feat(skill): add ancestry-risk-profiler

[Jake-BS]

Infers super-population ancestry (AFR/AMR/EAS/EUR/SAS) from a 23andMe/AncestryDNA file via AISNP Hardy-Weinberg likelihood scoring, then computes ancestry-stratified lifetime disease risk percentages using published GWAS effect sizes (GWAS Catalog, Pan-UKB, Biobank Japan).

Key outputs:

• Lifetime risk % per disease via Cornfield equation (P₀·OR / (1−P₀+P₀·OR))
• Ancestry Elevation Score (AES): how much ancestry amplifies risk vs EUR baseline
• Variant detail table with ancestry vs EUR OR contrast per rsID
• Optional matplotlib AES bar chart

Curated panel covers T2D, CAD, hypertension, CKD (APOL1), lactose intolerance, breast cancer, atrial fibrillation, Alzheimer's, and more. 32 TDD tests passing. Wired into CLAUDE.md, catalog.json, pytest.ini.

Summary

Skill affected

Checklist

• SKILL.md is present and complete (YAML frontmatter + methodology)
• Tests included and passing (pytest -v)
• Demo data included for reviewers to test
• No patient/sensitive data committed
• Follows CONTRIBUTING.md conventions

Test output

[manuelcorpas]

Audit verdict: request changes (safety)

Thank you for this contribution. The engineering is clean (no secrets, no unsafe code), but this skill needs rework before it can merge, on scientific-safety grounds.

1. Low-marker ancestry inference. The skill infers a super-population ancestry label from roughly 80 AISNPs and forces a single call with no abstention. On zero panel hits it silently defaults to EUR and continues. ClawBio's standing rule is that a thin AIM panel must not be presented as someone's ancestry, because it over-calls; the skill should abstain or use a high-resolution reference panel, and state documented ancestry rather than a fabricated label. SKILL.md line ~266 ("Do not refuse to run" on low confidence) is read and obeyed by the agent and directly defeats that guardrail. Please remove that instruction and add an explicit low-confidence abstention path.

2. Risk math. or_to_absolute_risk applies population-averaged ORs on top of population baseline prevalence, which double-counts the allele contribution and inflates the result (the demo "61.3% T2D" is an artefact). Risk alleles are also combined multiplicatively with LD ignored. For ancestry-stratified PRS, please use a validated, ancestry-appropriate polygenic score rather than summed log-ORs. The "Ancestry Elevation Score" is a novel metric with no literature backing and should not be surfaced as a clinical fact.

3. Genetic ancestry vs ethnicity. The trigger keywords and routing ("ethnicity disease risk") conflate inferred genetic super-population with ethnicity. Please separate these concepts explicitly.

4. Missing data files. data/aisnp_panel.csv, data/ancestry_risk_associations.json, and the demo patient file are referenced but absent from the PR, so the AIM count, allele-frequency sources, population labels and every OR are unverifiable, and the test suite cannot run. Please include them (with provenance) so the numbers can be checked.

Happy to help reshape this toward an abstain-by-default, PRS-backed design.

---

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[Jake-BS]

Audit verdict: request changes (safety)

Thank you for this contribution. The engineering is clean (no secrets, no unsafe code), but this skill needs rework before it can merge, on scientific-safety grounds.

1. Low-marker ancestry inference. The skill infers a super-population ancestry label from roughly 80 AISNPs and forces a single call with no abstention. On zero panel hits it silently defaults to EUR and continues. ClawBio's standing rule is that a thin AIM panel must not be presented as someone's ancestry, because it over-calls; the skill should abstain or use a high-resolution reference panel, and state documented ancestry rather than a fabricated label. SKILL.md line ~266 ("Do not refuse to run" on low confidence) is read and obeyed by the agent and directly defeats that guardrail. Please remove that instruction and add an explicit low-confidence abstention path.

2. Risk math. or_to_absolute_risk applies population-averaged ORs on top of population baseline prevalence, which double-counts the allele contribution and inflates the result (the demo "61.3% T2D" is an artefact). Risk alleles are also combined multiplicatively with LD ignored. For ancestry-stratified PRS, please use a validated, ancestry-appropriate polygenic score rather than summed log-ORs. The "Ancestry Elevation Score" is a novel metric with no literature backing and should not be surfaced as a clinical fact.

3. Genetic ancestry vs ethnicity. The trigger keywords and routing ("ethnicity disease risk") conflate inferred genetic super-population with ethnicity. Please separate these concepts explicitly.

4. Missing data files. data/aisnp_panel.csv, data/ancestry_risk_associations.json, and the demo patient file are referenced but absent from the PR, so the AIM count, allele-frequency sources, population labels and every OR are unverifiable, and the test suite cannot run. Please include them (with provenance) so the numbers can be checked.

Happy to help reshape this toward an abstain-by-default, PRS-backed design.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

I have made my best attempt to address your feedback here. Let me know if you are happy with the changes, or if more is needed. As a side note, I've noticed there are a few areas where code is repeated in the repo, like parsing 23AndMe data. I would be happy to do another PR making sure repeated functions like these are using what is available in common. And if it is not available there, moving that logic to common. Thanks

[manuelcorpas]

Re-audit verdict: changes still required.

Thank you for the substantial rework. Real progress: the silent EUR default is gone (now raises on insufficient coverage), the blanket "do not refuse to run" is narrowed, the absolute-risk double-count is removed, genetic ancestry is now explicitly distinguished from ethnicity, and the data files are included. Four of the five original points are fixed or partly fixed.

The headline safety issue, however, is not yet resolved, and the rework introduced a few new inconsistencies:

Prior issue	Status
Silent EUR default / no abstention	Partially fixed: abstains below the coverage floor, but still forces a single super-population label from ~80 AIMs with no abstention for the confident-but-wrong case
"Do not refuse to run"	Fixed (narrowed to a visible-limitation note above the coverage floor)
Risk-math double-count + invented score	Partially fixed: lifetime-% removed, but a multiplicative combined OR across LD-correlated loci is still surfaced (e.g. "4.39x") with LD not modelled
Ancestry vs ethnicity conflation	Fixed
Missing/unverifiable data	Files now present, but several OR citations do not match their variants (see below)

Blocking items:

1. Low-marker single-ancestry call. Inferring one super-population from ~80 AISNPs is the regime that over-calls non-European ancestry; "confidence" here is an uncalibrated log-likelihood gap, not a correctness guarantee. Please either abstain on low-separation calls (not just low coverage), or report ancestry as a soft posterior over populations rather than a single hard label, and avoid attaching disease ORs to a potentially-wrong single label.
2. Contradictory abstention thresholds. The code enforces 30 markers, the generated report's methodology says 16, and the error message states "30, which is 20% of the 80-marker panel" (30 is 37.5%). Please make the floor and its justification consistent in code, report, and SKILL.md.
3. Inflated combined OR. Removing the lifetime-% relabelled the figure rather than fixing it; multiplying ORs across correlated T2D loci with LD unmodelled overstates the magnitude. Please use an ancestry-appropriate validated PRS, or stop combining correlated ORs multiplicatively.
4. OR citation mismatches. Several pmid/OR pairings do not match the variant (e.g. an APOL1/CKD row cites an atrial-fibrillation PMID; a TCF7L2 diabetes OR cites a lactase paper). Please reconcile every effect size against its actual source so the numbers are verifiable.
5. Minor: the now-unused baseline_prevalences block in the associations JSON should be deleted rather than orphaned.

Happy to help reshape the ancestry call toward a soft-posterior, abstain-on-low-separation design.

---

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

Audit verdict: REQUEST CHANGES

Thanks for the careful framing here, the abstain-on-low-coverage logic, the soft confidence levels, and the "genetic super-population is not ethnicity" signposting all show you know the over-call trap. The intent is right, but three issues need fixing before this can merge.

1. Combined-OR math inflates risk (ancestry_risk_profiler.py:646). compute_disease_risks multiplies single-locus ORs into a headline figure (the SAS demo reports "4.39x Type 2 Diabetes"). This multiplies across loci that are in LD, three KCNQ1 markers (rs2237892/rs2237895/rs163184) and two 9p21 CAD markers (rs1333049/rs10757278), as if independent. The Methodology section even states LD is not modelled. Please either drop the multiplicative combined-OR or restrict to LD-pruned independent signals with an explicit uncertainty band.

2. Effect-size provenance does not match the cited PMIDs (data/ancestry_risk_associations.json). PMID 22158537 (a breast-cancer GWAS, per your own PROVENANCE.md) is cited as the source for SAS Type 2 Diabetes, CAD, Hypertension and Alzheimer's (11 entries). PMID 17603472 is listed twice as two different papers. The Safety claim "all effect sizes trace to cited PMIDs" is therefore not currently true. Please re-derive each OR from the correct ancestry-matched GWAS and fix the citations.

3. The coverage gate is internally inconsistent. The code enforces MIN_AISNP_COVERAGE = 30 but the generated report tells the user "minimum 16 markers (>=20% of panel)", and 30/79 is 38%, not 20%. The "30 = reliable continental assignment" floor is attributed to Kosoy et al. 2009, which validated a 128-AIM panel, not a 30-marker floor. Please reconcile the numbers and correct the citation. Separately, baseline_prevalences is still shipped in the JSON and the catalog entry still advertises "lifetime disease risk percentages", which contradicts the SKILL.md "no absolute risk" scope.

No tests assert the combined-OR magnitude is bounded, and the APOL1 ancestry-specific path is untested on real output. Adding boundary tests for both would catch regressions of this class.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

ClawBio domain audit — REJECT (request changes)

The automated scanner passed this PR, but a deep domain review found four blocking issues.

1. Ancestry over-call (critical). ancestry_risk_profiler.py:186-203 selects a single best super-population (max of log-likelihoods) and returns one hard label regardless of confidence; the only abstention is the <30-marker coverage floor. The bundled demo itself returns SAS, confidence low, coverage 64 and still emits a full risk report. A thin AISNP panel at low confidence must return a soft posterior over populations (or abstain), not a confident single label that then drives clinical-sounding output.

2. Inflated risk math (critical). compute_disease_risks (:269-290) sums log(OR)*dosage across loci and exponentiates, producing e.g. a T2D combined_OR = 4.39x that is the raw product of 7 single-variant ORs and is advertised as the headline (SKILL.md:167). A product of per-SNP ORs is not a validated aggregate; use a calibrated PRS or report honest uncertainty.

3. APOL1 modelled multiplicatively, but it is recessive (critical). rs73885319 (G1) and rs60910145 (G2) are scored as two independent per-allele entries and multiplied (ancestry_risk_associations.json:534-560). APOL1 nephropathy risk is recessive (two risk alleles). A per-allele log-additive product is biologically wrong for the two highest-OR claims in the panel. The demo is hom-ref, so it hides this.

4. Citations do not match their claims (critical). Verified against PubMed: PMID 23340282 (cited for the APOL1 CKD ORs) is a chemistry electrochemiluminescence-sensor paper; PMID 20686008 (cited for ALDH2 rs671 oesophageal cancer) is an anaesthesiology case report; PMID 17571276 is a Cdk5 neurodegeneration paper. PROVENANCE.md is scrambled against both the JSON and reality, so the SKILL.md "no hallucinated ORs / cited PMIDs" claim is inaccurate.

Also: catalog.json still advertises "lifetime disease risk percentages" that the code claims to have removed; the coverage threshold disagrees between code (30), report text (16 / 20%), and the error message; aisnp_panel.csv:16 has a trailing space in an rsid (a silent dead marker).

Requested before re-review: soft-posterior-or-abstain ancestry output; drop the combined-OR / AES construction for a validated PRS or explicit uncertainty; a recessive model for APOL1; full citation re-verification against PubMed / GWAS Catalog; reconcile catalog.json and the coverage threshold.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

Audit verdict: REJECT (request changes)

Deep domain re-audit (beyond the secret/lint scanner, which passed this PR as SAFE). The skill is well-engineered around its caveats, but the underlying data is not safe to ship.

Critical

1. Fabricated citations across the association panel. Every spot-checked PMID in data/ancestry_risk_associations.json / PROVENANCE.md resolves to an unrelated paper. Two verified directly against PubMed:

• 14507249 is cited as Enattah 2002 (LCT lactase persistence); it is actually Hashibe 2003, bladder-cancer survival.
• 22561518 is cited as Wu 2012 (ESCC / ALDH2); it is actually Jin 2012, a vitiligo GWAS.

The two PMIDs PROVENANCE.md says were "corrected from v1.1 audit findings" (APOL1, ALDH2) are themselves wrong. The Safety claim "No hallucinated ORs / cited sources" is therefore false, and every OR in the panel is unverifiable as-is.

2. Biologically incoherent LCT metric that ranks #1 in the skill's own demo. The LCT entries score the same C risk allele as protective in EUR (or=0.45) and as risk in non-EUR (or=6.80 EAS, 3.20 SAS, 4.50 AFR). The ancestry-effect ratio OR_anc/OR_eur then mixes two opposite outcome framings and manufactures a 7-15x "ancestry amplification." Running the bundled demo patient surfaces "Lactose Intolerance | AES 7.11 | N=1" as the top result, sourced to a bladder-cancer paper. The ## Example Output table omits this row. Separately, an EAS OR of 6.80 at rs4988235 is not credible: rs4988235 is the European persistence marker and is essentially ancestral/monomorphic in East Asians (who use rs4988236 / -13908).

3. Hard ancestry label gates disease risk off a self-flagged low-confidence call. compute_disease_risks filters assoc["ancestry"] != ancestry on the single best-match label, even when infer_ancestry returns confidence="low" on the demo. The soft posterior is displayed but not used; the gating is still hard. This is the documented low-marker ancestry over-call class, compounded into the fabricated ORs above.

Required before reconsideration

1. Re-source and verify every PMID/OR from the primary literature; add a test that binds each OR to a resolvable, topic-matched source.
2. Remove or rebuild the LCT entries using the correct per-population persistence variants; fix the AES denominator so it never compares opposite outcome framings.
3. Keep ancestry as a soft posterior end-to-end, or abstain below a marker-count threshold; do not hard-gate disease ORs on a low-confidence label.
4. Add a test covering the direction artifact (the current test_aes_gt1_for_sas_t2d does not).

Local-first handling is correct (no remote patient-data fetch) - that part is fine.

---

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

Re-audit (v1.3.0) verdict: Changes requested

Thanks for the substantial rework. Several of the previous findings are genuinely fixed, and that is real progress:

• LCT artifact removed — no rs4988235 entries remain; the spurious 7-15x AES is gone.
• AES opposite-framing fixed and absolute lifetime-risk % removed (no more baseline double-counting).
• ALDH2 (rs671) PMID set to null with an honest provenance note rather than a third wrong guess.
• Soft-posterior gating added (scoring skipped when top posterior < 0.35) and the AISNP >=30-marker abstention (Kosoy 2009) retained.

However, the original blocking reason — fabricated / mismatched citations — is not resolved. I re-resolved the high-reuse PMIDs in data/ancestry_risk_associations.json against PubMed. Five of six are wrong:

PMID	What it actually is	How v1.3.0 uses it	Issue
20566908	Tschiesner 2010, "Health professional perspective on disability in head and neck cancer"	APOL1 rs73885319 / rs60910145 (AFR), compound_recessive_groups, AND the formal Citations list as "Genovese et al. 2010 Science, APOL1"	Fabricated; unchanged from the previous review
22158537	Cho 2011, East Asian T2D GWAS (Nat Genet)	the SAS source for ~11 variants incl. rs429358 (APOE/Alzheimer's), rs10455872 (LPA/CAD), rs4343/rs699 (hypertension)	Wrong population (EAS, not SAS) and wrong trait for non-T2D variants
27005778	Kettunen 2016, European circulating-metabolites GWAS (Nat Commun)	the AFR source for rs2981582 (FGFR2 breast cancer), rs1333049 (CAD), rs4343/rs699 (hypertension), rs1800497 (ANKK1)	Wrong population (European) and wrong trait
23945395	Hara 2014, Japanese T2D GWAS (3 loci: MIR129-LEP/GPSM1/SLC16A13)	EAS source for rs429358 (Alzheimer's), rs2981582/rs3803662 (breast cancer), rs1333049 (CAD)	Wrong trait; none of these variants are in that paper
17478679	Helgadottir 2007, 9p21 myocardial infarction (Science)	reused for rs11591147 (PCSK9), rs4343 (ACE), rs699 (AGT), rs1800497 (ANKK1)	Cross-citation; unchanged from the previous review
16415884	Grant 2006, TCF7L2 T2D discovery	rs7903146 EUR (TCF7L2 T2D)	Correct

The pattern is systematic: one representative PMID per super-population is applied as the source for every variant in that population, regardless of the actual trait, or even the population the paper studied. The effect-size values may be plausible literature numbers, but the provenance does not hold: the cited paper does not contain the claimed variant / trait / population association. "Cited but wrong" is the same safety hazard as "missing". The SKILL.md statements "All effect sizes trace to the bundled JSON" and "most association entries include verified PMIDs" are therefore not accurate yet.

Required before reconsideration

1. Re-source every PMID from the primary literature, per (variant, trait, population). A single paper cannot be the source for an entire super-population across unrelated diseases.
2. Fix 20566908: the real Genovese APOL1 paper is PMID 20413513 (Science 2010, 329:841). Correct both the data file and the Citations list.
3. Add a test that binds each active OR to a resolvable, topic- and population-matched PMID, so this class is caught in CI rather than by manual resolution.

The structural improvements are good and worth keeping; the remaining work is provenance integrity on the association panel.

---

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

Correction to my previous review.

I cited the correct Genovese APOL1 paper as PMID 20413513. That is wrong: 20413513 is an unrelated diabetes drug trial. The correct citation is PMID 20647424 (Genovese et al. 2010, Science 329:841, "Association of trypanolytic ApoL1 variants with kidney disease in African Americans"), verified via PubMed E-utilities and the GWAS Catalog.

The substantive finding is unchanged: the PMID in the panel, 20566908, is still a head-and-neck-cancer disability survey, not the APOL1 paper. Only my suggested replacement number was wrong.

Apologies for the error, which is precisely the "asserted a PMID from memory instead of resolving it" failure the rest of this review is about. It is now corrected in the provenance-gate work (ClawBio/ClawBench#4), and is the reason that gate resolves every PMID against PubMed + GWAS Catalog rather than trusting any human or model to type the right number.

[manuelcorpas]

PR Audit: Request Changes

Thanks for this — the skill is genuinely well engineered against the usual ancestry pitfalls (79-AISNP panel, enforced 30-marker abstention, soft-posterior gate, recessive APOL1 model, local-first, and a thorough test suite that exercises the abstention paths). The blocker is provenance, which for an ancestry-stratified disease-risk tool is a core safety issue rather than a footnote.

Several association PMIDs do not match the variant/disease/population they are cited for (verified against PubMed):

• PMID 17478679 is Helgadottir et al. 2007 Science, a chromosome 9p21 myocardial infarction / CAD GWAS. In data/ancestry_risk_associations.json it is cited (6×) as the source for hypertension (rs4343/rs699) and Parkinson's (rs1800497) effect sizes. (PROVENANCE.md also mislabels the author as "McPherson".)
• PMID 22158537 is Cho et al. 2011 Nat Genet, "eight new loci for type 2 diabetes in east Asians." It is labelled here as "Kooner et al. — T2D South Asian" and cited for SAS hypertension, CAD, and Alzheimer's. Wrong author, wrong population, wrong phenotypes.
• PMID 27005778 is blanket-cited across six different AFR diseases (T2D, hypertension, CAD, breast cancer, AF, Parkinson's), and PROVENANCE.md itself labels several entries "proxy ... (re-verify)".
• rs671 ALDH2 ESCC ships "or": 1.89 with "pmid": null after two prior wrong PMIDs.

Secondary items to resolve in the same pass:

• Version strings disagree: SKILL.md/code say 1.3.0, ancestry_risk_associations.json metadata says 1.1.0, catalog.json says 1.2.0. Please reconcile — the correction history is load-bearing for trust here.
• catalog.json tags the skill lifetime-risk and adds the trigger "which diseases am I at risk for", which contradicts the skill's explicit "does NOT compute lifetime risk" scope and will misroute exactly those queries.
• The combined OR is a naive product across loci (demo T2D = 4.39×; an HFE C282Y homozygote → ~96×). The disclaimers and one-variant-per-LD-block design are good, but consider not leading the table with the aggregate figure.

Requested before merge: re-resolve every association's PMID to a paper that actually reports that variant–disease–ancestry effect (or drop the entry, as you correctly did for LCT); verify or remove the null-PMID ALDH2 entry; reconcile the version strings; fix the catalog.json scope tags.

Automated audit by ClawBio pr-audit, with PMID resolution verified against PubMed. Manual review is still recommended.

[manuelcorpas]

PR Audit: REJECT (request changes)

The static scanner passed this PR with zero findings, but a deep genomics domain review found four CRITICAL issues. The safety scaffolding (abstention gate, LCT removal, APOL1 recessive handling, disclaimers, Agent Boundary) is genuinely good; the problem is the data layer and the risk math.

CRITICAL

C1. Fabricated provenance (single-source back-attribution). In data/PROVENANCE.md / ancestry_risk_associations.json, one paper is reused as the source for every disease in an ancestry. PMID 22158537 (Kooner 2011, a T2D South Asian GWAS) is cited for SAS hypertension rs4343/rs699, CAD rs10455872, and Alzheimer's rs429358. A T2D GWAS cannot be the source of a hypertension or Alzheimer's odds ratio. Same pattern for EAS (23945395) and AFR (27005778). These ORs drive every "Elevated By Ancestry" call.

C2. EUR reference ORs traced to a wrong-disease paper (self-flagged). PMID 17478679 (a 9p21 CAD paper) is cited as the primary source for PCSK9 rs11591147, hypertension rs4343/rs699, and Parkinson's rs1800497, tagged "(re-verify)" in PROVENANCE.md. The eur_or denominator of the AES calculation is therefore wrong for these entries.

C3. HFE haemochromatosis modelled additively. rs1800562 "or": 9.80 has no model field, so compute_disease_risks applies it on the additive path: a homozygote (dosage 2) gets exp(log(9.8)*2) ~ 96x, and a clinically-benign heterozygous carrier gets a spurious 9.8x. APOL1 was correctly special-cased as recessive; HFE (the same bug class) was missed. No test bounds the combined OR magnitude.

C4. Harm surfaced to lay users. The headline table tells a South Asian user "Type 2 Diabetes 4.39x Elevated By Ancestry," where 4.39x is a naive product of 7 largely unverifiable non-EUR ORs. The exploratory caveats do not neutralise a confident ancestry-attributed disease multiplier shown as the primary output.

WARNING

• AISNP panel padded with near-zero-Fst markers. PROVENANCE.md claims selection for Fst > 0.3, but the panel includes rs4680 COMT, rs1801133 MTHFR, VDR, rs53576 OXTR, rs1800497 DRD2, all with Fst near 0. MIN_AISNP_COVERAGE = 30 counts these, so a file can clear the abstention gate on ~30 uninformative SNPs.
• Version drift. Skill reports v1.3.0, JSON metadata says 1.1.0, catalog.json registers 1.2.0.

Required before merge

1. Re-source every or/eur_or to a matching primary GWAS with a verified PMID whose study disease matches the association.
2. Move HFE (and any other known-recessive large-OR locus) to a recessive model with a magnitude-bounding test.
3. Restrict the coverage gate to genuinely high-Fst markers.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

ClawBio automated audit — verdict: NEEDS REVIEW (changes requested). The ancestry handling is genuinely well done: 79-AISNP panel with a 30-marker hard floor, soft posterior over super-populations (no hard label or admixture percentages), and disease scoring gated off at low confidence. That correctly avoids the low-marker over-call trap. The blocker is provenance, not ancestry inference.

Please address before merge:

1. Mismatched effect-size citations (blocking). Several disease-ancestry ORs carry PMIDs for papers that never studied that disease. Your own data/PROVENANCE.md documents one: PMID 17478679 (McPherson 2007, CAD 9p21) is attached to Hypertension and Parkinson's EUR ORs as a "proxy … re-verify". Same pattern with single-paper blanket citations spanning multiple unrelated diseases (e.g. PMID 27005778 across six AFR diseases; Kooner 2011, a SAS T2D GWAS, cited for SAS hypertension/CAD/Alzheimer's). In a tool that tells users their ancestry amplifies disease risk, a wrong citation is as serious as a missing one. Please re-verify each OR against its stated source, or set the PMID to null (as the ALDH2 entry already does).
2. rs1800497 (ANKK1/DRD2 TaqIA) as a Parkinson's locus looks invented — that is a dopamine/addiction marker, not an established PD GWAS hit. Please confirm the source or remove it.
3. Composite "4.39x (N=7)" T2D score is a naive product of per-SNP ORs (LD not modelled). The caveats exist but only in text while the multiplier is the headline. Consider softening the surfaced number or leaning on the validated-PRS signpost.
4. Housekeeping: version strings disagree (SKILL.md 1.3.0 / associations JSON 1.1.0 / catalog 1.2.0); the inert baseline_prevalences block is still in the JSON and is one wiring change away from re-introducing the removed double-counting bug — recommend deleting it.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

ClawBio re-audit (commit b68032e) — most of this is genuinely fixed; one requested item remains, so the change-request stays open.

Verified against the updated diff. Strong progress:

• Mismatched PMIDs resolved. All four flagged PMIDs (17478679, 27005778, 22158537, 23945395) are re-sourced or dropped, and — this is the best part — it's now CI-guarded: TestProvenanceIntegrity bans a _KNOWN_BAD_PMIDS set, forbids any single PMID spanning more than one disease group, and enforces OR-range sanity. Exactly the kind of gate that stops this class recurring. Thank you.
• rs1800497 (ANKK1/DRD2) removed from the AISNP panel and its PD association re-sourced to a real PD paper (PMID 21663922).
• Composite OR softened: the combined-OR column is dropped from the rendered report with an explicit "naive product … overstates risk" caveat.

Remaining blocker (the one requested change not done):

1. Delete the inert baseline_prevalences block from ancestry_risk_associations.json (still present ~L1541-1602, including diseases not even in the panel), and add a test asserting its absence. SKILL.md already claims it was removed, so right now the docs contradict the data, and the block leaves the previously-removed double-counting one wiring change away.

Two smaller doc-vs-data reconciliations while you're there:
2. Version drift — the data reflects v1.3.2-level edits (panel 79→72, EAS/APOL1/HFE changes) but every version string still reads 1.3.0. Bump them to match.
3. Stale rs671 claim — the JSON now sets pmid: "22960999", but SKILL.md and PROVENANCE.md still say rs671's PMID is null "pending re-verification." Update the docs to match the data.

Once the baseline_prevalences block is gone (with a guard test) and the docs are reconciled, this is good to merge. The core science is in good shape.

Automated audit by ClawBio pr-audit. Manual review is still recommended.

[manuelcorpas]

ClawBio audit — resolved and approving. The remaining blocker is fixed via maintainer edit (commit 37b2f65), rebased on current main:

• Inert baseline_prevalences block removed from ancestry_risk_associations.json (it was read by zero code paths after absolute-risk scoring was dropped in v1.3.0). A new CI test test_no_baseline_prevalences_block asserts its absence so it cannot silently return.
• rs671 (ALDH2 ESCC) docs reconciled with the data — the entry uses PMID 22960999 (Wu 2012 Nat Genet); SKILL.md and PROVENANCE.md no longer describe it as pmid: null. The GWAS-Catalog accession-to-PMID caveat is retained in the entry note.
• Versions aligned to 1.3.1 across SKILL.md, the associations JSON, and catalog.json; v1.3.1 changelog added.
• Merge conflict with main resolved (catalog.json now carries the full merged skill set plus this entry).

All 64 skill tests pass locally. The science content and the earlier TestProvenanceIntegrity CI gate (bad-PMID ban, no PMID spanning disease groups, low-FST marker exclusion) are unchanged and green. Thanks for the thorough iteration on the provenance work.

Automated audit by ClawBio pr-audit. Manual review is still recommended.