Skip to content

0.8 part 2: the build plan for the agreed scope

Status: concluded 2026-11-05, at the 0.8.0 cut. Every item it planned landed on the 0.8 branch; each record is in ROADMAP_HISTORY with its Residuals, and what 0.8 shipped unfixed is in INTEGRATION_0_8 § 4. It no longer overrides the roadmap files.

Written by the lead on 2026-11-04 from the pm's handout ("0.8 scope agreed: yours to plan and build") and its two addenda (the tag selector, then column=value selectors). This is RELEASE_CYCLE § Minors step 2: the plan the build runs from without asking. PT1 (the bugfix round) is concluded on the 0.8 branch. A fresh critic read it the same day against the decision lines and the code; its findings that held are applied, and the ones that need the maintainer are §4 questions 5 to 7. The pm's third addendum (the shape agreed after the second measuring run) is folded in.

How to read it. The decisions are the maintainer's and live as dated lines under each item in ROADMAP_0_8.md; where an item has several, the latest wins. This file does not restate them. It fixes what the decisions left open (§3), names what goes back to the pm (§4), and gives each item the contract a builder works from (§5). A builder that finds a decided column with no case, or a needed one missing, stops and sends it to the pm; it never builds around it.

1. Branching

  1. Push 0.8. The local branch is 150 commits ahead of origin/0.8 (last pushed 2026-09-25) and 0 behind; PT1's build (2026-09-30) is among the unpushed commits. The lead's grant covers a fast-forward push.
  2. Merge main into 0.8 (main is 80 commits ahead of the fork point, 0.8 is 43). A trial merge on 2026-11-04 conflicted in 13 files. Resolution rules:
  3. compiler/src/just_dna_compiler/compiler.py and enricher/src/just_dna_enricher/cli.py were split into packages on 0.8 (RM260) and edited on main since (93 lines added, 22 removed). Port main's hunks into the package modules that now hold those functions, then take 0.8's deletion.
  4. docs/ROADMAP_0_8.md: main's text wins wholesale (the scope table and every 2026-11-04 decision line). 0.8's "built" markers for PT1's items are re-applied on top.
  5. docs/ROADMAP.md, docs/RM_TOC_*.md, docs/CHANGELOG.md: the union. An item 0.8 closed (PT1) stays closed; an item main filed or decided keeps its newest line.
  6. .claude/hooks/*: main's versions (the charter gate is newer there).
  7. Code conflicts (orchestration.py, test_declared_use_recorded.py) are resolved by reading both sides, and the full suite runs before the merge commit.
  8. The polars 2 lock lands in wave 0, right after the merge (A1's second half, moved forward), so no lane builds on one polars and merges onto another. Its RELEASE_RECORDS line goes with it.
  9. Every item is built in its own worktree under the scratchpad, on a build/RMnnn branch cut from 0.8, and merged into 0.8 by the lead once its tests pass. main is merged into 0.8 again at each wave boundary (§2), so a patch triage lands on main reaches the builders within a wave.
  10. Nothing in this scope lands on main. RM338 is the exception that proves it: its patch half is main's (a patch, built by the lead there), and RM337's warning builds on 0.8 after the sync that carries RM338 across.

2. Order, and what runs in parallel

Four lanes, one builder each, serial inside a lane. A builder is a subagent (Opus) with this file and its lane's §5 block as its brief. The lead reviews and merges.

Wave Lane A: small schema adds Lane B: the meta-table Lane C: indel placement Lane D: derived inputs
0 the push, the sync and the polars 2 lock (§1), the lead alone; RM338's patch on main, the lead
1 RM342 (with P4's amendment), RM336, RM322, RM334 waits on §4 question 1 (built under its default if no answer) RM304 RM303 (with the Goals amendment)
2 the StudyRow batch: RM325, RM326, RM328, RM329, RM330's stratum, RM331's author half, RM332, and the locus-level derived columns (§3 item 10) RM28 / RM323 / RM324 / RM327 / RM330's sex, one item, without the study-orphan widening RM292 RM333's Open Targets lane, carrying RM239, RM236 and RM238's coloc (§4 question 2)
3 RM335, RM337 (after the sync that carries RM338), RM253 the derived columns (tradeoff, the resolved facts), then the study-orphan widening once the StudyRow batch has merged (both edit _cross_validate_studies) RM305, after RM303 RM331's computed LD pair
4 RM267, then RM270
5 reference examples (§6), INTEGRATION_0_8, the review wave and the dogfooding wave

Why this order and not the scope table's listing.

  • RM303 first in lane D: RM305's and RM333's lanes write release.json, and RM305's own text asks for RM303's model if it has landed. Building a lane before the model means writing a hand-built dict the model's guard then refuses.
  • Lane C runs RM304 → RM292 → RM305 → RM267 → RM270, not the scope table's order. RM292 makes every check name its witness, RM305 is the witness those checks name, RM267's comparison needs both, and RM270 carries the vrs_id RM304 versions.
  • The StudyRow batch is one builder, because seven items edit one model, one parquet and one duplicate check. In parallel they would conflict on every file.
  • The outcome strip and eur_only_evidence are lane A's, built with the StudyRow batch, because they derive from phenotype_class and ancestry and nothing in the meta-table.
  • The study-orphan widening waits for the StudyRow batch: both edit _cross_validate_studies.

Rough sizes, for the person's planning, not as a promise: wave 0 about 2 hours; lane A about a day; lane B one to two days; lane C two days; lane D two days, most of it the Open Targets lane. With four builders in parallel, waves 1–4 take about three working days, then the review and dogfooding waves.

3. Decided here (reversible until the cut, listed for review)

Each is the lead's call under LEAD.md step 3: a reversible choice with a recommendation. The rejected alternative is beside it. Names still get the P5 audit in §5, and the audit can overturn any of them.

The meta-table.

  1. File name meta.csv, parquet meta.parquet. The maintainer's and every document's working name. Rejected: statements.csv (accurate but nobody's word for it), combinations.csv (wrong for a one-member facts row), meta_conclusions.csv (a facts-only row concludes nothing).
  2. One grammar for every member: table[column=value, column=value]. A concrete row is the selector over its table's full key (variants[variant_key=rs429358, genotype=C/C]), a locus is variants[variant_key=rs10455872], a named allele is allele_function[gene=APOE, allele=ε4], a tag set is variants[flags=legacy_longevitymap]. The precedence the maintainer decided then falls out of the conditions: the full key is a concrete row, variant_key alone is a locus, otherwise more conditions outrank fewer. Rejected: a bare rs10455872 shorthand beside it (a second grammar; additive later if authors ask), and the typed prefixes of the audit's appendix (allele:APOE=ε4, dip:…), which were one grammar per member kind. A named allele is an allele_function.csv item (the pm's setting in the third addendum, which overrode this plan's first choice of haplotypes.csv). The cost accepted with it: a named allele the module uses as a member needs an allele_function.csv row, every column but the key empty, which is also where RM334's HLA alleles live.
  3. Members are joined by & at bracket depth 0, in one members cell. The parser is bracket-aware, so an & inside a value (RM253's DPYD names, c.1129-5923C>G&c.1236G>A(HapB3)) is data, and so is | in a phased genotype. The shared MULTI_SEP splitter is never used on this cell. A value may not contain [, ], , or =; such a selector is refused with a message naming the character. Rejected: member_a / member_b columns (no separator at all, but every corpus case is arity 2 only until it is not, and a third member would be a third column).
  4. The table key is the normalized member set plus trait_efo_id plus sex (plus drug, 2026-11-04). Members are sorted for the key, so A & B and B & A are one key, refused as a duplicate. Within a member, conditions are sorted too. sex is in the key because it scopes the whole row, facts included: a male row and a female row for one member set are two statements, not a duplicate. A row with sex outranks the same key without it, for that sex. Addendum 2026-11-04: drug joins the key (the member set plus trait_efo_id plus sex plus drug), at the maintainer's default, after cpic_gene_pairs found CPIC answering one phenotype pair once per thiopurine; a drug-scoped row outranks the same key without one, for that drug.
  5. Columns, all optional except members: trait_efo_id, trait_role (closed: mediator; empty means the trait is an effect, the maintainer's third addendum), direction (the existing vocabulary), conclusion, negatives, tier, age_of_action, age_of_action_confidence, flags, pmid, doi, sex (closed: male, female). A row stating none of the finding or fact columns is refused as empty. trait_role without trait_efo_id is refused. pmid and doi follow StudyRow's rules, the NONE sentinel included (decided in the third addendum for every citing table, A2).
  6. A per-trait row needs an allele-bearing member (the audit's gap 2): every member of a row with direction must fix an allele, meaning a variants member with a genotype condition, a diplotypes or pharm_variants member that names its alleles, or an allele_function member. A bare variant_key member also fixes an allele when every variants.csv row of that locus states the same effect_allele, and direction is then about that allele. This is the audit's second run (§6): v2 is uniform on 228 of 228 loci, and the rule halves the per-trait rows (about 134 against 268 with genotype keys), the reduction rule's direction. A locus whose rows disagree on effect_allele or state none draws the warning meta_member_allele_ambiguous and its direction is withheld (the maintainer: a warning, not a refusal). A tag selector on such a row is refused, naming it. The four loci v2 marks "allele not determined" honestly get no per-trait row. The member tables are variants, diplotypes, pharm_variants and allele_function (the pm's setting); studies is excluded by the maintainer's bound.
  7. tier is open text, and module_spec.yaml gains tier_scale (one free-text name, the weighting: pattern), copied into the manifest and listed in REVERSE_RECONSTRUCTED so reverse writes it back (else the recompile refuses every tier cell, a P7 break). A tier cell in a module with no tier_scale is refused (@weight-has-no-unit). Values are not checked against the scale: the scale is the module's, and v2's X, A-, B- are legal.
  8. The compile writes two parquets. meta.parquet: one row per authored row, the selector text kept verbatim for reverse, the match count, and the expanded matches as a list column of (table, row key). A matched combination replaces the findings of exactly the rows in that list, which is how a consumer knows what it replaces without evaluating anything (P1). meta_resolved.parquet: one row per single-statement row that any meta row reaches, carrying each fact column after precedence, tradeoff (item 9), and a withheld list naming the columns two incomparable, disagreeing matches withheld. Combination matching against a genome stays the consumer's join, as every table's is. Rejected: a third parquet for the expanded members (it folds into meta.parquet's list column), and resolved facts on weights.parquet and diplotypes.parquet (two published parquets widened for one concern). Specificity across member kinds (the audit's gap 1, decided in the third addendum: "the narrower condition wins across tables"): a diplotypes member is more specific than an allele_function member naming either of its two haplotypes in the same gene. This is an ordering of members, not a selector condition, so the maintainer's bar on cross-table selectors is not crossed. With it the ε2/ε4 diplotype row settles the ε2 and ε4 rows' disagreement, and without it that column is withheld.
  9. tradeoff is a column of meta_resolved.parquet, Kleene as RM327 says: true when an allele-bearing member has both a protective and a risk per-trait row, false only when every trait's judgement is stated and none opposes another, otherwise null. The name moves from RESERVED_NAMES_0_4 to MISPLACED_COLUMN_REASONS in the same commit.
  10. The two locus-level derived outputs (the audit kept both): eur_only_evidence (RM328) and the outcome strip (RM326) become columns on weights.parquet, repeated per genotype row, since a parquet column is approximately free and adds no file. Both Kleene: eur_only_evidence is true when every supporting study row states only EUR, false when one states another code, null when any states none; a strip cell (healthspan, lifespan, exceptional) is true when a study row of a mapped class exists and null (untested) otherwise, never false. The class-to-cell map is a format constant over the recommended classes, taken from v2's requirements §5.1 as written there and listed in the builder's report for the pm; MULTIVARIATE_AGEING and any class outside the twelve map to no cell, which v2's README states (the audit's gap 8).

The StudyRow batch.

  1. sample_size is the total analysed, cases included; n_cases is the case count. A row with n_cases greater than sample_size is refused (the two cannot both be true, and both columns are new, so no published row can hold the pair). sample_unit is closed: individuals, lives. A sample_size without sample_unit is refused (@weight-has-no-unit).
  2. ancestry is multi-valued over the 1000G superpopulations (AFR, AMR, EAS, EUR, SAS), closed. VALID_TRAINING_ANCESTRY (pgs.py) holds the same five plus multi, so the five become one constant and the PGS set is derived from it with multi added: one vocabulary, not two. A multi-ancestry study lists its codes rather than writing multi. population stays free text.
  3. phenotype_class is open, with v2's twelve as recommended members, spelled as RM326's list spells them (FAMILIAL, MR, BIOMARKER, not the long forms). A token outside them draws the new warning phenotype_class_unrecognized.
  4. age_min / age_max are numbers in years, either may be empty (an open band), and age_of is closed: participant, parent. No refusal ties them together.
  5. stratum / stratum_level are free text, both or neither. Rejected: a recommended vocabulary for stratum (a warning nobody asked for).
  6. The duplicate-citation narrowing (RM140) treats two rows for one (variant_key, pmid) as distinct when any of these is both stated and different: phenotype_class, the band (age_min, age_max, age_of), the stratum pair. Unstated is not different (@absent-is-not-different). Gap 12 decided these three and no more.
  7. effect_unit reaches weights.parquet from "the matching study row", and VariantRow has no field naming a study, so "matching" is undefined: sent to the pm (§4 question 5). Default meanwhile: the unit joins only when every study row of the variant that states a unit states the same one; otherwise the column is null, never a guess.

Elsewhere.

  1. pmid's sentinel is NONE, exactly. Other spellings an author might reach for (NA, n/a, -, none) are already refused today; the message gains a pointer to NONE. With NONE, doi is required and checked against spec.DOI_PATTERN, which exists; validate_pmid_cell has no DOI arm yet, and A2 adds it.
  2. RM342's toolchain block is manifest.compilation.toolchain: dict[str, str] | None, with the keys python and polars. A future library is a new key. Where its reason fires is §4 question 7.
  3. RM304's names: manifest.resolution.vrs_version (the pin) and a vrs_version column beside vrs_id in resolution.csv and in every parquet carrying an id we minted; frequencies.parquet's vrs_id is gnomAD's and gets none. The new reason vrs_version_differs belongs to the enricher's VRS verify pass (RM304's question 2), not to verify_manifest.
  4. RM303: the builder first lists what each reader of release.json reads (the item's own method), then types every key a reader reads, plus parents; the rest ride in extra="allow". The drifted pairs settle on row_count and license (the spelling module_spec.yaml already uses), and the old spellings are read on input, never written. builder against builder_version: read what acmg's builder means by builder first; default, builder_version is the field and builder is read as an alias if it holds a version.
  5. RM334's grammar applies where gene starts with HLA-: *, then colon-separated numeric fields, then an optional expression suffix (N, L, S, C, A, Q), IPD-IMGT/HLA's published grammar. "At a stated resolution" is read as: the name states its own resolution by its field count, and the check refuses a name with fewer than two fields; no column states it. A malformed name is a refusal, gated on the same published-module scan RM336's refusal owes (§5, A2).
  6. RM335's pmid and doi are optional and draw no warning, because 17 of v2's 77 rows have no PMID and the table is annotation only (R-ANIMAL-1); when a row cites, the pmid/doi pair's rules apply (A2). Rejected: a fourth warning code for an uncited row. Its concordance is closed (concordant, discordant, nonmonotone); an empty cell is unstated, which is what v2's 55 "unclear" rows are.

4. Sent to the pm, each with the default the build takes meanwhile

Answered by the pm, 2026-11-04: questions 2, 3, 4, 5 and 7 accepted as their defaults (RM238 moved to 0.9 in 3c18ea3); 6 answered in the third addendum. Question 1 answered by the person the same day: meta.csv, its grammar and its key as §3 items 1–4 write them, with no shorthand yet. Nothing in §4 is open.

  1. The meta-table's member grammar and name (§3 items 1–4). A one-way door at the cut. Default: as written. This is the block to read closely.
  2. One Open Targets lane for four inputs. The 26.09 release on EBI's FTP lists l2g_prediction, credible_set, colocalisation, variant and study side by side (listing read 2026-11-04), under the one CC0 marking RM333's probe checked. So RM333's lane can feed RM239's posterior (from credible_set, onto gwas_effects.csv), RM236's per-transcript consequence (from variant, as RM236's recommended derived sidecar), and the coloc behind effector confidence (from colocalisation) in place of RM238's per-tissue GTEx table, which its own entry parks on a consumer. One lane and one licence instead of three sources: the reduction rule's direction. Default: build it that way; RM238's GTEx table stays parked and moves to 0.9. Field-level contents (that variant carries transcript consequences) are unprobed and are the lane's first hand-run step.
  3. RM267's open half: may the enricher write the corrected spelling? Default: no. A disagreement between two authorities withholds the row and reports both placements; nothing is rewritten in 0.8 (@enrichment-is-validation).
  4. RM331's LD reference. Default: Ensembl REST's pairwise LD endpoint over the 1000 Genomes phase 3 superpopulations, which matches decision 12's codes and adds no new licence. Unprobed; its first step is a hand-run call.
  5. RM332's "matching study row" is undefined, since VariantRow names no study. Default: §3 item 17's unanimity rule (a unit joins only when every stating study row agrees).
  6. doi on the meta-table. Answered in the third addendum: yes, on every citing table (A2).
  7. Where RM342's toolchain reason fires. The decision says verify, but verify_manifest never recompiles: it re-hashes the files on disk against the manifest, so the running polars cannot change its answer (checked in integrity.py). The places that compare a fresh compile against a recorded digest are the publish collision check (upload.py, a re-publish whose digest differs) and the release sweeps (release_records.py). Default: the reason toolchain_differs fires there, when the digests differ and the two manifests' recorded toolchains differ; verify_manifest is unchanged.

5. Per-item contracts

Every block follows the shared rules in §7. "Touch" lists the places a builder must reach; it is a floor, not a ceiling, and the registry guards (_ALL_MODELS, the warning-code and feature-corpus equalities) will name any it misses.

Lane A: small schema adds

A1. RM342, the toolchain in the manifest. Add Compilation.toolchain (§3 item 19), filled at compile from sys.version_info and polars.__version__. The reason toolchain_differs fires where §4 question 7 puts it: where two manifests' digests differ and their recorded toolchains differ, it replaces the plain collision or mismatch; when either toolchain is unrecorded (an older manifest), the existing behaviour stands. Amend Principle 4's reproducibility sentence to "pinned via compiler_version, the recorded toolchain and the resolved reference", with the why in CONSTITUTION_AMENDMENTS_HISTORY.md, in the same commit. The polars 2 lock already landed in wave 0. Test first: two manifests for one module whose digests and recorded polars versions differ are reported as toolchain_differs, never as a collision.

A2. RM336, a DOI in pmid, and the NONE sentinel. Before building, scan the published modules (the registry catalogue through registry_search and reference_examples/) for a DOI in pmid; any hit goes to the pm, not into the build. Then: validate_pmid_cell refuses a cell whose only token is a DOI and names it (RM50's arm is the template); accepts NONE (§3 item 18), which makes doi required and validated; extract_pmids returns nothing for NONE and for a DOI, so the literature pass never looks up 10, 1101, and so on. The pair applies to every citing table (the third addendum): doi is added to pharm_variants.csv, the binning kinds, gwas_effects.csv (derived), meta.csv and model_organisms.csv, and the roster is read from _CITING_TABLE_KINDS, asserted by an equality over it, never listed by hand. Tests: the RM336 reproduction (10.1101/2025.03.01.640000 → refused), NONE with and without doi on every citing kind, and the literature pass's lookup list for a DOI cell. Lane A runs A2 before A5 and A6, which lean on it.

A3. RM322, DiplotypeRow gains negatives and category. The three touch points (model, compile, reverse) and the diplotype parquet. Both are free text, as on VariantRow (which binds no vocabulary to category). Round-trip test over a diplotype module carrying both.

A4. RM334, the HLA nomenclature grammar on allele_function.csv (§3 item 22). Gated on A2's published-module scan, extended to HLA rows. Tests: *15:01, *57:01:01, *02:01:01:02N pass; 15:01, *1501, *15:01X refuse with a message naming the expected shape.

A5. The StudyRow batch (wave 2, one builder, one commit per item): RM325 (sample_size, sample_unit, n_cases), RM326 (phenotype_class and its warning code), RM328 (ancestry), RM329 (age_min, age_max, age_of), RM330's study half (stratum, stratum_level), RM331's author half (proxy_rsid; _check_study_effect_alleles checks effect_allele against the proxy's alleles when it is stated, and the enricher resolves proxy rsIDs as resolution subjects so those alleles exist), RM332 (effect_unit, and its join onto weights.parquet, §3 item 17), and the locus-level derived columns (eur_only_evidence and the outcome strip, §3 item 10). Then the duplicate-citation narrowing (§3 item 16), with one test per narrowing column showing a pair that warned before and does not now, and one showing an unstated column still warns. Each column: the P5 audit line in the commit message, the three touch points, studies.parquet, and reverse.

A6. RM335, the model-organism table (model_organisms.csv, parquet model_organisms.parquet). Columns: gene (required, the human symbol), orthologue, species, model_type, sex, diet, lifespan_effect, concordance (§3 item 23), pmid, doi. Key: every column except lifespan_effect and concordance. A _TABLE_KINDS entry, a _TABLE_DUPE_KEYS entry, reference.py's model map, the authoring reference, TABLES.md's hand-written section. It never reaches an effect column or a tier.

A7. RM337, the warning for a file under logs/ that does not ship. After the sync that carries RM338's patch from main. New code log_file_not_shipped, one finding per file, naming it and the reason: not a recognized extension, or failed its content check (RM338's reasons). Both modes, never a refusal.

A8. RM253, the CPIC drafter's notation. Repeat alleles spelled as callers spell them (the RM253 decision line: *28 = C>CAT at GRCh38 2:233760233), DEL<bases> / INS<bases> spelled VCF-anchored with the anchor read from the reference the enricher already reads, the bare DEL still skipped, and DPYD names as CPIC writes them with spaces dropped and & for the comma. No row keyed on rs3064744 alone. Test against the provisioned CPIC snapshot: draft_gene("UGT1A1") defines *28, *36, *37, and the "used but not defined" warning no longer fires for them.

Lane B: the meta-table

B1. RM28 / RM323 / RM324 / RM327 / RM330's sex, one item. The table as §3 items 1–9 fix it, plus module_spec.tier_scale. The work, in order:

  1. The model (MetaRow, on AuthoredModel) with the selector parser, its refusals (an unknown table, an unknown column, a forbidden character, an empty row, trait_role without a trait, tier without a scale), and the bounds the maintainer set (no negation, OR, ranges, cross-table joins, or selectors over studies.csv; each refused with its own message, not the generic one).
  2. Expansion at compile: each member against its table's rows, list columns meaning "contains". A member matching nothing is the warning meta_selector_matches_nothing; the manifest records every selector's match count. Precedence and withholding (§3 item 8).
  3. The two parquets, tier_scale in the manifest and REVERSE_RECONSTRUCTED, reverse (selectors verbatim), and the round trip.
  4. The study-orphan widening (_cross_validate_studies), in wave 3, after the StudyRow batch has merged: a study variant is not an orphan when it is a variants.csv row, a defining variant in haplotypes.csv, or the variant_key of a meta member. APOE's 14 links are the test.
  5. The rsID-merge check (NCBI, enrichment) runs over meta members' variant keys too, so IL6R's rs8192284 under two keys is caught.

Tests first: the duplicate set (A & B against B & A); the ε2/ε4 nesting case (two incomparable allele rows disagree → withheld, and a diplotype row settles it); a combination replacing what it names; variants[flags=legacy_longevitymap] against a more specific locus row; CPIC's TPMT × NUDT15 written as one phenotype-pair row; a selector matching nothing. Each against the real compile path.

B2. The derived columns: tradeoff (§3 item 9) and the resolved facts' withheld list, with the Kleene truth table as a parametrized test.

Lane C: indel placement

C1. RM304, the VRS version. The pin and the per-row column (§3 item 20), the minting call recording its version, the enricher's VRS verify pass reporting vrs_version_differs. The pinned insertion test (rs72613567 → ga4gh:VA.Jml7SNku3QQBCVIj78BGiFvR21bNkos7) behind JUST_DNA_NETWORK_TESTS=1. The ga4gh.vrs<2.4 cap stays until vrs-python releases the rle_subunit_mode keyword; lifting it is the liaison's call, and when it is lifted our call passes the mode explicitly.

C2. RM292, every check names its witness. VerificationRecord gains the source a check reads and the source it checks against, and a verdict member self_consistent used when they are the same. The check registry carries both as fields, and an equality test over the roster refuses a check that states neither. A tolerance that makes a class undecided publishes the count of that class.

C3. RM305, the ClinVar-GKM lane. Addendum 2026-11-04, from the hand-run probe (probes/LANES_0_8.md §3): the set is on Cloudflare R2, not Google Cloud Storage, and one Parquet set is overwritten weekly, so the pin is (source_release, sha256) plus a drift check; it mints under VRS 2.1.1, so ids compare only at the same VRS minor. The lane was built to that. Through CACHE_SURFACE.md's checklist, every line with its test. Indels only (the ~270k ReferenceLengthExpression alleles and other indels). A pinned release recorded through RM303's model; the Google Cloud Storage download owes the client exception contract; a currency finding when the latest weekly delta is old. It is a witness RM292's records name, with no check of its own.

C4. RM267, two authorities compared. For each rsID two loaded authorities place, compare by minted vrs_id (no tolerance), else by parsimony_reduce over a reference window. A disagreement is a finding naming both placements, and the row is withheld (§4 question 3). The count of rsIDs with one witness is published beside it. A new optional resolution.csv column records the witness asked and its answer, tri-state. Tests: rs8176719 (two ids, withheld), rs72613567 (one id, written).

C5. RM270, the spelling-independent key reaches the artifact. The enricher left-normalizes the coordinates it fills; for an authored coordinate it reports the normalized spelling as a finding and never rewrites the cell. The per-ALT vrs_id (with C1's version) goes onto the variant parquets. The stamped verdict (left-aligned / not / unknown) is a resolution.csv column. Remove the SHOX pin from compiler/tests/test_consumer_join.py; the test then passes with an empty miss set, which is the proof.

Lane D: derived inputs

D1. RM303, the snapshot layout in just_dna_format.snapshot. As the item's split says, with §3 item 21's choices, the Goals amendment (rule in CONSTITUTION, why in CONSTITUTION_AMENDMENTS_HISTORY) in the same commit, and an AST walk over *_build.py refusing a hand-built release.json. ClinVar's per-lane label exception (clinvar_dataset_label) is removed once its builder writes dataset through the model.

D2. RM333's Open Targets lane, and what it carries (under §4 question 2's default). First, by hand: read one row of each of l2g_prediction, credible_set, colocalisation and variant from release 26.09, record their sizes and fields in docs/probes/, and only then build. The lane writes: - an effector-gene derived table at Open Targets' own grain, one row per (study locus, gene) with the study locus's lead variant_key, the L2G score and the release, and a sibling grain for coloc evidence as the probe finds it published. No row is collapsed to one per variant: several study loci per variant are kept and counted, never resolved to a winner (@multiplicity-is-a-finding). The probe's field list decides the exact columns, which go back to the pm before the build fixes them; - RM239's two columns on gwas_effects.csv, the posterior and the credible-set level, with the item's withhold for a study with no credible set (a nobody-asked third state, not a zero); - RM236's per-transcript consequence sidecar keyed (variant_key, transcript), so RM23 inherits the grain when it is built (it stays in 0.9). The SourceRow states Open Targets' terms with notice carrying its disclaimer verbatim (RM333's decision). The eight hand-curated causal/effector inputs go in overrides.csv as corrections; nothing authored names a C or E label.

D3. RM331's computed LD pair. After A5. For each study row with proxy_rsid, the enricher asks the LD reference (§4 question 4) for r² between proxy and lead in the study's one ancestry code (a mixed-ancestry study gets its row with r² withheld, the pm's setting; the audit's second run counts 3 such links), and writes a machine-written ld_pairs.csv keyed (proxy_rsid, variant_key, ancestry) with r², the reference and its release. A study with no ancestry gets no row and is counted (unknown is not zero). First step: one hand-run call, recorded.

6. Reference examples

The audit's thirteen candidates fold into six examples, each with a README naming what it exercises and what it broke while being built (@probe-becomes-example). Each is the real check for its lane: it compiles under --strict, reverses, and recompiles to the same digest.

Example Audit candidates Lane
apoe_klvs_meta 1 (facts-only row), 2 (allele facts, the ε2/ε4 nesting), 3 (KL-VS × ε4), and the legacy tag selector B
pleiotropy_tradeoff 4 (per-trait rows, tradeoff, the no-allele loci) B
cpic_gene_pairs none (the second addendum's case): TPMT × NUDT15 as phenotype-pair rows B
longevity_studies 5, 6, 7, 8, 9, 11, 13 (the StudyRow columns, the narrowing, the sentinel) A
hla_alleles 10 (HLA members, sex) A + B
model_organisms 12 A

7. The shared rules every builder follows

  • The contract before the code, a failing test first. Each test is shown failing on the branch before the fix, and the commit message says so.
  • Each new column: the P5 audit (collisions, the reserved namespace, likely future names) in the commit message; the three touch points plus the parquet; reverse; a round-trip test (P7).
  • Each new warning code: a features/ scenario (the corpus gate stays on although RM149 is frozen) and a line in docs/INTEGRATION_0_8.md, since an older reader refuses an unknown code. The codes this plan names: phenotype_class_unrecognized, meta_selector_matches_nothing, meta_member_allele_ambiguous, log_file_not_shipped. Any other a builder adds is listed in its report.
  • Each new check is broken on purpose (mutate the condition, watch the test go red) before it lands.
  • Real paths, real data: the compile and reverse paths, the provisioned snapshots, expected values computed at run time. No mocking of a transformation.
  • An item that calls an outside service (D2, D3, C3) is marked building only once its first hand-run call and what it returned are recorded.
  • Commit as one block: the behaviour, its test, its doc line, its CHANGELOG line. uv run ruff format on the changed Python files first.
  • Forbidden: building around a decided column that turns out to have no case (send it to the pm); network in the format or compiler tiers; a refusal on data a published module already holds without the scan A2 describes; a count stated in prose that a test cannot re-derive.

8. Release gates

The cut waits for all of these, checked at the cut commit, never read off this list:

  • both charter amendments landed (P4 with A1, Goals with D1), each with its history entry;
  • the polars 2 lock and its RELEASE_RECORDS line;
  • docs/INTEGRATION_0_8.md covering every new warning code, manifest field (toolchain, vrs_version, tier_scale, the selector match counts), parquet and reason;
  • RM270's test_consumer_join pin removed and green;
  • the six reference examples compiling, reversing and recompiling to the same digest;
  • the review wave and the dogfooding wave come back clean (RELEASE_CYCLE § Minors steps 5–6), with longevitymap v2 itself as the dogfood.