a minor with no zone to go in. Unparks on a run that reports a counted refusal, or a consumer that cannot inject its own client. Then it is a near-free field and rides whichever minor is open.

Residuals none yet.

RM302 — one conclusion serves every place a report shows it; a table cell, a details card and an explanatory mode want different wordings

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when a consumer's report shows a conclusion in a second register (a table cell, a card, an explanation) and asks for the authored wording rather than its own, or an author writes two registers into one conclusion, when a scope the maintainer highlights touches it, or when the roadmap is empty. The maintainer had placed it in 0.9 by name (the next paragraph); desk #347 asks whether it stays parked, gets a 0.11 with the other two, or goes into 0.10. Parked is that question's default.

0.9, decided 2026-11-04 by the maintainer, after the final measuring run. v2 never uses the two registers, because its conclusion is already lay (R-VOICE-1), so they leave 0.8. The shape decided earlier the same day (two columns on four kinds, inside content_signature, and language reserved) stands for when the item is built.

Decided 2026-11-04 by the maintainer: two optional columns, conclusion_lay and conclusion_concise. They go on VariantRow, DiplotypeRow, PharmVariantRow and the bin kinds, but not StudyRow, because a study's finding is not the report's claim. Both are inside content_signature, as authored claim text. The name language is reserved now, as a separate axis from register. Tools never write a register; that was decided on 2026-09-27.

Severity medium · Status open — a minor, parked; principle decided, shape kept open (see Decided) · Owner format (the authored models) + compiler (the parquet columns) · Motivating case the maintainer, 2026-09-27: a report shows a variant's conclusion in a table, in a details card and in an explanatory mode, and one text cannot be right for all three · related RM7, RM28, the "not format scope" note on lay-language rendering

For longevitymap v2 (2026-11-03): v2 writes for three readers (citizen, geneticist, ageing researcher), which is the register question. See the group.

The ask. Today every annotation carries one conclusion, the detailed wording. The report needs up to three registers of the same claim: detailed (today's text), lay (a simpler explanation for someone without the background), and concise (scientifically accurate but very brief, for a table cell). They are one claim written three ways, never three claims.

Where conclusion lives today, all authored: VariantRow (required), every MeasureBinRow kind (required), DiplotypeRow (required), PharmVariantRow (required) and StudyRow (optional). It is also inside annotations.parquet's key, (variant_key, conclusion, negatives) (USE_CASES § the S-key rework), which constrains every option below.

What the charter already settles.

  • Legal as a minor (P3/P8): new optional columns or a new optional table. conclusion itself stays what it is and stays required; it becomes the detailed register by declaration, not by rename (a rename is a removal plus an addition, major-only).
  • Not in any key. A register is a rendering of the claim the key already identifies. Keying on it would split one effect into several rows whenever an author adds a lay text, the dedup failure the annotations key exists to prevent.
  • Not the refused "lay-language rendering". That note (§ Not format scope) is about generating patient prose from a CURIE, a presentation job. This is text a curator writes and signs, which is annotation data like conclusion is. The entry adds a line there so the two are not read as one.
  • Priced at full (P9): every option is authored text, the most expensive layer. The argument for paying is that a consumer cannot derive a correct lay or concise wording from the detailed one without an LLM in the loop, which the charter keeps out of every tier.

Decided by the maintainer, 2026-09-27.

  • Tools never write a register; authors and curators may, if they see fit. No drafter, enricher pass or compiler step fills, rewrites or derives a lay or concise wording, and none is required: a module with only conclusion is complete. This is the line the § Not format scope note on lay-language rendering draws from the other side.
  • Shape: kept open, with two columns as the default. Unless something below changes it, the build is option 1: two optional columns on the carrying kinds.
  • The exception is meta-conclusions. If RM28 lands its own table, the registers may belong there instead (a meta-conclusion is the claim a report most needs in lay words), and the two items then decide the shape together.

Design options, as they stood before the decision.

  1. Two optional columns on each carrying kind (conclusion_lay, conclusion_concise, names to be audited under P5). Simplest to read and to author row by row; costs two columns on five kinds, and every future register is another column on all five.
  2. One optional side table (conclusions.csv: the row's key, a closed register vocabulary, the text). One CSV, one concern: an author who writes no registers never sees it, a new register is a vocabulary member rather than a column, and the key is the carrying row's own. Costs a join for every consumer and a key per carrying kind.
  3. A register vocabulary with detailed implicit either way, so conclusion never has to move.

Open questions.

  • Which of the five kinds need registers. StudyRow.conclusion is a study's finding, not the report's claim, and may not.
  • Does content_signature include the registers? They are authored claim text, so the default reading is yes; the counter-argument is that a lay rewording does not change what the module asserts (@provenance-beside-a-claim-is-outside-content-identity decides it).
  • Language is a separate axis from register (P5: no overloaded fields). A lay text in German is register lay, language de; a design that folds the two into one vocabulary would have to be unwound when a second language arrives. Reserve the language axis's name now if it is expected.
  • Does the reference consumer (just-dna-lite) need all three, or concise first? That orders the build.
  • Should drafting ever fill a register? The house answer so far is that a provider fills facts, never prose; a register is prose.

RM350 — HLA allele names are checked for grammar only; whether the allele exists is IPD-IMGT/HLA's list, and no lane reads it

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when an author or a consumer reports an HLA allele name in a module that IPD-IMGT/HLA never assigned; the licence probe this entry starts with needs nothing built and may run meanwhile, when a scope the maintainer highlights touches it, or when the roadmap is empty. The maintainer had placed it in 0.9 by name (the next paragraph); desk #347 asks whether it stays parked, gets a 0.11 with the other two, or goes into 0.10. Parked is that question's default.

0.9, filed 2026-11-04 by the lead: RM334's residual, placed in 0.9 by the maintainer's RM334 decision ("grammar now, the list in 0.9").

Severity low · Status open — a minor, parked, licence unprobed · Owner enricher (a new lane) · Motivating case RM334 · related RM334

RM334 refuses a malformed HLA name on allele_function.csv; a well-formed name that IPD-IMGT/HLA never assigned passes. The list is the witness. Its terms are unprobed, so the first step is the licence probe (@probe-the-real-file), then a lane through CACHE_SURFACE.md's checklist.

Residuals none yet.

RM294 — STRchive's evidence grade has no place in the module, so a Refuted locus is doubted only in a draft note

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when an author or a consumer asks for STRchive's own grade beside the module's tier, or a locus STRchive grades Refuted reaches a published module, when a scope the maintainer highlights touches it, or when the roadmap is empty. The maintainer had placed it in 0.9 by name (the next paragraph); desk #347 asks whether it stays parked, gets a 0.11 with the other two, or goes into 0.10. Parked is that question's default.

0.9, decided 2026-11-04. The maintainer chose one tier column with the scale declared per module, so a source's own grade cannot sit beside the module's tier. This item waits for 0.9, where a separate home for a source's grade can be weighed (a machine-written side table, or the manifest's provenance block).

Severity low · Status open — a minor, parked · Owner format / enricher · Motivating case RM276's minor half, split off when the warning shipped on 2026-09-27 · related RM276, RM170, RM165

For longevitymap v2 (2026-11-03): a source's evidence grade travelling into a module is the same column family as RM324's tier. See the group.

RM276 names a drafted locus STRchive grades Refuted or Disputed, in a note the drafting run prints once. Nothing in the module keeps it, so the next reader of repeat_alleles.csv sees DMD's bands with no trace of the source's doubt. Carrying it is a minor: an authored column costs full price (P9), and STRchive's vocabulary is open upstream (combobox: true), so a closed frozenset would have to be priced against a source that adds members. The alternatives to weigh first are a derived sidecar (half price) or the manifest's provenance block, and whether RM170's CIViC answer already fixed the shape. Decide on the 0.8 branch.

RM345 — SourceRow cannot state five licence shapes: AlphaGenome's four, and a host licence that defers to owners a row cannot name

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when a module drafts rows from a source whose licence takes one of the five shapes and the compile gate or the registry's publish gate decides it wrong, or the registry asks for one of the shapes as data, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Severity medium · Status open — a minor, parked (new optional sources.csv columns or a sibling table; authored layer, full cost under P9), design first, placed in 0.9 by the pm, parked 2026-11-16 · Owner format + enricher · Motivating case RM333's Open Targets L2G read (ROADMAP_0_8.md) and probes/ALPHAGENOME_ATLAS.md §2.6 · related RM333, RM27, RM191, @a-hosts-terms-are-not-its-contents-terms, @licensing-as-data

The gap. SourceRow carries share_alike, commercial_use, redistribution (recorded, gated nowhere) and declared_use. Five real licence shapes have no field:

  1. Holder eligibility (AlphaGenome §2.1): "a commercial entity may not hold this at all" is not commercial_use=False, which only says the data may not be sold.
  2. A field-of-use bar (AlphaGenome §2.4/4): no training of models on it, which binds a non-commercial lab too.
  3. Revocability with a delete obligation (AlphaGenome §2.3): every source so far, at worst, stops being fetched. This one can compel deletion of derivatives already distributed, against P4's frozen digests.
  4. Notice of modification must travel (AlphaGenome §2.4/5): notice carries the terms notice, but a threshold slice is itself a modification the terms ask to be declared.
  5. A host licence that defers to the original owners (Open Targets, from RM333): data marked CC0, with "no restrictions other than those provided by the original data owners". @a-hosts-terms-are-not-its-contents-terms writes the host's terms as the floor and lets each record override them. An L2G row cannot name its owner, though: it reaches its study only through studyLocusId. So the override has nothing to key on, and the floor states less than the terms impose.

Not decided here. Whether each shape is a column, a vocabulary member on an existing axis, or a sibling table. Which of them a compile gate reads, given that redistribution is still ungated (RM27). And how shape 5 says "deferred to owners not named per row" without claiming a value: a three-valued reading, where unknown is withheld rather than written as permissive. AlphaGenome §2.6's (e) and (f) are existing gotchas, not gaps, and are not reopened here.

Residuals none yet.

RM286 — PharmVar's research-use-only, personal-key restriction lives in notice prose, and RM27 closed without the axis RM38 handed it

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when the same as RM345's, whose design it shares, or a PharmVar-drafted module reaches the registry's publish gate, when a scope the maintainer highlights touches it, or when the roadmap is empty.

Stays in 0.9, the pm (2026-11-05): it is the same design question as RM345 (a licence shape SourceRow cannot state), which stays for longevitymap v2's Open Targets rows. The status line below predates this.

Severity low · Status open — a minor, parked · Owner format (SourceRow) + enricher (the PharmVar terms constant) · Motivating case RM38's "belongs to the RM27 design round", a pointer to an item that closed without it, found by the 2026-09-27 postmortem sweep (P13) · related RM38, RM27, @pgx-research-only, @gated-source-caches

RM38 declined a SourceRow column for PharmVar's restriction (research use only, a non-transferable personal key). The restriction is narrower than commercial_use=False, so today it lives only in notice as prose. RM38 handed the axis to "the RM27 design round, which already owns 'the recorded axes do not cover every real restriction'". RM27 shipped in 0.6 as a record-only redistribution verdict with the registry named as enforcer. It designed no further axis, so the handoff landed nowhere.

One premise to correct before designing. RM38 sized the column as 1.0 because "a new column on an existing parquet moves every compiled module's digest". A digest moving is not by itself a reason to defer (the house coding standards say so), and a new optional column is minor-legal under P3 and P8. So this is a minor if built.

What to decide. Whether research-use-only and personal-key are one axis or two, and whether either is tri-state like share_alike/commercial_use (an unestablished permission is never a permission). Also whether the compile gate reads it, or it stays recorded-only like redistribution. A new authored-schema column costs the full P9 price, though sources.csv is mostly machine-written.

RM358 — a DOI-only literature answer is never pinned, and bioRxiv's own full-text route is unused

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when a module that reaches compile carries a DOI-only (pmid=NONE) citation a downstream check needs pinned, or an author asks for a bioRxiv full text the tools cannot reach, when a scope the maintainer highlights touches it, or when the roadmap is empty.

0.9, filed 2026-11-04 by the lead: RM349's two residuals, found by its builder.

Severity low · Status open — a minor, parked · Owner enricher (literature) + compiler (the literature.csv checks) · Motivating case the RM349 builder's report, 2026-11-04 · related RM349, RM336

One class: a pmid=NONE citation is checked, but the answer goes nowhere a downstream check reads. Two instances:

  1. RM349 checks a DOI-only row's existence and quote, but the answers are not pinned in literature.csv, which keys on PubMed ids. The compiler's quote counter, its licence check and its title-as-quote check read literature.csv, so none of them ever sees a pmid=NONE row.
  2. The quote is matched only where Europe PMC holds the preprint's text. bioRxiv's own API publishes a jatsxml link for every bioRxiv and medRxiv preprint, a wider route than Europe PMC's subset, and no client reads it.

The fix, as one rule. A DOI-only answer is pinned in literature.csv under its DOI, so every check that reads the table reads it too; bioRxiv's jatsxml becomes a full-text source for the quote check, as a new source with its own SourceRow and licence probe, not a branch inside an existing client.

Residuals none yet.

RM360 — four longevitymap v2 rows the 0.8 shape cannot hold, found by writing 0.8's reference examples

Parked, the pm (2026-11-16), out of ROADMAP_0_9.md, under the maintainer's rule for smaller minors (2026-11-15: "lets make em smaller"), desk #326 item 3, applied on its default. Unparks when the next version of longevitymap v2, or another module, needs one of the four rows written, when a scope the maintainer highlights touches it, or when the roadmap is empty.

0.9, filed 2026-11-04 by the pm, from the lead's report on the 0.8 reference examples. None is built in 0.8.

Severity low · Status open — a minor, parked, each gap weighed when it unparks · Owner format (columns) · Motivating case the 0.8 reference examples, 2026-11-04 · for longevitymap v2 · related RM323, RM328, RM334, RM335

One class: a v2 fact has a value and no cell to put it in. Four instances:

  1. An allele's odds ratio has no numeric home on meta.csv. A composite row states a direction and a conclusion. It cannot state the number its citation gives for the allele.
  2. A tag SNP standing for an HLA allele has no column. proxy_rsid points from one variant to another. It cannot point from a variant to an allele_function.csv allele such as an HLA type.
  3. A GenAge entry cannot be cited in model_organisms.csv. The row has pmid/doi and no slot for a database entry id.
  4. An r² cannot say which lead allele the proxy allele travels with. The computed LD pair carries r² and the population, but not the allele phase, so the effect allele cannot be carried across the proxy.

How to weigh them (the standing rule). Fewest entities first. Before any new column, check whether an existing column or a meta-table row already holds the fact: for (1), a citing table's effect columns joined by the member; for (4), a computed field, since the phase comes from the same LD lookup that fills r². A gap whose only case is one example row waits for a second case.

Residuals none yet.

RM450 — a snapshot pull takes no lock on the cache directory, so two pulls into one cache share their .part names and the pull record

Severity low · Status open — a patch, parked · Owner enricher (download._provision_snapshot, ensure_snapshot, cache pull) · Motivating case the pm's check request from RM437 on 0.9 (2026-11-17): the groups' pull record is rewritten without a lock, "as the core's is" · related RM437, RM445, RM231

Checked on main (5ff00802). download.py takes no lock of any kind. The one flock in the enricher is transaction.spec_lock, over a spec directory. A pull stages each file at a fixed name (<file>.part, release.json.part, manifest.json.part), renames it, and rewrites the pull record with atomic_write_text. Two pulls into one cache, such as an enrich provisioning while a cache pull runs, or two module runs sharing a cache, write to the same .part paths. One may rename or remove the other's partial file, and the last record written wins. The record is never torn (atomic), but it may describe the other pull's file set.

No incident. Nothing here was reproduced, and no run is known to have failed this way (@argue-from-incidents). That is why the severity is low and the status asks the pm whether to keep it. If kept, the shape is @flock-not-a-lockfile: an advisory flock on the cache directory's own descriptor for the whole pull, as spec_lock does for a spec, with the same warning where flock cannot hold. RM437's groups on 0.9 inherit it.

Class. A patch: no schema, column or vocabulary moves (P3), nothing authored (P9).

Parked, the pm (2026-11-18) (@argue-from-incidents-not-mechanisms): nothing reproduced it and no run is known to have failed this way. Unparks when a pull is seen to leave a wrong file or record, or a host reports running two pulls into one cache at once (the registry's proxy hosts are the likely first). The fix is then the shape above, about 30 minutes, and RM437's groups on 0.9 inherit it.

Residuals none yet.

RM453 — offline, the placement witness reads a merged rsID's ClinVar record as no record, so RM449's "witnessed neither" leaves rs80356821 unpositioned

Severity medium · Status open — a patch, a declared corrected derivation, built on main 2026-11-18, uncut, for 0.8.10 · Owner enricher (placement.compare_placements, the ClinVar witness's rsID key) and compiler (RM449's neither arm, resolution._settle_mirrored_spellings) · Motivating case the lead's 0.8.9 run (2026-11-18): pathogenic_clinvar enriched offline compiles strict except one row; the 0.8.9 cut states it (c8f3f3ed) · related RM267, RM398, RM449

What was seen. An enrich --offline of pathogenic_clinvar (the table deleted, ClinVar clinvar_2026-11-15, the release the committed table records) writes placement_verdict=unplaced on both spellings of rs80356821 at 11:5226762, the insertion C>CAAAG and the deletion CAAAG>C. The committed online table has the deletion same_event. RM449 reads "a witness was asked and placed neither" and fills neither, so --strict refuses the row as unresolved. 12 of the 13 mirrored sites agree between the two runs.

Why, read 2026-11-18 04:21 UTC. It is not an indel the offline run could not normalize: the pm's first hypothesis does not hold here. ClinVar clinvar_2026-11-15 files the deletion 11:5226762 CAAAG>C as VCV 15417 under rs281864900 and holds nothing under rs80356821. dbSNP merged rs281864900 into rs80356821 (/variation/v0/refsnp/281864900: merged_into: [80356821]). ClinVar's label is the retired id. The witness is keyed by rsID (load_witnesses(…, fresh_rsids)), so offline it holds no record for rs80356821, no allele has a counterpart, and both rows read unplaced. That is the witness's honest answer for "nothing filed under this id". It is not "the witness placed the rsID elsewhere". Online, the merge check (RM398's dbSNP record) is available and the deletion reads same_event; which leg supplies it is to be confirmed, since the row names clinvar.

Where the defect is, two candidates (the builder decides).

  1. RM449's neither arm. unplaced conflates "the witness holds this rsID at other alleles" with "the witness holds nothing under this rsID". Only the first is a witness that placed neither spelling. The second is closer to "not asked", and RM449's not-asked arm keeps the expansion. The compiler sees only the row's verdict, so telling them apart needs the enricher to say which.
  2. The witness's key. ClinVar's records under a retired rsID are invisible to the witness offline. The Ensembl snapshot carries synonyms (rs281864900 among rs80356821's?), which could let the offline witness follow the merge without the network.

The first is the safer patch, and the second changes what the witness compares. Neither changes a published value beyond un-refusing the row.

Class. A patch: no schema, column or vocabulary moves (P3). Severity medium: a module enriched offline (or by an older enricher) can refuse under --strict on a mirrored indel site whose ClinVar record carries a merged rsID. It is loud and names its rsID, and an online re-derivation clears it.

Sized (the pm, 2026-11-18): candidate 1, in the compiler, and RM449's neither arm goes. unplaced on every spelling of a mirrored site says the witness filed neither spelling under this rsID: no evidence for either, the same knowledge as undecided or an empty cell (@unreachable-not-absent; @absence-is-weighted-by-the-base-rate, since ClinVar files only some variants). So a mirrored site fills the spellings a witness placed same_event; with none, it fills every spelling, under RM449's warning, as the not-asked and undecided arms already do. No arm fills neither. This needs no new verdict member (the vocabulary is closed, and a member would be minor class), and nothing the enricher writes changes. It is a corrected derivation (rows that were withheld are written), declared in 0.8.10's entry. Tests: pathogenic_clinvar enriched offline compiles strict (0.8.9's xfail becomes a pass), and a site witnessed on one spelling still drops the other.

Candidate 2, the offline witness following dbSNP merges through the snapshot's synonyms, is a residual of this entry: it would make more offline verdicts same_event, but after the change above nothing refuses for want of it. It becomes an item if an offline placement elsewhere is seen to misread a merged rsID.

Built (triage, 2026-11-18). Candidate 1, compiler side, as the pm sized it (f3e6d6a1). RM449's neither arm is gone. At a mirrored site where no spelling was witnessed same_event, any mix of unplaced, undecided and empty, both spellings stay filled, said once per rsID under the borrowed rsid_expanded_to_multiple_loci ("placed neither as the same event", (RM453)). Nothing the enricher writes changes, and no verdict member is added. A one-sided witness still fills only the witnessed spelling. MirroredSettlement.withheld_all is removed, since no settlement leaves a row unpositioned any more. Tests in test_a_mirrored_indel_genotype_names_one_event.py:

  • pathogenic_clinvar with rs80356821's two verdicts set to unplaced, as the offline run writes them, compiles strict;
  • the both-unplaced and undecided/empty-beside-unplaced cases fill both;
  • the summary counts no rsid_no_hosting_locus.

Four fail on origin/main. Two of the lead's 0.8.9 tests pinned the removed arm and now state the new rule. The feature scenario moves to the borrowed code, and COMPILER.md, AGENT_NOTES and the CHANGELOG say it. 0.8.9's entries keep what 0.8.9 did.

Residuals candidate 2, the offline witness following dbSNP merges through the snapshot's synonyms, is left as the pm decided: this item no longer needs it, and it would change what the witness compares

RM454 — the enricher's local key compares never went through the one normalizer: constraint by gene, lookup_loci by rsID and contig, and SQL's lower(trim()) beside gene_key

Severity low · Status open — a patch, after 0.9 merges (its home, key_compare.KEY_COMPARES, is 0.9's) · Owner enricher (gene_metrics.lookup_snapshot, clinvar.lookup_loci, resolver.lookup_loci, key_compare) · Motivating case the RM430 follow-up's deferred findings (a) to (c), the lead's 0.9 run (2026-11-18, merged as 15668c1e) · related RM363, RM430, RM431

What was seen. 0.9 made each keyed function compare a key the same way on the local and the remote route (one registry, key_compare.KEY_COMPARES, feeds both routes' SQL, and a walk over REMOTE_LANES asserts equal answers). Its builder found three compares that are equal across routes but not across functions:

  • (a) the constraint snapshot compares a gene exactly (gene IN (?) in gene_metrics.lookup_snapshot), though RM363 lists gene_metrics.py among the sites moved to gene_key;
  • (b) clinvar.lookup_loci and resolver.lookup_loci compare an rsID and a contig exactly, so RS334 and chr11 miss, where RM431's keyed functions normalize the same keys;
  • (c) SQL's lower(trim(gene)) and Python's gene_key (casefold, every whitespace) differ on a tab and on non-ASCII text. Older than 0.9.

Why low. No incident: an authored row's rsID is one spelling (^rs\d+$) and VariantRow folds chr12 to 12, so a module never sends the spellings (b) misses. Only a direct caller of a lookup (a hint, a tool's user typing a key) can, and a miss there reads as absent. RM363 decided one normalizer per value; these are the sites its AST guard cannot see, because the compare is inside SQL.

The fix. Every keyed SQL compare in the enricher is declared in KEY_COMPARES, not only the routed ones. The SQL form and the Python normalizer give the same key, or the compare normalizes in Python before the query. One walk asserts both: over every declared compare, a fixture key in another case, with surrounding whitespace, with a chr prefix where it is a contig, answers as its normalized form does. A patch after 0.9 merges; triage builds it. It may widen what a direct lookup returns, never what a module resolves, so no derived table moves.

Residuals none yet.

RM455 — the pre-0.9 instances of 0.9's review classes: an unreadable pull record escapes its readers, three settings check batch_size only when built, mitomap_miss calls itself present over moved parents

Severity low · Status open — a patch, built on main 2026-11-18 (e63a8d8c, 5799c2c9, 78b8fb3f), uncut, for 0.8.10 · Owner enricher (download._pull_record, locations.pull_record_repo and pulled_from, the gnomAD, E-utilities and literature settings, mitomap_miss's status, cli/cache_commands._payload) · Motivating case 0.9's review round 1 fixes (the lead, 2026-11-18, merged as ea3dcf07), whose builder found these on code older than 0.9 · related RM430, RM431, RM437

What was seen. 0.9 fixed four error classes in its own code, and the same classes have older instances on main:

  • download._pull_record, locations.pull_record_repo and locations.pulled_from catch JSONDecodeError but not UnicodeDecodeError, so a pull record that is not UTF-8 escapes as a traceback (RM437's class 3);
  • the gnomAD and E-utilities settings and both literature clients never check batch_size, so one set to 0 answers nothing, silently (RM404's class 8);
  • mitomap_miss declares stale_parents but no stale reason, so cache status calls it present over moved parents (RM431's class 11). main has no stale state at all, so this one is RM431's on 0.9, not this patch;
  • cache_commands._payload keys on lane.name == "strchive" rather than a field of the lane (class 4).

The fix. Each takes the form 0.9 gave its own instance: catch UnicodeDecodeError beside JSONDecodeError; refuse a batch size below 1 where it is used; a stale reason for mitomap_miss, so status says stale; a lane field in place of the name. Each gets a test, and nothing a module resolves moves. Built on main and merged into 0.9 by the next sync, never on 0.9 (the trial rule: a defect older than the release is triage's patch).

Residuals mitomap_miss's stale reason → RM431 on 0.9 (main has no stale lane state, so it has no form here). Built with the rest: a fourth unreadable-record reader, locations.snapshot_record_missing, took the same fix; acmg took payload_filename beside strchive, since its resolver also finds a named root file.

Not format scope

Listed so they are not mistaken for format scope, and so nobody re-proposes them.

RM7 — Evaluation-output / report-card schema

Severity — · Status not format scope — a consumer contract · Owner consumer (just-dna-lite) · Motivating case verification harness (§1a)

For the verification harness — NOT a format task. Per-sample results are a measurement, so by the data-agnostic north star this is a consumer contract (just-dna-lite), listed here only so it is not mistaken for format scope.

Corpus (S115, 2026-09-26). The first output shape a shipped caller produced, recorded here as ground truth for whatever RM7 settles on (in the consumer, not the format): a four-state status (called / ambiguous / not_assessable / no_match — never silence, never a reference default); phenotype set only when every consistent diplotype agrees; the consistent candidates; a tri-state per-site evidence list (called / restored_hom_ref / no_call); and a phase_would_decide predicate. The one artifact question it raised — a per-gene defining-site set — resolved (b): it is the union of the gene's haplotypes rows, derivable by any reader, so it stays out of the format (@derived-not-stored).

Annotating core, not format scope (the 0.5 source assessment)

RM7 and RM13 are listed above so they are not mistaken for format scope. The same needs saying about roughly half of every annotation source assessed in 0.5 — the half that calls or interprets. A module supplies annotation tables; the measurement arrives from the consumer at query time, so none of the following can land in these libs no matter how useful it is:

  • Star-allele callers — PharmCAT, and Cyrius / PyPGx for the CYP2D6 case PharmCAT punts on. These turn a VCF or a BAM into a diplotype call: measurement. What does belong here is their data — the CPIC allele definitions PharmCAT ships — which is why the drafting helper reads them. Note that routing through PharmCAT does not launder the terms: its definitions are CPIC's, so the ClinPGx no-sale clause still applies.
  • Running splice or missense predictors, and choosing their thresholds. A SpliceAI delta of 0.2 vs 0.5 is an interpretation policy, not an annotation; RM23 carries the score and the dataset, never a verdict.
  • ACMG rule application and incidental-findings reporting policy. The format carries acmg_sf as a flag and (since 0.5) validates it against the published list; deciding what to report to whom is the consumer's.
  • Lay-language rendering. A module already carries the ontology CURIE and a human conclusion; turning a MONDO term into patient-facing prose is a presentation concern. Author-written lay text is a different thing, annotation data a curator signs, and is filed as RM302.

Cross-repo (tracked elsewhere): just-dna-marketplace — take just-dna-compiler as the M4 publish dependency; serve logs via the files endpoint; render the cross-version provenance union (aggregate.aggregate_provenance) on the module-detail view.

Trackers

The 1.0 cleanup (candidate tracker)

The compatibility policy — additive within a major, breaking cleanup only at a major bump, the two-step deprecate→remove default — is a durable rule in CONSTITUTION.md (Principle 3). This is the living tracker of concrete items queued for the → 1.0 break; add candidates as they surface.

Additivity has two axes. A new version may expand the column-set (new optional columns) and the row-set (one authored row compiling to several — e.g. a one-to-many rsid → one row per locus). Both are minor-legal: a new optional column leaves the authored identity untouched (it is omitted from content_signature) and moves only a recompile's artifact.digest, which P4 already scopes to a fixed compiler_version. Row-set expansion changes identity cardinality but is not a schema break: it is resolver behavior pinned on the compiler_version axis (P4 already pins the digest to the resolved reference), so the GRCh38 expansion ships now. Only the build-aware generalization (which/how-many loci per build, cross-build annotatability) is RM15. The idea is to pile genuinely rule-tripping edge-cases (requiredness demotions, retypes, identity-key semantics changes) on the 1.0/RM15 piles instead of forcing them into a minor.

The cadence changed on 2026-08-12 and this tracker is read under the new one (CONSTITUTION § 0.6 amendment). Retirement is deprecate in a minor, remove at the next major — so an item whose replacement already exists gets its warn-only deprecation in a 0.x release and disappears at 1.0, rather than being deprecated at 1.0 and lingering to 2.0. The exception is written into the principle: a deprecation must be actionable, so anything Principle 8 still makes mandatory — VariantRow.state, the pathogenic/benign booleans — cannot be deprecated while an author has no way to stop setting it. Those keep the old shape (demoted and deprecated at 1.0, removed at 2.0) because P8 blocks them, not because the cadence does. Every entry below that says "deprecate at 1.0" should be re-read with that distinction in mind, and moved forward where nothing blocks it.

Every item here also owes an upgrade line under RM52 — written when the item lands, not when the release is assembled.

Version-axis note: schema_version is "1.0" while the packages are 0.x (now 0.5.0). At 1.0, either align them or document explicitly that they track different things (wire format vs. package release).

RM135 — ProvenanceItem.outranks is superseded by the overlay, and one of them has to go

Severity low-medium · Status queued for 1.0 — filed 2026-08-28 by PROPOSAL_0_7, which decided the succession rather than the merge · Owner format

RM124's overrides.csv records an authored value beating a source, with prose. ProvenanceItem.outranks — {column: why}, shipped by S52 — records an authored value beating a source, with prose. They are one concept in two files, split on authored-versus-derived, and Principle 5 says decide before either grows a second field. The proposal decided: both stand in 0.7 and the unification lands here.

The rule that survives is the simpler one — the existence of an override in an authored table auto-beats the derived value, with no separate declaration to write. By Venn diagram the overlay's logic is a partial superset of what outranks allows and reaches it more directly, so once an author controls both derived overrides and authored tables the outranks knob has nothing left to do.

Why the deprecation is not in 0.7, and this is the part to re-read under the cadence note above. The two-step default would put a warn-only deprecation in a minor and the removal here. It cannot, yet: the overlay covers derived tables and outranks covers an authored cell, so until the overlay's semantics reach authored tables — which is this item — an author warned off outranks has nowhere to go, and Principle 3 says a deprecation belongs in a minor only where its audience can act on it. The warning ships with the replacement, in whichever minor extends the overlay to authored tables; the removal is here. That ordering is the item's own upgrade line under RM52.

module.version — refusing a digitless version

Severity low-medium · Status queued for 1.0 — filed here on 2026-08-21 by the RM103 split, and filed as a charter question rather than as a fix

Split off RM103, whose additive half stays open as a minor. normalize_version("abc") returns "0.0.0" — a legal SemVer and a plausible pre-release marker — and it reaches identity.version in a published manifest.json. Refusing it instead is the cleanest end state and is the reporter's implicit ask; the coercion for every digit-bearing case (v2, 3, 1.5, v1.2.3-beta) is RM17's decision and stays.

It is here because the two readings of the charter disagree, and nothing had made them argue. Precedent says minor: RM50 and RM48 both shipped new refusals in 0.6.0 as minor work, and INTEGRATION_0_6 § 1 lists them under "two checks can newly refuse an author's spec". Principle 8's purpose says major: its forbidden-moves clause exists so that nothing previously valid becomes invalid, and a spec that compiles today failing tomorrow is exactly that. The clause is written about fields — requiredness and retyping — and a new refusal on a value is an axis it does not name, which is why both readings are defensible and why neither is written down.

Filing it here takes the conservative branch by default. That is a decision worth revisiting deliberately rather than by precedent, because the answer governs RM50 and RM48 retroactively and every future check: if a new value-refusal is minor-legal, say so in the Constitution; if it is not, two 0.6.0 changes were mis-sized. Whoever picks this up should settle the general rule first and let module.version fall out of it.

trait_efo_id is named for one ontology and accepts every ontology

Severity low · Status queued for 1.0 — filed 2026-08-31 by the maintainer after the name misled a provider in this tree · Owner format (+ every consumer reading the column by name)

The column takes any ontology CURIE and always has: validate_trait_ids enforces only a PREFIX:LOCAL / PREFIX_LOCAL shape, the cell is multi-valued, and the field's own description reads "EFO/MONDO/OBA/HP trait ontology id(s)". HP:0000006, MONDO:0005265, DOID:1612 and OBA:2040158 are all legal today, singly or together. The name says EFO because it matches just-prs's column, and that is the whole of the reason.

A name is a claim, and this one is read as a restriction. The CIViC provider withheld a DOID: id from the column on the reasoning that a DOID in an EFO column would be a wrong identifier, and put it in conclusion prose instead — losing a joinable id on every row it drafted, since every CIViC germline row carries a DOID. That is the same class as an analogy in a field description being taken for a rule: the field was doing its job and the label was not. It has been fixed at the provider and the rule is now stated in SCHEMAS.md § Conventions, but the durable repair is the name.

It is here rather than in a minor because a rename is a removal plus an addition, and removal is what P3/P8 reserve for a major — a consumer selecting trait_efo_id by name breaks the moment the old name goes. The additive half (a new trait_id column) is minor-legal on its own and is deliberately not proposed: two spellings of one fact is the overloading P5 forbids, and a deprecation window that leaves both columns readable is exactly the state this tracker exists to end. So it waits, and lands as one rename with the removal.

Two things for whoever takes it. The successor name should not encode an ontology at all — trait_id — and the description should keep naming the accepted prefixes, because the next provider will read the name first. And the rename has to move with the reserved-namespace and authoring_reference surfaces together, since a consumer's column list is generated from the model rather than hand-kept.

stats' scalar counters read 0 where the table is absent, and cannot say "inapplicable"

Severity low-medium · Status queued for 1.0 — filed here on 2026-08-24 from S72, because the fix is a retype of published fields and nothing smaller reaches it

variant_stats derives variant_count, unique_rsids, study_count and the ClinVar counts from variants.csv alone, so a module led by any other kind publishes 0 for all of them. Measured on a 1,482-row pharm_variants.csv module: variant_count: 0, unique_rsids: 0, study_count: 0.

unique_rsids: 0 is the one that is simply false rather than merely narrow, and the reporter is right about why: rsid is the first authored column of pharm_variants.csv and 1,482 rows carry one, so the counter reports none of something that is there. variant_count: 0 for a module with no variants.csv is true and unhelpful; this one is untrue.

The ask is None where the table is absent, and it is a retype. Stats.variant_count and its siblings are int with a default of 0 in a published manifest.json, so widening to int | None breaks any reader that compares or sums them — P3 names retyping as major-only precisely for that, and this is the ordinary case rather than a stretch of the rule. The same holds one layer down for ValidationResult.stats: the dict is dict[str, Any], so nothing retypes formally, but its keys are a documented contract and changing 0 to None breaks the same arithmetic. Sizing it as a minor because "the field is only advisory" is the move RM127 recorded as the tempting one.

It is the right end state, though, and that is why it is filed rather than refused. The reporter's framing is our own rule in our own output: VerificationRecord's docstring says "subjects=0 with no skipped means the check ran and had nothing in scope, which is not the same as not running", and stats inverts it — a consumer cannot tell counted, and the answer is none from this counter does not apply here. A registry keying a facet off either inherits the collapse, which is the S57 failure one field over.

What shipped instead, and it is not a substitute. row_count (documented, family-independent) and table_rows (promoted from de-facto to contract) give a caller the honest number today, and the field description now says in terms that 0 in the scalar counters means no variants.csv rows and never no data. That makes the counters readable; it does not make them correct.

Decide it with the module.version refusal above, which is the other charter question on this tracker: both turn on how far P3/P8's field-shaped clauses reach, and answering one without the other is how two defensible readings stay unwritten.

VariantRow.variant_key / authored_ident are inside content_signature; the 0.6 stamped fields are not

Severity low · Status queued for 1.0 — align the two, one way or the other

Surfaced by RM43 (0.6) rather than designed: the three positional models gained stamped, parquet-only variant_key and authored_ident columns, and they had to be declared Field(exclude=True). Declaring them plainly moves content_signature on all five positional-table modules, because integrity.content_signature hashes model_dump(exclude_none=True) and a stamped field is never None — so it lands in the authored identity, which is exactly what the stamped-column mechanism exists to keep it out of. _build_table reads the values off the row instead, so the columns still reach parquet.

VariantRow's own two are not excluded and therefore are inside its content_signature, grandfathered: they predate the mechanism being generalized, and changing them now would move the authored identity of every SNP-core module ever published. So one model hashes its stamped fields and three do not, for no reason a reader could derive. Disposition: at 1.0, exclude VariantRow's two as well (the honest shape — a compiler-stamped value is not authored content) and accept that every 0.x content_signature moves once, under the major's documented upgrade procedure. The alternative — un-excluding the three new ones — is strictly worse: it would put machine-filled coordinates into the authored identity and defeat RM43's whole reason for existing. Owes an RM52 upgrade line either way, since a moved content_signature breaks content-dedup across the boundary.

VariantRow.state

Severity medium · Status queued for 1.0 — deprecate; remove at 2.0

Overloaded legacy field; a derived alias of direction since 0.3. Disposition: Deprecate at 1.0 (still read) → remove at 2.0, once consumers read direction/stat_significance. This keeps the pre-amendment shape for a reason, and is not stale: the field is still required, so a deprecation warning in a 0.x minor would fire on every module in existence and name nothing the author is permitted to stop doing. P8 is the blocker, not the cadence — the demotion and the deprecation land together at 1.0, and removal falls to 2.0.

state values alt / ref

Severity low · Status queued for 1.0 — drop from the read-vocabulary

Genotype-relative descriptors that never belonged; recoverable from ref/alts/genotype; not emitted since 0.3. Disposition: Drop from the accepted read-vocabulary at 1.0.

VariantRow.pathogenic / benign booleans

Severity medium · Status queued for 1.0 — deprecate; remove at 2.0

Lossy (can't express likely_*/uncertain); derived aliases of clin_sig since 0.3 (now materialized tri-state). Disposition: Deprecate at 1.0 → remove at 2.0. (clinvar provenance boolean stays.) Same P8 blocker as state above — required/authoritative today, so the deprecation cannot move into a minor however cheap warn-only is.

StudyRow.p_value: str

Severity low · Status queued for 1.0 — retype (p_value_num shipped in 0.5)

Untyped string holding a number; can't be compared/sorted numerically. Disposition: Add a numeric companion in 0.x if needed; retype/remove the string at 1.0 (breaking).

weights.parquet end column

Severity low · Status split by the charter amendment — wiring it is 0.6; removing it is 1.0

Always set equal to start — no source column feeds it. Disposition: wire it to a real end coordinate (an additive change to an existing optional column, so a minor) or remove it outright at 1.0 (removal is what the amended rule reserves for a major). Re-examined in 0.5 and deliberately left here rather than wired inside the window: wiring a second coordinate buys an off-by-one unless the first one's convention is unambiguous, and half of that was still open — every tier stored Ensembl's 1-based position while VariantRow.start's own description said "0-based", which is the text describe/requirements/reference print at an author.

That half is now closed: the authored start descriptions say 1-based VCF POS, and schema/tests/test_coordinate_convention.py pins the prose to what the minting code actually does. What remains is the genuine design question — whether an end is interbase-half-open (VRS) or inclusive (VCF-ish), which is the same choice RM15 has to make for a build-agnostic identity, so the two stay paired.

weights.parquet likely_pathogenic / likely_benign

Severity low · Status queued for 1.0 — remove; wiring rejected in 0.5

Always False; no CSV column feeds them — dead output. Disposition: Remove at 1.0, or wire to the clin_sig tier. Re-examined in 0.5: removal is the answer, and wiring was rejected. clin_sig is itself materialized into weights.parquet and derive.pathogenic_from_clin_sig already maps likely_pathogenic → True, so a wired column would tell a consumer nothing it cannot already read. That argument is unchanged by the charter amendment — wiring is cheap now and still pointless — while the removal stays major, which is the half the amendment does speak to.

VariantRow.weight vs effect_size

Severity low · Status queued for 1.0 — review only

Potential confusion — module-local score vs published magnitude (both kept, documented). Disposition: Review at 1.0 whether weight stays or is subsumed by effect_size. Evidence for that review is in § D5 of the consumer-note triage below: a real module (superhuman) authors weight on none of its 190 rows and declares no weighting: block.

sources.csv — the name, and the source column it collides with

Severity low (nothing is wrong; a reader has to be told three times) · Status queued for 1.0 — rename with the two-step, or decide explicitly to keep it

The file is a licensing and attribution ledger: one row per (source, layer), carrying the terms, the attribution text, license_sha256, and the three tri-state permission axes. It is the only file the compile licence gate reads. Nothing in the name says any of that, and it collides twice over — with the source column, which in resolution.csv/frequencies.csv/gene_metrics.csv/ literature.csv means "which link answered" (the overload RM33 already had to split, adding authority so the compiler had something to join on), and with the ordinary English sense in which studies.csv and literature.csv are also "sources". SCHEMAS.md now carries a three-row table disambiguating them, which is the tell: a name needing a table is a name doing no work.

The input half is done — RM51 shipped in 0.6.0: licensing.csv is an accepted spelling, sources.csv is deprecated (warn-only, read exactly as before), and four reference examples already carry the new name. Its ledger line is in RM52. What stays here is the half that genuinely breaks a reader: sources.parquet is in _OUTPUT_FILES and therefore inside artifact.digest, and consumers read it by name; manifest.sources is a published key. Renaming either is a removal, so both are major-only. The old CSV spelling retires on the amended cadence (Principle 3, 0.6 amendment): deprecated in the 0.6 minor that adds the alias, removed at 1.0 — deprecation is warn-only and needs no major, and an author carrying sources.csv can act on it the day they read the warning, which is the condition that makes a minor the right place for it.

Disposition: at 1.0, rename sources.parquet → licensing.parquet and the manifest.sources block → manifest.licensing, and drop the sources.csv spelling deprecated in 0.6 — with the upgrade line RM52 makes mandatory, which here is a file rename and a recompile. licensing.csv is the recommendation: it names what the file is for and what the gate reads it for, and it cannot be confused with a source cell. data_sources.csv is the conservative alternative (a smaller change in meaning, but it keeps the collision with the column and only lengthens it). provenance.csv is out — it collides with manifest.provenance, which is a different thing — and attribution.csv is out because it undersells the half that refuses a compile. Renaming the column is a separate and larger question and is not proposed here: SourceRow.source is inside its own fact set, so it is the row's key, and every other table's source already means what RM33 settled it to mean.

Note what this does not unblock: the file's shape is already right. A rename is legibility only, which is exactly why it waits for a major rather than justifying one.

fetched_at — the column says fetch, the value means write

Severity low (nothing has broken; the name mis-describes what is in the cell) · Status queued for 1.0 — bundle with the sources.parquet rename above, or decline explicitly

Seven sidecar models carry fetched_at — ResolutionRow, FrequencyRow, GeneMetricsRow, LiteratureRow, GeneValidityRow, ClinicalAssertionRow, SourceRow. The name says the row records when a source was fetched. It does not. Its own field description has to correct it in prose — "records when this row was last written by a pass, not when the source published anything" — and a description that opens by contradicting its field name is the tell, the same one the sources.csv entry above is built on.

What the value actually is, measured rather than argued. Every sidecar merge is never-clobber, so an already-recorded row wins and its stamp is never rewritten. record_source_terms run twice against one spec directory leaves the file byte-identical; only deleting the sidecar re-stamps (2026-08-16T02:02:24Z → …02:02:27Z, with source_signature unchanged across all three states). So on an ordinary re-run — including one that really did go and ask the source — the column records no fetch whatever. It records when this row's facts were first set. That is a useful thing to have and a reasonable thing to publish; it is simply not what it is called. Established independently at S7, which probed the same setdefault and answered the behaviour question; the naming question was never put.

Be honest about the evidence: no incident is attributable to the name. S7's proximate cause was SCHEMAS.md calling artifact_digest the "content identity", and it was fixed there. Nothing has misread fetched_at itself. That is exactly why this is low severity and why it does not justify a major on its own — and why the disposition below is ride along, not schedule.

What it costs, checked. fetched_at is outside all seven fact sets (verified against RESOLUTION_FACT_FIELDS and its six siblings), so no signature moves — not content_signature, not any *_signature. It is a column in six parquets (resolution.csv alone has none by design), so artifact.digest moves on every module carrying a sidecar. Blast radius at 0.6: 36 occurrences in schema/src, 5 in compiler/src, 34 in enricher/src, 51 across the suites, and 27 reference-example files.

The rename obliges one semantic decision, and it is the substance of the item. Two writers rewrite a recorded row without touching the stamp: licensing.withdraw_stale_dataset blanks dataset, and provenance.stamp_draft_digest re-labels draft_digest. Under fetched_at that silence is plainly correct — nothing was fetched. Under updated_at the row was updated and the stamp did not move, which is a new small untruth. Recommendation: leave both silent and say so in the description — the value dates the row's facts, and a provenance-column rewrite is not a new fact. That reading is also what keeps the delete-and-re-derive drift check sharp (see MODULE_LIFECYCLE § 5.1): a stamp that moved whenever any cell was rewritten would stop separating "these facts are from that moment" from "somebody touched this row". If that reading is adopted, the honest name is arguably recorded_at rather than updated_at, and the choice should be made deliberately rather than by reaching for the database convention.

No 0.x deprecation step, and the cadence's own condition is the reason. Principle 3's 0.6 amendment puts a deprecation in a minor only where its audience can act on it. Nobody authors this column — all seven writers are machine passes — and an author cannot pre-emptively rename it, because extra="forbid" rejects the unknown column. A 0.x warning would therefore be a finding no authored edit can clear, which this project treats as a defect wherever else it appears (P5). So: a straight rename at the major, mitigated on the input side rather than by a warning.

Mitigation — accept the old spelling on read through the 1.x line. The writer emits the new name; the loader keeps accepting fetched_at as a deprecated input spelling, so a hand-maintained or downloaded 0.x sidecar keeps loading and the author-side upgrade route is no action needed. This is RM51's shape one layer down, and it is what makes the ledger line below cheap.

Why this is not the column rename the entry above declines. That one refuses to touch SourceRow.source because it "is inside its own fact set, so it is the row's key". fetched_at is the exact opposite: outside every fact set, keyed by nothing, joined on by nothing, derived from by nothing. The stated reason for declining there is the reason this one is cheap.

Disposition: rename at 1.0, in the same change as sources.parquet → licensing.parquet. That item already moves artifact.digest on exactly the modules this one would, so bundling spends a cost already being spent and the two share one upgrade line; taken alone this item would be a digest move for legibility, which is not a trade worth making. Decide updated_at vs recorded_at when the semantic question above is settled — the two names encode different answers to it. If the sources.parquet rename is declined, decline this one with it.

Deprecated flag/vocab aliases

Severity low · Status queued for 1.0 — collapse to the canonical vocab

Any transitional vocab kept for 0.x compat (e.g. the trimmed-vs-full state set). Disposition: Collapse to the canonical vocab at 1.0.

ModuleManifest.authors: list[str] + free-form curator

Severity medium · Status queued for 1.0 — fold into RM14's record

Flat and overloaded — no role (created/edited/audited), no kind (AI/human); Defaults.curator smuggles kind via its "ai-module-creator" default. Superseded by the structured authorship record (RM14) once it ships. Disposition: Keep both as derived projections through 0.x (P8); at 1.0 fold authors into the structured record and drop the kind-smuggling curator default.

StudyRow.pmid required + PMID-shaped

Severity medium · Status queued for 1.0 — a requiredness demotion

Mandatory pmid (must parse to a real PubMed id) rejects DOI-only provenance — preprints (bioRxiv/medRxiv), books, theses, datasets. Demoting a required field is P8-forbidden in-major, so adding optional doi (RM11) alone can't unblock it. Disposition: doi-first at 1.0: make pmid optional/legacy and require ≥1 of {doi, pmid} (every citation has a stable id, not necessarily a PMID; the reverse holds). Requiredness change → major-only. Pairs with RM50, which shipped the PMCID axis in 0.6 and left the LiteratureRow key question open (its re-key on a general citation id option), so this entry now carries both questions: what a study row may be authored with, and what the pmid-keyed sidecar does with a row that has no PMID. Settle them in one release.

Compiler ensembl_cache deprecated shim

Severity low · Status queued for 1.0 — remove the parameter

0.5 already moved the whole DuckDB resolver + cache-location into just-dna-enricher and dropped duckdb/platformdirs/python-dotenv from the compiler (it is now pure-Python; resolution is the resolution.csv table). What remains is the compile_module(ensembl_cache=…) surface, kept as a deprecated shim that emits DeprecationWarning and routes to the enricher via a guarded import. Disposition: Remove the ensembl_cache/resolve_with_ensembl params outright at 1.0 (internal call, not the wire/artifact contract, so additive-within-major does not protect it).

Coordinate-first identity (option C)

Severity — · Status ✅ resolved in 0.5 by VRS — kept for traceability

The objection was that a coordinate key is build-baked. A VRS allele id is not: it names its reference sequence by refget accession, so it satisfies RM15's own reconsideration condition. variant_key now derives from the VA for a resolved substitution; rsid-keyed, position-only, indel and multi-allelic rows keep their previous keys. Disposition: Done, in 0.5.0's pre-publication window — an identity-semantics change is major-only because variant_key sits in artifact.digest, and that gate is publication, not the version number: 0.4 is the published line and 0.5.0 never shipped, so it rode the same one-time re-baseline as the alt-carrying key. No published artifact moved.

Parking conditions (the gate audit)

A deferred item states what would unpark it. A gate is only honest when the thing that satisfies it can exist while the item is still parked — and the failure mode this tracker exists for is the gate that cannot: one whose only possible satisfier is a tool built against the very table the item is refusing to build. Read narrowly, such an item is not deferred, it is closed, and nothing in the entry says so.

The rule, for whoever files the next parked item. Write the gate, then ask who or what satisfies this, and can they exist today? If the answer names something downstream of the unbuilt thing, the gate is circular and the entry needs a different one — usually the shape question hiding behind the demand question, which is answerable now.

Swept 2026-09-13 across ROADMAP.md, ROADMAP_0_8.md and ROADMAP_1_0.md, prompted by the maintainer after RM16 was caught. Every item carrying a parked-on / would-unpark clause was checked. One circular, one marginal, one possibly already satisfied, and the rest sound — the sweep's value is mostly the eight it cleared.

Re-sized 2026-11-05, after the 0.8.0 cut (the pm). 22 items were parked out of ROADMAP_0_9.md into Active items. The 15 rows from RM261 down are new. RM16, RM23, RM28, RM68, RM122, RM164 and RM238 already had rows here; their verdicts stand, and their entries now live in this file.

Re-sized 2026-11-16 (the pm), for smaller minors (desk #326). RM430 and RM431 left this table for 0.9's scope, where their gate, the maintainer naming remote reads a zone, was met on desk #261. Seven items were parked out of ROADMAP_0_9.md; their six rows (RM345 and RM286 share one), and one for RM412's parked probe, sit above the 1.0 queue's row, all judged Sound.

Item Gate as written Verdict
RM16 — authored PRS weights "a real consumer" — one that combines authored weights into a score Circular. No such consumer can exist before the table does. Re-parked on shape 2026-09-13. Restated 2026-11-05 by the pm as the shape question, which needs nothing built: Sound now.
RM28 tail — the two remaining cofactor classes (ancestry, family structure) "Neither is built until a real module needs it" Marginally circular, the same family. A module cannot need a cofactor class it has no way to express, so read strictly nothing satisfies this. It survives on the looser reading — an author saying "I want to write this and cannot" — which is what the clause should say. Low stakes: both classes are on-demand-only and shapeless by design.
RM122 — the measure lookup as a public function "what it waits on is a caller, not a decision" Sound but possibly already met, and nobody has checked. The rule is normative prose in SCHEMAS.md today, so a consumer can implement it without the function — which means callers may already exist and be silently disagreeing, which is the exact failure the item predicts. Ask just-dna-lite whether it implements the measure lookup before treating this as waiting.
RM238 — per-tissue eQTL "reopen this with a consumer, never with an argument" Sound. Somebody can need a per-tissue eQTL answer without our table existing; they would be asking for the data.
RM164 — heteroplasmy.csv with no source "Reopen it with a source, never with an argument" Sound, and the cleanest of the set — a gate on a measured fact about the world, independent of anything here.
RM28 — the predicate half parked on a corpus Sound. A corpus is counted off other people's published data; it has three entries and none of them needed us to ship anything.
RM23 — predictor scores "the acquisition measurement done, and a decision on per-transcript grain" Sound. Both are actions available today; the item is unstarted, not blocked.
RM68 — a drafting provider off GRCh38 "an author with a non-GRCh38 module saying which outcome they wanted" Sound. Such modules exist — reference_examples/grch37_build is one.
RM279 — a conclusion checked against its own row (filed 2026-09-27) a measured precision for the locus-restricted rule Sound, and met on 2026-09-28: 10 of 10 real over 653,706 rows, shipped as a hint.
RM309 — a state its own conclusion negates (filed 2026-09-28) a hand-judged precision over the reporter's six curated modules and the registry corpus Sound. The corpus and the rule both exist today, and RM279 met the same gate the day this was filed.
RM388 — placement readers bypass usable_loci (parked 2026-11-05) a real table reaching compile with a row off the module's build or a not_found row with coordinates Sound. Anyone who hand-writes or merges a resolution.csv can produce one today; nothing needs the fix to exist first.
RM390 — the enricher does not say what compile will refuse (parked 2026-11-05) an author reporting an enrich that passed and a compile --strict that refused Sound. The refusal already fires at compile, so the report can come in through Sn today.
RM261 — strip comments from the enricher wheel (parked 2026-11-05) the wheel's size becomes a reported problem, or the packaging is reworked Sound. Anyone installing the wheel can report its size today.
RM248 — a declarative report schema (parked 2026-11-05) a consumer asks for a report a module cannot declare Sound. The renderer holding the problem (just-dna-lite) exists; it would describe a report, not our schema.
RM181 — a moved digest says nothing about what moved (parked 2026-11-05) a consumer reports a moved artifact.digest the signatures could not explain, or the 1.0 hash-family review Sound. Digests move today.
RM188 — the competitor survey (parked 2026-11-05) the maintainer names competitor parity as a minor's zone Sound, and not this kind of gate. A decision, available any day.
RM237 — a recurrent-CNV region has nowhere to be written (parked 2026-11-05) an author wants to write a recurrent-CNV region and cannot Sound on the looser reading, as RM28's tail: an author saying "I want to write this and cannot".
RM240 — the ClawBio nutrition panel in the corpus (parked 2026-11-05) the corpus is asked for a nutrigenomics example, or RM188 unparks Sound. The translation is already run.
RM266 — the phase-confusable pair set is dropped (parked 2026-11-05) a consumer asks for the pairs as data rather than as a warning Sound. S116 already came in through Sn.
RM281 — clinicalVariants.zip unread (parked 2026-11-05) an author needs a row that archive carries and summaryAnnotations.zip does not Sound. Both archives are public today.
RM419 — pacing gates per instance beyond NCBI (parked 2026-11-08) a rate-limit refusal from a listed service in a run where two of our clients shared a process, or the registry running enrichments as threads in one process Sound. Both can happen with today's gates; nothing needs the fix first.
RM282 — the PGx lane has one licence class (parked 2026-11-05) someone runs the source survey its entry describes Sound. The survey needs nothing built.
RM284 — no run-level place for a skip in verification.json (parked 2026-11-05) a reader needs to know, from verification.json alone, that a run could not reach a source Sound. Skips happen today.
RM287 — ClinGen dosage has no snapshot lane (parked 2026-11-05) an offline deployment reports a dosage pass that never ran Sound. skipped_offline already says so in every offline run.
RM296 — a PRS has no condition → conclusion table (parked 2026-11-05) an author writes a PRS module that needs a score → conclusion band Sound. The author holds the problem; pgs.csv exists.
RM298 — no per-row ClinPGx provenance on pharm_variants.csv (parked 2026-11-05) the currency check withholds on a real mixed module whose operator needed the answer Sound. RM297's warning fires today.
RM306 — no exon structure in the tier (parked 2026-11-05) a real module's AlphaGenome rows need an exon-level gene attribution Sound. The AlphaGenome lane ships.
RM378 — errors carry no code (parked 2026-11-05) a consumer needs to tell apart refusals RM377's two phrases do not cover Sound. S129 already came in through Sn.
RM405 — the offline floor is four unindexed snapshot queries (parked 2026-11-07) a consumer reports offline enrichment time as a problem, or the snapshot builder offers an rsID-sorted layout Sound. The Sn inbox is open, and the layout is another repository's to offer.
RM406 — dbSNP witnesses only REST's subjects (parked 2026-11-07) RM398's build counts dbSNP-versus-snapshot disagreements on the registry's replay, or a consumer reports a misplacement outside REST's subjects Sound. RM398 is the next patch and its build step produces the count; the Sn inbox is S117's and S132's channel.
RM424 — a verification record carries no binding of its own (parked 2026-11-09) a reader asks which checks are current without re-running them, or wants the history RM423 drops Sound. Readers of verification.json exist today, and S139's reporter already proposes approximating the first from checked_at.
RM429 — GwasResult carries no count of page-size refusals (parked 2026-11-11) a run that reports a counted refusal, or a consumer that cannot inject its own GwasCatalogClient Sound. The client counts refusals today, and the Catalog can lower its ceiling at any time.
RM302 — one conclusion for every register (parked 2026-11-16) a consumer's report asks for the authored wording of a second register, or an author writes two registers into one conclusion Sound. A report already shows one conclusion in a table, a card and an explanatory mode (the entry's motivating case), so a consumer can ask without the columns existing.
RM350 — HLA names checked for grammar only (parked 2026-11-16) an HLA name in a module that IPD-IMGT/HLA never assigned, reported Sound. A module can carry such a name now; RM334's grammar check passes it. The licence probe needs nothing built.
RM294 — STRchive's grade has no place (parked 2026-11-16) a request for the source's grade beside the tier, or a Refuted locus in a published module Sound. Both can happen with today's tables.
RM345 + RM286 — licence shapes SourceRow cannot state (parked 2026-11-16) a gate deciding a real module's licence wrong, or the registry asking for a shape as data Sound. Modules drafting from AlphaGenome, Open Targets and PharmVar exist or can be written today.
RM358 — a DOI-only citation is never pinned (parked 2026-11-16) a pmid=NONE citation a downstream check needs pinned, or a bioRxiv full text an author cannot reach Sound. Both occur with today's literature tools.
RM360 — four longevitymap v2 rows without a shape (parked 2026-11-16) the next v2 version, or another module, needs one of the four rows Sound. An author meets the gap while writing, before any column exists.
RM412 half 1's per-lane release probe (parked 2026-11-16) an operator reports a rebuild that downloaded a source only to learn nothing had changed Sound. Every rebuild today downloads first; an operator can meet this before the probe exists.
RM450 — a snapshot pull takes no lock (parked 2026-11-18) a pull seen to leave a wrong file or record, or a host that runs two pulls into one cache Sound. Either can happen while it is parked; nothing needs the lock to exist first.
RM15 / RM69 / RM52 and the 1.0 queue a major-version bump Sound, and not this kind of gate. A release boundary is a schedule, not a satisfier.

What the sweep did not find, stated so it is not re-run for a while. No gate anywhere is parked on a consumer report that would have to describe the unbuilt thing — the Sn inbox is a channel consumers can use before we build anything, and every item that names it is asking for a report about a problem, never about a solution. That is the structural reason only RM16 failed: it is the one entry whose gate named the user of the output rather than the holder of the problem.

Audit findings parked without a number (until a real run meets them)

The maintainer's rule (2026-11-05): "judge how frequently would any bug would realistically fire. For once per known universe scenario defer to 'when roadmap is empty/someone steps on this defect'." The findings below were confirmed by a pre-cut audit, but no real run or example has hit them. Each gets an RMn only when it unparks: a run, a consumer or a test meets it, a scope the maintainer highlights touches it, or the roadmap is empty. One line each, so the finding is not re-reported as new.

Audit round 3 of 0.8 (2026-11-05, at 64c835b9), sized by the pm:

  • A client built in a function and never closed: clinvar_gkm.py lane_currency, caches.py, clinvar_gkm_build.py, cli/clinvar_gkm_commands.py, currency.py (two sites). Nothing leaks past a short process. The guard in test_cli_surface.py skips a function with no close() at all. Unparks when a long-lived process (a server) holds the enricher.
  • The S85 warning "recorded as unresolved" fires for an rsID+coordinate pair (written authored/resolved; since 0.7) and for an rsID-only subject whose placement --rederive carries (RM382). Wrong words on a correct table. Unparks on a report of it.
  • RM391's test fence lets a loopback HTTPS_PROXY through. No proxy is set on any machine that runs the suite. Unparks when one is.
  • The regex timeout counts the worker's start-up (0.4 s under forkserver), so test_literature.py's catastrophic-regex test cannot fail. Since 0.7. Unparks when that bound guards a real regex.
  • A ClinVar witness on another contig reads unplaced (placement.py events.counterparts needs the same contig). Measured: 0 of 145 corpus rows. Unparks on the first real one.
  • test_acmg.py depends on the working directory (fails from enricher/). Since 0.7. Unparks when the suite runs per package.
  • Europe PMC's two DOI lookups raise the outage type on a 410 (pmids_for_doi, record_for_doi); the recorded reason is still retired, so only the type is wrong. RM361's guard checks the module, not each method. Unparks when a caller catches by type.
  • net.py's retired-endpoint message advises "upgrade just-dna-enricher" even when the operator chose the release or URL (RM361's amendment, 0.8). Rare; the advice is wrong only for a hand-picked release. Unparks on a report of it.
  • Two leftover import guards for a core dependency in download.py. Dead code, no effect.
  • A malformed legacy rows scalar rides along as an extra in snapshot.py (RM303). No instance.
  • STRchive's snapshot has no prefix and no pull record (SNAPSHOT_FILE_GLOBS has no strchive). Unparks when STRchive is pulled through the shared path.

Reserved namespace

Because backward-compat makes column names and vocabularies permanent within a major (CONSTITUTION Principle 5), a name expected to become a real module column later is reserved against the one-way door and must not be claimed early or smuggled in as flags. This list is only for genuine anticipated module-side axes — it is not a catalogue of names that "may not appear" (that space is unbounded and pointless to enumerate; barring caller would be as arbitrary as barring pasta_recipe). Audit every new name against this list before adding it.

Enforced now (the live set is just_dna_format.vocab.RESERVED_NAMES_0_4). Every authored model inherits AuthoredModel, which sets extra="forbid" (rejects any unknown column) and runs the reject_reserved before-validator, so a reserved name fails with a specific diagnosis — what it is reserved for + that a release may claim it (vocab.RESERVED_NAME_REASONS) — while a random/misspelled column gets the generic "extra inputs not permitted": - reference_db — a module-side hint naming which reference database the app should join this annotation against when several exist (implicit Ensembl for variants / ClinVar for clin_sig today; a module may pin it, e.g. a specific PharmVar release). Annotation-side addressing, a real future axis. - callable_element / quality_element — added in 0.6 by RM54, which built source_element on the binning tables and deliberately did not build these two companions on VariantRow's pointers: no module points callable_from or quality_from at a multi-valued field, and an authored column is full cost. They are reserved rather than merely absent because the symmetry makes them guessable — an author reasoning "if source_field has one, callable_from must too" should hear what the name is held for, not the generic stray-column message. (tradeoff was reserved here from 2026-11-03 and MOVED on 2026-11-04 on 0.8 (RM327): the verdict that an allele is protective on one trait and risk on another, which the maintainer decided is computed, never authored. Since the compile writes it on meta_resolved.parquet it sits in vocab.MISPLACED_COLUMN_REASONS, the map for a column real on a generated table, and an authored tradeoff is still refused, with that reason.)

(callable_from was reserved here through 0.4 and is now BUILT as a VariantRow column in 0.5 (RM6). A built name must leave this list: reject_reserved refuses a reserved column at author time, so leaving it would make the very column the release added unwritable.)

(caller / caller_version were reserved through the 0.4 draft as a "provenance triple" (round-2 Q2) but are dropped: they name which tool produced a call — a consumer-side measurement, never module annotation — so there is no future module axis to hold, and barring the bare name is arbitrary. A consumer records them on its own call data; a module never carries them, and extra="forbid" rejects them like any stray column. reference_db stayed because it has a real annotation-side meaning above, not the caller-provenance one it was first reserved under.)

Planned future annotation axes (documented intentions, not yet in the enforced set above — they are rejected generically by extra="forbid" today, and get a slot + a specific diagnosis only when a release actually commits to building them): - consequence — VEP molecular consequence (Sequence-Ontology term, e.g. missense_variant). Distinct from direction (phenotypic) and clin_sig (clinical). Never repurpose the bare word effect for it. - impact — VEP impact {HIGH, MODERATE, LOW, MODIFIER}, derived from consequence. (Built in 0.8 as a table, not a column (RM236): consequences.csv → ConsequenceRow, one row per (variant_key, transcript), carrying impact (closed, the four classes) and the consequence as Sequence Ontology accessions in consequence_so_ids, the frequency precedence again: a consequence is per transcript, and a column on the variant row would have to pick one. The bare name consequence stays held for a term label (missense_variant), which a mapping from the accessions would fill.) (allele_frequency + af_population were listed here and are now built in 0.5 as a table, not a column — frequencies.csv → FrequencyRow, one row per (allele, ancestry group). A column pair could carry one number for one population; frequency is inherently per-group, and flattening it onto the variant row would smear two axes together. So the planned axes are retired rather than shipped. Gene-level constraint arrived beside it as gene_metrics.csv. See SCHEMAS.md and USE_CASES.md §6.)

(doi, provenance_quote, and provenance_regex were reserved here for RM11/RM12 and are now built as optional StudyRow columns in 0.4 — so they are absent from this list. The doi-first flip that relaxes the mandatory pmid remains a 1.0 item; see the 1.0-cleanup tracker.)

(The ploidy / non-SNV quantities that were reserved through 0.3 — allele_fraction / heteroplasmy, repeat_count + repeat_unit, copy-number dosage — are built as the 0.4 binning primitive; the hemizygous genotype case ships via the widened single-allele genotype. Symbolic/structural alleles remain open as RM5.)

The idea-book

Freeform suggestions — the 0.5 idea-book

The consumer's grounded 0.5 ideas (kept inside the one constraint: VCF-based, possibly augmented on top) came from the round-2 thread, retired on 2026-08-18 with its per-item disposition recorded in history/CONSUMER_SUGGESTIONS_HISTORY_PRE_0_6.md and its prose in git at 635da8c; each idea was run through the what-blocks lens in USE_CASES.md §1. Standing dispositions:

  • 3a — module declares where its measurement lives in a VCF. ✅ Taken early: source_field shipped in 0.4 (an optional, |-alternatable field-name token on every binning table — a declarative pointer, not an expression, inside Principle 1), and qualified with its namespace since 0.6 (RM53). The "zero glue" claim as first written was falsified by the schema it described, and the prose knew something the data model did not: it spells the pair INFO/RU and FORMAT/REPCN with the namespaces attached, which is how a VCF user writes them and how the reader of this line understood it — while the column accepted a bare token only, so RU and REPCN reached a consumer with the half that disambiguates them stripped off. 0.6 accepts the qualified form, warns on a bare key that INFO and FORMAT both define, and adds source_element for the second half the claim also assumed away: REPCN carries both repeat alleles, and the clinical rule for a dominant expansion is the larger, which no pointer could previously say.
  • 3b — modules as a deterministic verification harness (run a panel against N VCFs, emit a byte-diffable report-card). The strongest idea, and it needs nothing from the format: a panel is already a module, source_field names the field to read, artifact.digest makes the before/after diff trustworthy, and the mandatory unresolved/callability contract stops a no-call masquerading as a mismatch. It is a consumer feature (just-dna-lite); the format only supplies properties it already froze. Recorded as an enabled use case, not a gap.
  • 3c — augmented-VCF as the landing pad for cracked short-read loci (a synthetic <STR> record with INFO/RU + FORMAT/REPCN + custom evidence fields, consumed through the same source_field=REPCN path). Endorsed as the interface — the format binds to the VCF, it does not reinvent it. Consuming the symbolic alleles themselves is RM5.
  • 3d — smaller VCF-native ideas: callability three-state → RM6; phasing-aware panels → already expressible (the phased flag + VCF PS/HP); trio/de-novo assertion → RM10.

Parked in 0.5 (recorded so they are not re-proposed as if new)

  • Enricher co-authoring (permission-gated writes to authored files, not just sidecars). Attractive — it would let a stale rsID or a missing DOI be fixed where it actually lives instead of only being reported — and deliberately not taken, for a reason stronger than tidiness: content_signature is defined as pre-resolution and reference-independent ("computed from the rows before resolution, so recompiling against a different/complete reference does not change it"). If a network fetch could edit variants.csv, the content-dedup identity would become network-dependent and that documented property would simply be false. A secondary problem: authorship records who wrote the module, and an enricher that edits rows either falsifies that record or must add itself as an ai/agent contribution — coherent, but a much larger design than it first looks. Revisit only with both answered.

The drafting helper is not this item, and the line between them is one word: mutate. The 0.5 helper appends rows a source publishes into an authored CSV; it never rewrites a cell that is already there. Appending happens at authoring time and leaves content_signature a function of the authored bytes exactly as before — the property that would break is the one where a fetch changes the meaning of rows the author already wrote. So a row whose natural key is already present is reported, never overwritten (drift on existing rows is the cross-check pass's job, pgx.enrich_pgx), and the helper stamps no authorship: it transcribes a published table, and the human owns the module. Dedup keys on the compiler's own _TABLE_DUPE_KEYS, so an append can never produce a row the compiler would then reject as a duplicate, and rows are appended at the end — authored row order is preserved through compile → reverse → recompile, so re-sorting an existing file would move a compiled module's digest.

This bullet is a refused mechanism, never a home for a defect (added 2026-09-27, after the postmortem of that day). Most documents cite it for what it is: the reason a repair that would rewrite an authored cell is refused. One used it as a home. RM31's genotype-frame residual pointed here, and that defect went unowned for 55 days. Its consumer workaround is now in CONSUMING.md's indel warning, and the artifact-side answer is RM270. A defect that needs an authored cell rewritten gets its own RMn, and names this bullet only as the repair it cannot use.

  • Escalating the ClinVar clin_sig cross-check when the disagreement is with an expert panel. Tempting, because a VCEP or practice-guideline assertion genuinely is a different kind of claim from a one-star submitter's, and the snapshot already carries review_status/review_stars to tell them apart. Not taken: this is the one check that warns in both modes on purpose, because failing a compile over a clinical disagreement makes the format arbitrate a clinical dispute, which the data-agnostic charter forbids — and a curator who has read the primary literature and disagrees with a submission is doing their job. The tier is already surfaced (clinical.ClinSigFinding.confidence puts it in the message), and persisting it as queryable data is RM25. Surface it, let the consumer route on it, do not decide for them.

  • An offline allele-frequency snapshot. The obvious symmetry with the ClinVar and gene-constraint snapshots, and it does not work: gnomAD v4.1's sites VCFs are 58 GB (exomes) and 742 GB (genomes), so there is no slice to ship at any useful coverage. Frequency is therefore the first and only online-only link in the chain. This is not a reproducibility hole — once frequencies.csv is written it is the pin, and every later compile reads it offline and deterministically. Revisit only if gnomAD publishes a small pre-aggregated frequency release.

  • HGVS string generation (c./p. notation). ga4gh.vrs's extras pull hgvs transitively, so it would be available — but HGVS generation is its own feature with its own argument (which transcript, which reference, how to present ambiguity), and taking a dependency for indel normalization does not commit to shipping it. Deferred as a feature, not blocked by tooling.
  • Multi-build VRS minting. A second refget table beside REFGET_GRCh38; the remaining half of RM15.
  • ~~dbSNP obsolescence / merge checking~~ — built in 0.5 as identifiers.check_rsids, wired into enrich() behind --verify-rsids. Two corrections to what this entry used to claim, both found by probing: it is not "detectable two ways" — Ensembl REST resolves some merges (rs77121243 → rs334) and returns HTTP 400 on others (rs3216883, which dbSNP correctly reports as merged into rs3051860), so Ensembl alone would misclassify a merged rsID as unresolvable. NCBI esummary db=snp is the oracle, batched and authoritative. See the stale-identifier collision below for what is done with the answer.
  • Sex-stratified frequency counts. gnomAD serves nfe_XX/XY; sex is a second axis, and folding it into population would be the state-overloading mistake again. A future sex column on FrequencyRow is the additive shape if it is ever wanted.
  • google-re2 for provenance_regex matching (candidate, only if the current bound proves insufficient). The enricher matches provenance_regex against fulltext with stdlib re inside a killable child process (literature.regex_matches). That is a bound, not a linear-time guarantee, and the honest reason it is enough today is that the threat model here is a curator writing a slow pattern by accident — the pattern comes from the module being enriched and the document from a public archive, on the author's own machine, not an attacker meeting an arbitrary document.

The reason not to switch pre-emptively is capability, not cost. Real fulltext has periodic structure — repeated section headers, boilerplate, tabular runs — and pinning a quote inside it often needs a lookahead or lookbehind. RE2 does not support either, so adopting it would narrow the pattern language the format offers authors, in exchange for a guarantee the subprocess already approximates. Revisit if re exhibits problems the process bound does not contain (a pattern that wedges a worker often enough to matter, or memory blow-up rather than time). If it happens, the shape is: keep re as the default, add google-re2 as an optional accelerator, and record which engine ran — never silently change which patterns match. - Fulltext beyond the open-access subset (candidate, cost-driven). OpenAlex and Unpaywall can point at a green-OA repository copy for some closed articles. Probed and not taken: the closed paper tested (10.1038/s41580-019-0134-2) has is_oa: false with no location at all, and the copies that do exist are PDFs — which would mean a PDF-parsing dependency in the network tier for a partial improvement on a check that is already labelled partial. The abstract fallback shipped instead, which costs nothing (the abstract is already in the Europe PMC response the pass makes) and covers four of five non-OA papers. Revisit only if authors report quotes that live in the body of closed papers often enough to matter. - ~~Google Scholar for citation existence~~ — rejected, not deferred. It publishes no API, and automated querying violates its terms of service and is IP-blocked in practice. Crossref (DOIs, including preprints/books/datasets) and PubMed (indexed literature) cover the same ground through supported interfaces.

The stale-identifier collision (design note, 0.5)

An obsolete authored rsID forces a choice that Principle 7 and "keep the module current" pull opposite ways on, and it is worth writing down before anyone implements the lookup.

weights.parquet carries both variant_key and rsid, and for an rsid-authored row both are the authored label. Writing the updated label into the artifact is not a one-time digest move — it is an identity migration performed by a network lookup: reverse would then emit the new rsID into variants.csv, the next compile would key on it, and variant_key itself would change. The module's identity would drift without any authored edit, and the round-trip would stop being a fixed point.

So the rule is the one every other check here follows: report, never repair. Severity follows the mode — best_effort warns and compiles with the authored label (digest stable, round-trip intact), strict refuses, on the grounds that an all-or-nothing artifact should not be built on an identifier its own source has retired. Failing is the honest move because it pushes the fix to where it belongs: an authored edit.

That last clause is now load-bearing rather than rhetorical, and it is what qualifies this check for a mode ladder at all. This entry used to cite "the VRS-unverifiable decision" as its precedent; that decision was reversed in 0.5 for the half of it where no authored edit could clear the finding (an indel the compiler cannot recompute stays a warning in both modes), which is precisely the test this item passes and that one did not. An obsolete rsID is a cell a human can rewrite.

Two refinements, both now settled by the 0.5 implementation:

  • Merged ≠ withdrawn — and withdrawn is not observable. This entry used to say "probe the withdrawn shape before deciding", on the assumption that a withdrawn rsID deserved failing in both modes. The probe was done, and it dissolved the question rather than answering it: rs11273140, a genuinely withdrawn id, returns a response byte-identical to rs2000000000, which was never assigned — the same error string from esummary, the same count=0 from esearch, the same Ensembl 400. Routes checked and rejected for separating them: esearch has no withdrawn filter (the phrase is not indexed), and latest_release/misc/rs_unsupported_b157.txt looks like a withdrawn registry but is a one-off build-157 ClinVar-parsing incident list that does not contain rs11273140. Separating them would need a historical dbSNP dump, not the live API. So the vocabulary is live|merged|absent, not live|merged|withdrawn, and absent's message names both readings and asserts neither — because guessing "typo" sends an author to fix the wrong thing when the truth is that the variant itself was retracted. Severity is the same ladder for both (warn / fail in strict), not escalated beyond it, since absent has benign causes too (a very new rsID, or API lag).
  • The new columns are provenance, not facts. rsid_current + rsid_status sit outside RESOLUTION_FACT_FIELDS, beside rsid_alternates. They describe time-varying external state; inside the fact set they would make resolution_signature change when dbSNP merges something, with no change to the module — the signature would stop being reproducible from the module's own content. (Shipped as specified, with a test that pins it.)

One consequence worth recording, since it was previously filed as a loose end: reverse_module does not carry these columns, and that is correct rather than a gap. Reverse rebuilds resolution.csv from weights.parquet, which by design holds no provenance — it already resets source to reversed, status to resolved and blanks fetched_at. rsid_alternates/rsid_current/rsid_status are out of the fact set precisely so they never reach the artifact, so the information does not exist for reverse to emit; adding the column names would produce a permanently empty header. Recovering them after a round-trip means re-running the enricher, which is where a statement about a reference at a moment belongs. What reverse does now carry back correctly is the resolved facts and the authored shape — see COMPILER.md § Resolution for the enumerated round-trip contract that replaced the old "reverse emits position-only" rule. Addendum 2026-11-05 (RM380): the decision stands and the rationale moved. The parquets now carry every ResolutionRow column but a reasoned few on their resolution loci, status and authority among them, so "holds no provenance" is no longer true; these three stay out as statements about dbSNP at a moment (parquets.RESOLUTION_LOCUS_NOT_WRITTEN).

The strategic reading: this whole class of problem is label drift, and it exists only for rsid-keyed rows. A coordinate-authored row keys on a VRS allele id, which is content-addressed and cannot drift. The obsolescence check is therefore the standing cost of the rsID key, and the format already offers the escape — author coordinates and carry the rsID as data (reverse already emits coord-keyed rows as position-only). A strict failure is the nudge toward the drift-proof key.

Freeform suggestions — the 0.6 idea-book

  • ~~IUPAC ambiguity codes in ref/alts — expand Y to C,T~~ — probed and rejected as specified; one small real defect survives it. (the code table: R=A/G, Y=C/T, S=G/C, W=A/T, K=G/T, M=A/C, B/D/H/V for the three-base sets, N=any, ./- a gap.) Recorded in full because the reasons it fails are reusable, and because the first draft of this entry asserted a premise nobody had checked.

The load-bearing premise has no instantiation. The proposal rested on a code in an ALT column being a compressed ALT set — Y written once instead of C,T. Probed: zero occurrences of R/Y/S/W/K/M/B/D/H/V in REF or ALT across all 4,439,382 ClinVar GRCh38 records, and zero across all sixteen modules in this tree. Genuine ambiguity codes live in sequence contexts (consensus FASTA, array-manifest probes) and in genotype contexts (a Sanger heterozygote written Y) — the second of which is a measurement, so it is the consumer's by charter, and AuthoredModel's genotype validator already refuses it with a clear message. Not one of them is a variant record's ALT. This is the "mechanically possible, never instantiated" anti-finding the dogfooding rule exists to catch, and the first draft of this entry walked straight into it.

The non-ACGT alleles that are real are N, and they are two different things, neither expandable. 35 ClinVar records carry a single-base A>N — the substituted base is unknown, so expanding to A,C,G,T would assert four alleles ClinVar never stated. 633 more carry N inside a longer allele (TAAAAAT…TTTGG + NNNNNNNNNN + AAAA…) — unknown interior of a known-length insertion, not an ambiguity code at all, and 4¹⁰ expansions of nonsense. A rule keyed on "every character is a nucleotide or an IUPAC code" files the second as an ambiguity code, which is precisely the false claim cpic.unusable_allele_reason was already repaired to stop making about DELTCT.

And it is already solved, in the right place. clinvar_build filters ^[ACGT]+$ on both alleles at the snapshot builder and counts what it skipped, so none of those 668 records ever reaches a drafted module. Skip-at-the-source-boundary is the pattern; it is implemented; for the only non-ACGT ALT that exists in real data it is the correct answer.

Both halves of the proposed repair were also wrong on their own terms, and these are the parts to remember:

  • "Normalize at the enricher boundary, like ClinGen's dosage codes." The analogy does not hold. Those are decoded while reading a source into rows the enricher authors. An ambiguity code in variants.csv is authored data, where the enricher's standing rule is report, never repair — rewriting an authored value destroys the evidence of the upstream bug. The only legitimate site is a drafting provider at the moment of transcription, and every provider that meets one already refuses correctly.
  • "Have the compiler reject the code by name." Far larger than it sounds, and pointed the wrong way. No nucleotide grammar exists on any of the eleven ref/alt/alts columns across six models — vocab.validate_allele has two users, HaplotypeRow.allele and VariantRow.effect_allele (this said "exactly one" until 0.6; the count was wrong, the argument is not). Introducing one would reject <DEL> and N alongside Y, i.e. tighten the very field RM5 exists to widen. It is also Principle 3-illegal on the published line: a module with alts="Y" compiles today under best_effort (the locus is dropped with a warning), so a grammar would stop an existing module validating. The first draft claimed such a module "is already broken by it" — checked, and false.

What survived was small, and it shipped. hosting_verdict("C/T", "T", "Y") returns a confident False: a substitution locus has no spelling freedom, so a non-nucleotide alt reads as a positive contradiction. The author was told their genotype does not fit their own locus — true of the cell, false of the variant, and three steps from the actual mistake. A diagnosis defect rather than a grammar one, so fixing it needed no decision about what Y means: alleles.non_nucleotide_reason/non_nucleotide_alleles (format tier, the single definition cpic.unusable_allele_reason now delegates to) classify the offending allele, and both "cannot host" call sites say which of the two it is. Additive, digest-neutral, tightens nothing, orthogonal to RM5. The verdict itself is untouched — False was never the wrong answer, only the wrong explanation.

  • What an artifact should carry of the 0.3 axes — the residue of the direction report, and a 1.0 question. The documentation half shipped in 0.5.2: COMPILER.md's coverage row now names the tier each tick belongs to, and § Upgrade derivation says outright that weights.parquet carries the authored direction only, that an empty one on a legacy module is correct, and that a parquet-side consumer applies derive.direction_from_state(state, weight) itself. What is not settled is whether the artifact should ever carry the derived axes at all — a design question, not a version one: what bars it is that filling a blank asserts what no curator wrote. The candidate repairs and why each is wrong today — populate at compile asserts an axis no curator wrote (state='significant' has no direction, so one gets invented from the weight sign), trim state to a derived mirror on load is upgraded(), which belongs to the publisher's needs_upgrade flow rather than to the compiler, and state stays required under P8 regardless. Note for whoever picks it up: there is no numbered RMn for the 0.3 orthogonal-axes work — it shipped in 0.3 and is tracked only in COMPILER.md, which is part of why this gap sat unattended. Reported as S5 in CONSUMER_SUGGESTIONS.md.

  • Rename the compiler's resolution master switch. compile_module(resolve_with_ensembl=False) now warns when an injected resolution.csv is present and being ignored (0.5.2), which closes the silent-success half. The name is still wrong — it says Ensembl and means resolution of every kind — and a rename is a published-signature change, so resolve / resolution=off|table|cache is a 1.0 item rather than a patch.

New ideas enter here as freeform suggestions, then graduate through the design cycle (feedback → USE_CASES lens → PROPOSAL → shipped or parked as an RMn above).

  • Fixed by RM121 (2026-08-20): stats.genes now unions the gene column of every family a module carries. Kept for the triage record. manifest.stats.genes was derived from variants.csv alone, so a table-only module published gene_count: 0. Relayed from just-dna-lite (2026-08-21), originally measured by just-module-creator; filed here because neither of us owns variant_stats and we could not tell whether it had already been reported. compiler.py computes genes = sorted({v.gene for v in variants if v.gene}), so a module whose gene is stated only in a PGx or binning table — pharm_variants, haplotypes, diplotypes, allele_function, copynumbers, repeat_alleles — publishes gene_count: 0, genes: [] despite naming its gene on every row. The registry indexes version_genes straight off that field, so such a module is unfindable by gene in the catalog. Reported measured on a CYP2D6 activity_phenotype module, the corpus's own CYP2C19 example, and the shipped HTT manifest.

Two candidate homes and we have no view on which is right: widen variant_stats to union the gene column of whichever families the module carries, or leave the manifest alone and index the PGx tables' gene directly registry-side. The first makes stats.genes mean "genes this module is about" rather than "genes in variants.csv", which is a semantic change to a published field even though it is additive in type — worth a Principle 3 read before it is treated as a minor.

Consumer-side context: the symptom surfaces in just-dna-lite's discovery path, but nothing there can fix it — we read the field, we do not compute it. Full triage in that repo at docs/reviews/consumer-handoff-triage.md.

  • An authored report: block — let a module say how its rows should be presented. Proposal from just-dna-lite (2026-08-21), from a survey of everything our report decides for a module. Full write-up with wiring points and the constraint list at /data/sources/just-dna-lite/docs/MODULE_REPORT_CONFIG.md; this is the format-side half.

The bulk of what we found needs no format change at all — a module's Display block already carries title/description/report_title/icon/icon_set/color, and we never read it at report time: a locally compiled or registry-installed module gets it copied once into our own config at registration, and a module discovered remotely gets nothing at all, even though we fetch and validate its whole manifest to decide its kind. That is ours to fix and is planned on our side; noted here only because it is the reason the corpus looks like it has no display metadata. (One thing worth knowing on your side: icon_set is dropped on every path we have, so the vocabulary you validate it against has had no consumer here.) What is left over is a small set of decisions the author is better placed to make than any consumer, and which have nowhere to live:

  • categories: — key → {title, description, order}. Our report carries five hardcoded longevity pathway headings with hand-written prose, reachable only through a surviving elif mod_name == "longevitymap" name gate. The prose is a property of the module. Given a place to put it, the special case becomes a data check, and a second module wanting pathway sections needs no consumer code.
  • sort_by: — closed vocabulary over columns that already exist (weight_abs, evidence_level, effect_size, clin_sig, priority, p_value, gene, authored). We currently guess the ranking axis from the lead table: abs(weight) for weights-led, ClinPGx evidence rank for pharm_variants. That guess is right twice by luck and has no answer for the binning families.
  • preview_rows: / a row budget — how many rows lead before the fold. Ours is one global constant applied identically to a 20-row curated module and a 53k-locus ClinVar panel.
  • weight_display: + a declared weight range — strictly presentational. We are aware this borders on Weighting, whose docstring explicitly refuses a typed precedence field, and we are not asking to reopen that: nothing here blends, combines or reinterprets weights. The concrete defect is that our colour ramp hardcodes min(|w|*200, 200), an implicit 0–1 assumption, while RM92 correctly says the scale is free text — so a log(OR) module is coloured wrong and we cannot tell. A machine-readable range, or simply do not render a weight column, is the narrow fix.

Shape we would expect, and which we think keeps it cheap: advisory, same class as panel/authorship/license/weighting — out of artifact.digest, out of content_signature, not reconstructed by reverse_module, additive-optional and therefore a minor under Principle 3 as amended. Absent is tri-state and must mean the module has not said, never a default; every module in the corpus predates it, so absent is the normal path for the whole 0.x tail.

One thing we would ask the format to say out loud if this lands, because it is a contract point and not a consumer preference: a report: block governs the presentation of the module's own content and nothing else. It must not be able to suppress a consumer's integrity apparatus — an inferred/hom-ref-restored badge, a locus_count > 1 ambiguity caveat, licence attribution, module provenance, or the record that a module was skipped. Each of those exists in our report because its absence previously read as a positive claim, and a knob that can turn one off is a knob that lets a module launder a caveat. Stating the boundary in the field docs is what stops a future consumer from implementing it the permissive way.

No view from us on whether this is one block or several, or whether categories: belongs beside module: rather than inside a report: block — the placement is yours.


  • An absence that is a decision has no way to say so — weight, weighting:, license, authorship, and every sparse column. From just-dna-lite's 2026-08-21 dogfooding pass (D5, D6, D12/D26), dispositioned below. One module authors no weight on 190 rows and declares no weighting:; six declare version: null, license: null, authorship: []; direction, stat_significance and trait_efo_id are empty across the whole corpus. In every case "deliberately none" and "forgot" are the same bytes — which is the house tri-state rule aimed at authoring completeness rather than at data, and the one place we do not apply it.

Two halves that should not be answered by accident, because one is cheap and one is a schema decision. A per-column fill count in stats is one pass over rows the validator has already loaded, turns what is there to curate from a question into a table, and commits to nothing. And a way to declare an absence intentional — which is a new authored surface, full cost under P9, and needs a shape before it needs a field: weighting: already exists to say a module has no weights, so the question may be narrower than it looks (does the presence of the block, rather than a new field, discharge it?). The compile gate's precedent is that licensing is data rather than a flag, and declared_use's is that a third state beats a mode.

  • A resolved ALT set that is a tandem ladder is a row in the wrong family. From the same pass (D13). superhuman authors rs56185968 as one variants.csv row with no coordinates and genotype G/G; resolution expanded it to a 23-allele ladder (G, GTGTG, … up to 45 bases), which is where 22 of that module's VRS warnings came from. repeat_alleles.csv exists for exactly this and the authoring skill states the rule, but nothing notices. The signal is already computed, which is what makes it attractive.

What stops it being an item today: looks like a ladder is a heuristic, and a false positive tells an author to move a row that belongs where it is. Wants a rule sharp enough to be an info at worst — and note it would have fired on a module whose other 22 warnings were the S67 flood, so the two interact.

  • Nothing checks a conclusion against the other cells on its own row. Numbered as RM279 on 2026-09-27; the genotype rule shipped as a hint on 2026-09-28, with RM308 and RM309 left. From the same pass (D14), and the reporter measured it before proposing it: 20 hits in 1,418 rows, ≈60% precision by hand inspection, which is why they asked for warning rather than error. Two of the finds are worth quoting because they are exactly what a curation check is for — coronary rs17514846's C/C and A/A conclusions are swapped, and rs11591147 is scored protective +1.2 on T/T under text saying GG is protective. Two candidate rules, neither needing an external source: a conclusion naming a genotype built from alleles at this row's own rsID locus that is not this row's, and a state of risk/protective that the conclusion negates.

The locus restriction is the idea that makes it tractable — an earlier version of theirs flagged "raised plasma triglyceride (TG) levels". What has to be designed is the false-positive rate: a 40% miss on a warning an author reads on every compile is how a channel stops being read, which is S68's problem arriving from the other direction.

Freeform suggestions — the 0.7 idea-book

  • A resolver-ladder rung that consults a configured peer (S94, just-dna-registry, 2026-09-11 — a suggestion, not a request, in their words, and filed here rather than as an RMn for that reason). The ladder every snapshot reads is explicit argument → $JUST_DNA_<LANE>_CACHE → shared base → live → None; the idea is a rung between the local snapshot and the live source that asks a configured registry, so a thin client with no caches has every enricher command work unchanged instead of each consumer coding against an HTTP surface separately — the shape just-dna-lite's Source discovery already has for modules. The consumer has built the serving half (S91–S93 are the field notes from it) and named the asymmetry rather than asking for the client half.

The one question worth keeping, and it is the gate, not a detail. A peer serving answers drawn from a licence-gated snapshot is a different act from a client downloading that snapshot under its own declared_use: check_declared_use gates a fetch, and a read of an operator-built snapshot is not a fetch (@acquisition-gate-is-not-a-read-gate) — but a rung that fetches answers from somebody else's snapshot is neither of those two, and the licence table has no row shape for it. ClinPGx and PharmVar are CC BY-SA + no-sale with a personal key; whether an answer served over HTTP is a redistribution is outside what RM27 designed: RM27 shipped redistribution as record-only, enforced by the registry at publish, and a served answer is neither a publish nor a fetch (@redistribution-ungated). So the design order is: settle what a served answer is under each gated source's terms, then the rung. Not before, because a rung that works for Ensembl and ClinVar and silently also works for PharmVar is the failure mode.

Two smaller things a design would have to say: a peer's answer is a fourth provenance beside snapshot / ensembl-rest / ensembl-live and has to be labelled as one in checked (S93 made the labels the payload's contract); and --offline has to mean no peer either, since the operator's word for "reach nothing" cannot quietly exclude the one hop that reaches something. Nobody has asked for this; when somebody does, start from the licence question.

  • A declared dependency graph over a module's derived sidecars (S112, just-module-creator, 2026-09-24 — "offered as an observation, not a design", and filed here rather than as an RMn for that reason). expression_effects.csv is computed from resolution.csv's coordinates and, since RM259, can also be computed from the MANE lane's candidate genes. Nothing records that, so a consumer re-deriving one sidecar cannot tell that another one built from it is now stale, and orders its passes from knowledge nobody wrote down.

The shape exists one layer out. A cache lane's parents are a field (@a-derived-lane-has-parents-and-an-absent-parent-is-not-an-empty-result), and a parent that moved is reported, never silently rebuilt. Spec-directory sidecars have no such field. Today's settled answer for downstream staleness is delete-to-regenerate (@sidecar-authoritative), which costs nothing since 0.7. So this is not a defect to re-file, and it is not a cascade either.

What a design would have to answer first. What does it mean for expression_effects.csv to depend on resolution.csv when it is keyed on variant_key, which is rsID-first and does not move when a coordinate does? A row can be stale in its distance column while its key is still right. So a dependency is on columns, not tables, which is what the reporter wrote. Second, a staleness signal needs a baseline, and a merge-not-clobber sidecar's rows were written on different runs. Per row, the question is which parent state each row was computed from, and no sidecar records that (@currency-cannot-be-a-column). Nobody has asked for the cascade itself; the observation is what was offered.

  • ~~A measure kind for base-modification fractions~~ — promoted 2026-09-30. The VCF 4.5 audit (§ 6) parked it for lack of demand. The maintainer then named demand (PacBio samples arriving), and METHYLATION_PACBIO measured what PacBio users actually hold. The per-base VCF 4.5 M5mC shape this bullet described has no PacBio producer. What exists became RM317 (TRGT FORMAT/AM, FMR1), RM318 (imprinting regions in BED/TSV) and RM319 (TRGT's MC count), all in ROADMAP_0_8.md.

Consumer note (just-dna-lite, 2026-08-21) — a dogfooding pass over ten modules, and the eleven findings that are yours rather than the plugin's

Nothing here is a request to change an artifact, and none of it is urgent. We ran a revision/curation pass over all ten modules in data/interim/v1_port/ driven entirely through the installed authoring surface (just-module-creator 0.18.0 against format / compiler / enricher 0.6.6), and wrote up 26 findings in just-dna-lite/docs/MODULE_DOGFOODING.md. Seven of them are the plugin's. The rest are compiler- or enricher-tier and are listed here so the plugin maintainers do not have to relay them. All ten modules validate clean with zero errors — every item below is about what a green run does and does not tell an author.

The corpus is a fair sample of the awkward cases: six hand-curated Gen-I ports (8–528 variants), three ClinVar-drafted panels (57k / 70k / 309k variants), and one pharm_variants-led PGx module.

The one we would rank first

_verify_vrs_ids emits one warning per allele, and _vrs_coverage aggregates — so the better-resolved module produces the flood. (compiler.py:2648 / :2680; § D2.)

module resolution rows with vrs_id indel rows indel rows with vrs_id warnings from validate_spec
superhuman 101 101 47 47 85, of which 80 are one VRS line per allele
cardio 57,595 30,785 26,810 0 7, of which 1 is the aggregated coverage line

An absent id is "nothing to check" and lands in one tidy coverage sentence; an id that is present and not offline-justifiable gets its own warning. Same underlying fact — an indel identity needs the reference sequence — reported as one line or as eighty depending on whether the enricher happened to mint something. A 190-row module is the worst case in our corpus, and it is the size a human is most likely to be working on interactively. _vrs_coverage's grouping already exists one function away.

Presentation of findings (compiler)

  • § D3 — warnings is list[str] with no code, no count and no cap. On superhuman the three warnings an author can act on (8 het / 31 hom-alt / 46 ref-hom genotypes with no row) are items 83, 84 and 85. A warnings_summary: {code: count} beside the list would be enough and breaks no caller.
  • § D4 — nothing marks a warning an author cannot clear. The VRS ones say it themselves ("minted upstream by the enricher, not recomputable here") and sit at the same level as a real curation gap. compiler.py:2634 already reasons about exactly this — "a finding no authored edit could clear is not a strict matter" — and applies it to severity but not to presentation. The blame discriminator is already computed.

Messages that mislead

  • § D19 — the panel: deprecation says "nothing else is lost", and for a drafted panel that is false. (compiler.py:3425.) cancer's block carries 425 genes, a significance filter and a reference_sha256; the replacement — the licence row's dataset — carries clinvar_2026-06-27, a release name. validate_spec reports gene_count: 298 for that module, so 127 panel genes yielded no variant and only the block being deleted distinguishes not in the panel from in the panel, nothing found. Worse, cardio's dataset cell is empty — it was drafted 2026-08-10, before the drafter filled it — so following the instruction deletes its only record of which snapshot it came from, and refresh_sidecar will not backfill a sidecar that is already present. A conditional warning, or carrying panel into manifest.json the way weighting is, would settle it. Update: S69 shipped the message half (the warning stopped claiming nothing else is lost); the home-for-the-three-fields half is RM265 (S114), which showed the replaced-branch warning is unclearable for a module that needs them.

Absences that pass in silence (compiler)

Each would be one extra clause on a warning that already fires, and the closure warning is the model — it fires correctly on all ten.

  • § D5 — superhuman has 190 rows with weight empty on every one and no weighting: block. Green strict compile, weights_rows: 190, no remark. "Authors none deliberately" and "forgot" are the same bytes, and weighting: is the field that exists to tell them apart. We sum weight per module, so it renders 0.0 across the board while showing 190 findings.
  • § D17 — eight of ten modules carry no verification.json at all and nothing says so, while the two that do get a closure warning. "A check that could not run is not a check that passed" has no counterpart for the check that was never run.
  • § D12 / D26 — the six curated modules declare version: null, license: null, authorship: [], and their sources.csv carries only an ensembl / resolution row — nothing at layer: annotation. All six render Not stated for version, digest and terms in our report's Modules in this report table, which is the table that exists to tie a report to the bytes behind it. The licence compile gate is scoped to PGx sources, so originally-curated content gets no signal in either direction — no nag, and no way to record deliberately unlicensed for now.

stats (compiler)

  • § D7 — for a pharm_variants-led module the scalar counters are 0, not null. pharmgkb (1,482 rows) reports variant_count: 0, unique_rsids: 0, study_count: 0, with the real number only in table_rows. unique_rsids: 0 is wrong rather than inapplicable — the table is keyed on rsid. This is the tri-state rule the format enforces everywhere else ("null never means zero"), inverted in its own output. A family-independent row_count would also help.
  • § D6 — stats reports fill for genes/clinvar/pathogenic/benign and nothing else. Five of our six curated modules have category empty on every row and categories: [] is reported without comment; direction, stat_significance and trait_efo_id are empty corpus-wide. A per-column fill count is one pass over rows the validator has already loaded, and it turns what is there to curate from a question into a table.
  • § D9 — IFNL3;IFNL4 counts as a gene. pharmgkb reports gene_count: 33 including all three of IFNL3, IFNL4 and IFNL3;IFNL4 (33 rows carry the joined cell, straight from the ClinPGx export). Nothing flags a delimiter in a single-valued cell.

verification.json (enricher)

  • § D16 — the record counts findings and does not keep them. cancer: clinical_significance, 141,616 subjects, 20 findings, detail: null. pathogenic: 618,629 subjects, 32 findings, detail: null. Fifty-two rows in our two largest modules assert a clin_sig ClinVar disagrees with, and there is no way to learn which — no sidecar, and review_queue covers overrides rather than check findings. The MCP instruction block is emphatic that a mismatch means check both sides, and neither side is checkable for a finding you cannot name. A verification_findings.csv keyed on variant_key with both values would close it. Related: nothing rolls the count up into validate_spec, so a pass driven by the validator never learns the record exists.
  • § D20 — producer is top-level over a merged record set. check_identifiers on cancer correctly preserved the 0.6.4-produced clinical_significance record and appended three — and rewrote producer to just-dna-enricher 0.6.6, so the file now attributes to 0.6.6 a record it did not produce. checked_at survives, so it is recoverable. producer looks like it belongs beside source, release and checked_at on the record.

Two the linter could take, and one routing question

  • § D14 — nothing checks conclusion against anything, including the other cells on its own row. It is required, redundancy_bearing: null, and absent from attestation_bearing — all correct, and none of that implies it cannot be checked. lint_rows over twelve real thrombophilia rows returned errors: 0, warnings: 0 on a set containing rs1799963 A/A → "GA carriers have 6.74x risk" (the A/G row above it says "GA carriers have 2.8x"), rs2519093 C/T → "TT genotype is associated…", and rs1799889 G/G with state: risk under text reading "…is not increased". Two rules, no external source:
  • a conclusion naming a genotype built from alleles at this rsID's own locus that is not this row's — measured 20 hits in 1,418 rows across the six curated modules, precision ≈ 60% by hand inspection, so warning not error. (The locus restriction is what makes it usable: an earlier version flagged "raised plasma triglyceride (TG) levels".) The two worst are coronary rs17514846, whose C/C and A/A conclusions are swapped, and coronary rs11591147, scored protective +1.2 on T/T under text saying GG is protective.
  • state is risk/protective and the conclusion negates it.
  • A third shape we are not proposing as a rule but would like somewhere to record a decision about: 480 longevitymap rows share a conclusion across genotypes carrying different weights. Plausibly correct for a GWAS port — the association is one statement and only the dose differs — but a reader sees identical prose and different numbers, and nothing in the module says whether that was intended.
  • § D13 — nothing notices a row in the wrong family. superhuman authors rs56185968 as one variants.csv row with no coordinates and genotype G/G; resolution expanded it to a 23-allele GT ladder (G, GTGTG, … up to 45 bases), which is where 22 of that module's VRS warnings come from. module-tables states the rule (a repeat count is a binning table, not a variant row) and repeat_alleles.csv exists for it. The signal — a resolved locus whose ALT set is a tandem ladder — is already computed.

Two small ones

  • § D11 — just_dna_compiler/__init__.py exports nothing. dir(just_dna_compiler) is [], so from just_dna_compiler import validate_spec raises ImportError; the working import is just_dna_compiler.compiler. Our own CLAUDE.md documented the top-level form, so the drift was on both sides — corrected on ours. Flagging only in case the empty __init__ is unintended.
  • § D10 — logo.png travels through a compile and is not on the registry's roster. just_dna_registry.specfiles.RECOGNIZED_SPEC_FILES has 24 names and no image and no log; compile_module copies both into the output (verified — a compile of superhuman emitted logo.png and v1_port.log beside the five parquets). All ten of our modules ship a logo.png and we consume it in the module cards and the report. Whichever way it should go, it currently survives a compile and would not survive a server-side rebuild. No view from us on which is right — we would just like it decided and written down.

Everything above is a note, not a request, and we are not tracking any of it. Where an entry is already covered by an item in PROPOSAL_0_6_PT2 or the minor-deferral file that we have not read closely enough, the existing item wins.

Disposition of the note above — every finding, 2026-08-24

The note arrived in this file rather than in the inbox, so the triage ledger could not see it: an inbox the ledger cannot see is a backlog nobody sees, which is the whole reason CONSUMER_SUGGESTIONS.md is the one door. Their prose is left byte-for-byte above and this is the answer beside it. Eight of the sixteen findings arrived a second time as Sn items from just-module-creator and were answered there; the other eight had no ledger entry of any kind and are dispositioned here, because the reporter says they are not tracking them and an undispositioned note is one nobody reads again.

Answered as Sn, in CONSUMER_SUGGESTIONS_HISTORY.md: D2 → S67 (shipped, the VRS flood is grouped by reason). D3 + D4 → S68 (filed as RM131; the blame discriminator they name is the item). D19 → S69 (the message half shipped — the warning is per-branch and stopped claiming nothing else is lost; the home-for-the-three-fields half is RM265, from S114). D16 → S70 (shipped: detail on the clin-sig record and a findings warning; the sidecar is RM130). D20 → S71 (shipped, producer per record). D7 + D9 → S72 (row_count and the delimiter warning shipped; the 0→None retype is queued for 1.0).

D17 — "eight of ten carry no verification.json and nothing says so" — does not reproduce, and this is the one worth reading. A module with no attestation at all does warn: _closure_warning is keyed on the outcome, not on the file, and its docstring says so — one sentence covers all three ways a compile publishes no closure, because what an author needs to know is the same in each. Verified by deleting verification.json from apoe_epsilon and re-validating: the closure warning fires. What is genuinely absent is a statement that no checks were run, which is a different claim from authoring was never declared finished — and it is deliberately not made, because warning about an unverified module would fire on nearly every module in this repository.

D11 — just_dna_compiler/__init__.py exports nothing — reproduces exactly (dir() is [], from just_dna_compiler import validate_spec raises) and is intended. CLAUDE.md's own rule is avoid __all__ / pure re-export __init__.pys — they obscure where a symbol lives, so just_dna_compiler.compiler is the supported import path and the empty __init__ is the policy rather than an oversight. The reporter corrected their side already. Related and now fixed: S74 shipped load_spec, because the symbol they actually could not reach was a private one.

D5, D6, D12/D26 are one finding seen three ways, and it is the strongest thing in the note. A module that authors no weight on 190 rows and declares no weighting:; a stats that reports fill for four columns and nothing else while direction, stat_significance and trait_efo_id are empty corpus-wide; six modules declaring version: null, license: null, authorship: []. In each, "deliberately none" and "forgot" are the same bytes, which is the tri-state rule pointed at authoring completeness rather than at data. Filed below as an idea-book entry rather than an RMn because the shape is undecided — a per-column fill count in stats is cheap and mechanical, while declaring that an absence is intentional is a schema question, and the two should not be answered by accident.

D13 — a repeat ladder authored as a variants.csv row — is real and is the sharpest of the remainder. superhuman's rs56185968 resolved to a 23-allele tandem ladder, which is where 22 of that module's VRS warnings came from, and repeat_alleles.csv exists for exactly that. They are right that the signal is already computed. Idea-book below; it is a check with a clear input, and what stops it being an immediate RMn is that a resolved ALT set that looks like a ladder is a heuristic, and a false positive tells an author to move a row that belongs where it is.

D14 — nothing checks conclusion against the other cells on its own row — is real, is measured, and the measurement is why it is not filed as a check yet. Their own number is 20 hits in 1,418 rows at ≈60% precision, and they proposed warning rather than error for that reason. Two of their finds are serious enough to quote: coronary rs17514846's C/C and A/A conclusions are swapped, and rs11591147 is scored protective +1.2 on T/T under text saying GG is protective. A 40% false-positive rate on a warning an author must read every compile is the thing to design against, and their locus-restriction is the idea that makes it tractable. Idea-book below. Their third shape — 480 longevitymap rows sharing a conclusion across genotypes with different weights — they explicitly did not propose as a rule and we are not treating as one; it is the "nothing says whether that was intended" problem again, which is D5's shape.

D10 — logo.png and the log survive a compile and are not on the registry's roster — is a real cross-repo inconsistency and is not ours to settle alone. Verified on our side: compile_module copies both into the output deliberately (logo_file/log_files are parameters, manifest.logo and manifest.logs[] are published entries). RECOGNIZED_SPEC_FILES is the registry's list. Either the roster grows or the copy stops, and the reporter is right that the current state means a module survives a compile and would not survive a server-side rebuild. Raised with the registry rather than decided here; recorded so it is not lost.