PubMind — what it competes with, what it complements, and how we would adopt it¶

Status — assessment complete; adoption decided 2026-08-28 in PROPOSAL_0_7 § RM134, which is authoritative where the two disagree. This document stays the evidence — what was probed, measured and read — and is not rewritten to match the decisions; where a decision changed something below, an inline note says so and names it, marked Note (date). Written 2026-08-28 against the paper's accepted version and against the database's own bytes, both probed the same day. The item was RM134, all four sections of which shipped in 0.7 on 2026-08-28; the complement this assessment kept asking for arrived as RM167 on 2026-09-01.

PubMind (Wang & Wang, Nat Commun, 20 August 2026, doi:10.1038/s41467-026-76834-4) runs a fine-tuned DistilBERT triage stage over 41.7 M PubMed abstracts and 5.4 M PMC full texts, hands the surviving paragraphs to LLaMA-3.3-70B with per-variant-class prompts, and extracts variant–disease–pathogenicity associations with the model's own reasoning attached. Diseases are normalized to MONDO and phenotypes to HPO by PubMedBERT cosine similarity at a 0.9 threshold; coordinates come from PyEnsembl against Ensembl v111. The output is PubMind-DB: 1,473,920 literature-derived records consolidated into 765,260 unique variants across SNVs, gene fusions, SVs and CNVs, of which 916,538 SNV rows carry genomic coordinates.

The short answer¶

We barely compete, and the overlap is not on the axis either project cares about. Their own discussion calls PubMind "a literature-grounded complement to human curated databases", and that is also the correct description of its relationship to us: PubMind is a source, of the same kind as ClinVar, gnomAD or the GWAS Catalog, and our enricher already consumes four such sources. It produces per-variant assertions; it does not produce a signed, versioned, licence-declared annotation artifact that a consumer joins a genotype against, and nothing in the paper suggests it wants to.

The one genuinely contested surface is a sentence in their introduction: "PubMind empowers healthcare systems to construct secure, institution-specific variant-interpretation databases directly from clinical notes and reports." That is the same customer we would reach with an authored module — an institution that wants its own interpretation table. But the thing PubMind hands that institution is a SQLite database behind a Flask app, with no identity model, no integrity binding, no round trip, and no licence axis. The competition is for attention, not for the artifact.

Where we complement is much larger, and it is one-directional: they are upstream of us. Their strongest claim is coverage of what nobody curated — only 10.6 % of their coordinate-mapped SNVs overlap ClinVar, and the bioinformatics predictors they benchmark against return "unknown" for most variants they cover. Our strongest claim is everything downstream of an assertion: variant_key and VRS identity, the three-valued algebra, the compile gate on declared use, the round trip, the signature family. Those are disjoint. A PubMind row is exactly the sort of evidence an author would want surfaced while curating, and exactly the sort of claim our checks exist to cross-examine.

What we actually probed¶

Everything below is measured, not read off the paper. The inputs are git-ignored under data/input/; re-fetch them with:

What	Where	Pinned at probe time
Accepted manuscript	`https://www.nature.com/articles/s41467-026-76834-4_reference.pdf`	sha256 `1002f535468393ee57e18c81a41aff9ba4a37c690c7cb80816cdded8763e586a`
Coordinate table	`https://www.openbioinformatics.org/annovar/download/hg38_pubmind_db.txt.gz`	sha256 `ca224ebf5ef3aa516d9f14b91ae96ce2371ea785670f4723fd5103e6f94dc047`, ETag `"63275d-659cb3f35fd80"`, Last-Modified Mon 24 Aug 2026 13:48:54 GMT, 6,498,141 bytes
Summary API	`https://pubmind.wglab.org/api/summary`, `/api/pvid/<PVID>`	probed live
Source + terms	`https://github.com/WGLab/PubMind` (`README.md`, `LICENSE.md`)	probed live

There is exactly one per-variant channel, and it is the ANNOVAR table¶

The web API has two endpoints and neither takes a variant. /api/summary accepts gene, MONDO_ID_09, MONDO_name_09 and formatted_reference (with pmid/pmcid/mondo_id/mondo_name aliases) — an rsID query is refused with the allowed-field list — and it returns aggregate counts and mean scores per variant class, nothing per row. /api/pvid/<PVID> returns one variant's summary but requires you to already know the PVID, which only the website will tell you. Both responses carry an explicit note: "Detailed per-record information is not returned through the API."

So the only channel that answers "what does PubMind say about chr6:26092913 G>A" is the bulk table ANNOVAR redistributes as hg38_pubmind_db, dated 2026-08-24. It is small (6.5 MB gzipped, 909,224 rows) and carries five useful columns: PVID, pathogenicity_sum, paper_level_pathogenicity_score, confidence (0–3), on top of Chr/Start/End/Ref/Alt.

Its coordinates are VCF-style, not ANNOVAR-style, despite the packaging. There is not a single - allele in the file; a one-base deletion appears as 1 1014264 1014265 CC C, anchor base retained. That is our chrom/start/ref/alts convention exactly, start being the 1-based VCF position (@start-1based), so a join needs no coordinate translation. The indels are not left-normalized (answered 2026-09-27, RM273): 5,158 of the 20,006 single-base-anchored indel rows have an anchor base equal to their payload's last base, so each spells an event that also sits one base further left.

The channel is smaller than "1.3 million variants" implies¶

909,224 rows sound like coverage. They are not, because two thirds of them are enumerated codon alternatives rather than observed variants. When PubMind recovers only a protein change from the text, it back-maps through the transcript and writes out every codon that could encode it:

Row shape	Rows	Joinable?
single-base substitution (`1/1`)	289,615	yes, directly
codon triplet differing at exactly one base	160,090	yes, after decomposition
codon triplet requiring 2 simultaneous substitutions	312,875	no — not a single event
codon triplet requiring 3 simultaneous substitutions	126,508	no
length-changing (indel)	20,131	yes, but normalization unverified
longer MNV blocks (4–19 bases)	5	no

Decomposing the one-base codons onto their differing position gives 342,209 distinct chrom:start:ref:alt keys over 305,935 distinct loci — the honest size of the joinable SNV layer. The remaining 439,388 rows assert that some codon change produced the reported amino acid, which is a statement about the protein, not about a position a consumer can genotype.

PubMind has record identity, not variant identity¶

This is the deepest structural difference and it is worth stating precisely, because it is the reason a PubMind row can never be adopted as a fact without a check in front of it.

Note (2026-09-03). The half this limit leaves open — which papers name this allele — is what LitVar2/PubTator3 answers, and RM167 (shipped 2026-09-01) adopted it as litvar coverage: the identity there is genuinely allele-level (BRAF rs113488022's three ClinGen allele ids carry 31,276 / 99 / 41 papers), and the answer is recorded with the tier that gave it, allele-resolved, position-only or absent. The measurement in this section stands; what changed is that the fan-out is no longer an unanswered structural limit but the reason two sources are adopted for two questions.

Consolidation into a PVID is keyed on the text the model extracted — gene symbol plus cDNA change, protein change or rsID — never on a coordinate. So one physical variant fragments into many PVIDs. 68,744 coordinate keys (8.4 %) carry more than one PVID; the worst carries 35. At chr6:26092913 G>A — HFE C282Y, one of the best-characterised variants in medical genetics — the table holds eight PVIDs, verdicts split four Conflicting, two Uncertain, one Benign, one more Benign at a G>G row that is not a variant at all. (523 rows across the file have Ref == Alt.)

One of those eight is instructive. PVID926871 carries rsid: rs1800562 and gene: TMPRSS6. rs1800562 is HFE, chromosome 6; TMPRSS6 is chromosome 22. Two iron-metabolism papers discussed both genes, the model crossed them, and because rsID→coordinate resolution is authoritative the row still landed on the correct C282Y position — carrying the verdict Benign, score 0.0, confidence 2. Our _gene_locus_conflicts check (identifiers.py, reported as gene_locus_agreement) catches exactly this shape (@gene-locus-relationship: check the relationship, not the members, at chromosome granularity). It is a clean demonstration that PubMind's output wants our checks pointed at it, and a clean demonstration of why we would never carry one of its verdicts into an artifact unexamined.

What it is worth against our own corpus¶

Joined against every GRCh38 resolution.csv in reference_examples/ (11 modules, 423 resolved loci; cyp2c9_warfarin_grch37's 5 GRCh37 rows are skipped, and five modules — cyp2d6_structural, fmr1_cgg_repeat, grch37_build, htt_repeat_expansion, mt_heteroplasmy — carry no resolution sidecar):

173 of 423 loci (40.9 %) are known to PubMind at position level.

Note (2026-09-03). The 423-locus denominator did not reproduce when RM167 (2026-09-01) re-derived the roster from the DRAFTABLE registry over the same eleven modules: it found 389 loci and 388 distinct rsIDs. The percentages on this page keep their original denominator as the evidence they were, and the comparable figure for a re-run is RM167's. - 190 of 589 authored ALTs (32.3 %) match a PubMind row exactly. The gap between the two numbers is the usual one: an rsID is position-level and a verdict is allele-level (@rsid-not-per-allele). - Over every variants.csv in the corpus that authored a clin_sig and resolved to a GRCh38 position — a slightly wider set than the sidecar join, since some modules carry coordinates on the row itself — PubMind returned a verdict for 134 loci and the two agree on 83 (62 %). Every disagreement runs the same direction: we say pathogenic, they say Uncertain (32) or Benign (19), which is what a literature-wide aggregate should do to a corpus deliberately seeded from ClinVar's pathogenic end.

Our 40.9 % locus coverage against a ClinVar-derived corpus sits comfortably beside their reported 10.6 % ClinVar overlap: the two measure opposite directions of the same asymmetry. PubMind adds breadth we do not have and would not gain by curating harder; it subtracts confidence we already have. That is the profile of a cross-check source, not of a fact source.

How PubMind places a variant, read from its code (2026-09-28)¶

PubMind's site has a Variant Normalizer (Beta) page (pubmind.wglab.org/variant_normalizer), which it says is "derived from PubMind's variant normalization module". It takes one gene:variant line per input (BRCA1:c.68delA, BRAF:p.Val600Glu) and offers a JSON download. It is not the VICC variation-normalizer that METAKB_VARNORM_SURVEY.md ran. The two share a name and nothing else: VICC mints VRS ids through cool-seq-tool, UTA and SeqRepo, while PubMind maps positions with pyensembl.

The mapping step is normalization/4-genome_coordinates/variant_genome_coordinates.py in WGLab/PubMind, read at the default branch on 2026-09-28 (last push 2026-08-24; GitHub reports its licence as NOASSERTION). What the code does:

The engine is pyensembl EnsemblRelease(111) (a comment above it still says release 104), with Biopython to translate the CDS. This matches the Ensembl v111 already stated above.
No transcript is chosen. cdna_to_genomic(gene, pos, ref) and protein_to_genomic(gene, pos, ref_aa) walk every protein-coding transcript of every gene the symbol names. They keep each one whose CDS has the stated reference base (or amino acid) at the stated position and return all of them. A text mention names no transcript, so a c. number is tried against every isoform, and each isoform that happens to carry the same base there becomes a separate placement.
It maps a position, not an event. The cDNA path maps one base, the protein path one reference codon. Nothing in this step reads an indel's extent or normalizes one, which fits the un-normalized indels RM273 measured.
Nothing is identified. No VRS id, CAID or rsID comes out of this step.

Why it matters to the lane. Our PubMind snapshot receives coordinates this code produced. A PubMind row sitting on an unexpected locus can therefore come from an isoform the author never meant, whose reference base happened to match, rather than from a wrong extraction. The two causes call for different remedies, and the table alone cannot tell them apart. Not measured: how often one PubMind mention (one PVID) fans out to more than one locus through this multi-isoform match. The 68,744 keys counted above are the opposite direction, one locus carrying several PVIDs. The fan-out count is the number that would size this, and it can be taken from the snapshot we already build.

What we cannot adopt at all¶

The part of PubMind that is genuinely novel — the per-record LLM_reasoning, the formatted_reference list, the supporting evidence passage behind each call — is not available through any open channel. The website serves it to a human; the API explicitly withholds it; the README routes anyone wanting it to CHOP's licence terms. So the material closest to a provenance_quote is the material we cannot reach, and no plan below should imply otherwise.

The terms need stating carefully, because two different things are licensed and only one of them clearly.

The software is covered by LICENSE.md: academic, non-commercial research use only, commercial use prohibited without a separate agreement from CHOP's Office of Technology Transfer. That is the same shape as the dbNSFP and OMIM terms licensing.py already names in the comment above CLINPGX_TERMS — academic-use-only, which bars passing data on at all, and which commercial_use=False alone understates.
The paper is CC BY-NC-ND 4.0, which expressly forbids sharing adapted material.
The ANNOVAR-distributed coordinate table carries no stated terms of its own. The README offers it as the open substitute for the licensed full database, and ANNOVAR redistributes it, but neither publishes a licence for those bytes. Under the house rule that is unknown, not permissive (@no-named-licence): unknown commercial terms warn, they never gate, and we record None rather than guessing, because licensing.py defines null as "the terms could not be established".

Running the PubMind pipeline ourselves is not a scoping judgement, it is barred: the Constitution's non-goals forbid pulling LLM SDKs into any tier of this workspace, and AI-assisted authoring is just-dna-pipelines' job by name. The cost argument — multiple H100/A100-class GPUs, by the authors' own account — is true and beside the point.

What the charter decides¶

Read in full before the plan below was written, because P3/P8 decide whether any of it is even legal.

Most of "do we compete" is settled by a non-goal rather than by an argument. The Constitution says this workspace does no gene–disease inference: the format catalogs curated annotations that consumers join against variant data, and interpretation belongs to those consumers. PubMind is an inference engine over the literature — that is its entire contribution. So the two projects are not adjacent by accident; the thing PubMind does best is a thing we are constitutionally not in the business of doing, which is precisely what makes it a good source and a poor competitor.

All four sections are legal, and the batch is a minor. Nothing below adds, removes, retypes or promotes a schema field, so Principles 3 and 8 are not reached — no already-published module becomes invalid, and there is nothing for a major to gate. All of it lives in just-dna-enricher, the only tier permitted to fetch (Principle 2), and none of it writes into the compile path, which never imports it. What makes the batch a minor rather than a patch is new commands, a new snapshot kind and new warning phrases, every one of which a consumer can grep for; a corrected derivation would be the other case and this is not one.

The derived snapshot is priced at approximately nothing, and that is Principle 9 doing its job. A parquet the machine writes and no author ever types is the free layer. The section that would cost something is the one nobody is asking for: a pubmind_score authored column would be full price, and Principle 5 rules on it before the price matters. A single score covering how many papers spoke, how confidently, and in which direction is three axes in one field — the anti-pattern Principle 5 exists to name, made permanent for the rest of the major by Principle 3. The snapshot keeps them apart, clin_sig / pathogenicity_score / confidence, and the split is not cosmetic. (Written as pubmind_sig; renamed before it shipped — see the supersession note below.)

Drafting is where Principle 3's one-way door actually bites. Drafted rows land in variants.csv using columns that already exist, so the draft itself adds nothing — but --min-confidence, derivation and the pubmind entry in DRAFT_PROJECTIONS are names that become permanent the moment they ship, which is exactly what Principle 5's "audit every new name against likely future additions" asks for. derivation in particular wants checking against the possibility of a second literature-derived source arriving later.

Nothing here touches Principle 7. No section changes what compile → reverse → compile carries. A drafted row round-trips because it is an ordinary authored row; the snapshot is not part of the artifact at all. That is what keeps the whole batch additive.

The plan¶

Directed 2026-08-28: build both halves. An earlier draft of this section stopped at a report-only cross-check and recorded everything downstream as blocked. That is now overtaken — the ask is a ClinVar-shaped derived table with PubMind as the annotation authority, plus a three-way cross-check reporting concordance and discordance against ClinVar. The licensing constraint has not changed and is not waved away; it moves from reason not to design to precondition on shipping, and Gate below says exactly which step it holds.

One word needs pinning before anything else. "Authority" here means an authoritative annotation source in the sense ClinVar is one — never resolution.csv's authority column (@source-vs-authority). Nothing PubMind produces may ever enter resolution.csv. Its coordinates are PyEnsembl back-mappings of text the model extracted, codon-enumerated and multi-PVID; treating that as a resolution link would put a derived guess where a resolved identity belongs. PubMind annotates loci that something else resolved.

A. `pubmind build` — the derived snapshot¶

Mirror clinvar_build.py in structure, and mirror @duckdb-vs-polars in mechanics: builder in polars, fixed column order so a rebuild is byte-identical (P7). Registered as a pubmind sub-app beside clinvar, so the surface reads just-dna-enrich pubmind build / … publish, matching the clinvar build/clinvar publish pair a reader already knows.

Input is the ANNOVAR-redistributed hg38_pubmind_db.txt.gz, downloaded by --download or supplied by --table, exactly as clinvar build takes --vcf or --download. Output is a snapshot directory: one parquet under data/, plus release.json, laid out through locations like every other snapshot (@snapshot-layout-locations).

Schema. _empty_schema()'s split is the model — resolver-link columns first, annotation after — except that here no column is a resolver link, which is the point above made structural:

Column	From	Note
`chrom`, `start`, `ref`, `alt`	`Chr/Start/Ref/Alt`	the join key; VCF-style already, `start` is the 1-based POS (`@start-1based`)
`pvid`	`PVID`	their record id, not a variant id — see the fan-out rule below
`pubmind_sig` → `clin_sig`	mapped `pathogenicity_sum`	normalized into `VALID_CLIN_SIG`. Renamed 2026-08-28: the ClinVar snapshot already uses the unprefixed name, so the source is the file and a prefix restates it — and one column vocabulary across every snapshot is what lets an N-authority check read them all without per-source mapping
`pubmind_sig_raw` → `clin_sig_raw`	`pathogenicity_sum` verbatim	so the mapping stays auditable and nothing is lost, exactly as `clin_sig_raw` does for CLNSIG
`pathogenicity_score`	`paper_level_pathogenicity_score`	nullable, and null means not computed, never 0.0
`confidence`	`confidence`	0–3, the `review_stars` analogue
`derivation`	computed	`direct` for a `1/1` row, `codon` for a decomposed triplet, `indel` for a length-changing one

Superseded 2026-08-28 (PROPOSAL_0_7 § RM134): one normalizer, not a second map — and it needs two fixes first. Treating _CLIN_SIG_MAP as a precedent to copy gives two maps for one vocabulary, and since the three-way check's whole output is a comparison of two normalizations, any drift makes it report a disagreement between our own mappings. But reusing _normalize_clin_sig unchanged is wrong today: its keys are underscored and PubMind's tokens are spaced, so Uncertain significance → other and Conflicting → other — measured, both concepts ClinVar normalizes correctly, and Uncertain is 32 of the 51 disagreements counted below. Fix: a whitespace→underscore pre-step (a no-op on ClinVar's tokens) plus a bare conflicting key.

_CLIN_SIG_MAP is the precedent for the mapping and its default: a token with no module axis becomes other rather than being dropped. The six PubMind values map cleanly — Pathogenic → pathogenic, Pathogenic/Likely pathogenic → pathogenic, Benign → benign, Benign/Likely benign → benign, Uncertain significance → uncertain_significance, Conflicting → conflicting — and the two composite tokens are the ones worth arguing about, since collapsing P/LP to pathogenic loses the distinction our own vocabulary can carry. Keep both: pubmind_sig_raw holds the composite and pubmind_sig holds the mapped call, and no consumer has to re-parse a slash.

Superseded 2026-08-28 by PROPOSAL_0_7 § RM134, in three places. Kept because the reasoning above is why the decision went the other way.

One normalizer, not a second map. The paragraph proposes a hand-written PubMind map citing _CLIN_SIG_MAP as precedent. Two maps for one vocabulary drift, and since the check's entire output is a comparison of two normalizations, a drift would make it report discordant on its own mapping rather than on the authorities. The shipped form is one shared clin_sig normalizer both sources call.

Benign/Likely benign → likely_benign, not benign. Under the shared normalizer a composite splits on / and the severity order picks the answer, so B/LB reaches likely_benign — the same answer ClinVar's Benign/Likely_benign already gets, which is the point of sharing. P/LP is unaffected and still reaches pathogenic. This is a deliberate result, not an unimplemented line: do not "restore" benign.

The columns are clin_sig/clin_sig_raw, not pubmind_sig/pubmind_sig_raw. The source is the file, so naming a column after it restates the filename — and one column vocabulary across every snapshot is what lets an N-authority check work with no per-source mapping at all.

Normalization is where the measured defects get handled, and every drop is counted. Silent truncation reads as full coverage, so each of these lands as a number in release.json (@dont-discard-computed):

decompose a codon triplet differing at exactly one base onto that base's position (160,090 rows in the 2026-08-24 file), stamped derivation=codon;
drop triplets needing ≥2 simultaneous substitutions (439,383) and longer MNV blocks (5) — they assert a protein change, not a genotypable position;
drop Ref == Alt rows (523);
keep length-changing rows (20,131) but stamp them derivation=indel, because left-normalization is unverified and a consumer needs to be able to exclude them without re-deriving why;
keep every PVID on a contested coordinate as its own row. 68,744 coordinates carry more than one, worst case 35, and their verdicts disagree. Collapsing them would mean choosing a winner by an ordering nobody defined — mode() over an unsorted group is the exact shape the ordering rule bans. The multiplicity is a finding, and section B is where it surfaces.

release.json carries the provenance and the aggregate. Source URL, the file's sha256, its ETag and Last-Modified (all three available, all three recorded, so an upstream revision becomes a finding rather than a silent change of answer), the PubMind-DB version, our builder version, and the drop counts above. _merge_release_block is the existing shape for adding a block without clobbering a sibling snapshot's (@snapshot-accumulates).

pubmind publish is refused, and that is the design, not an omission. PharmVar is the standing precedent (@gated-source-caches): every gated source gets a cache, and PharmVar's is deliberately unpublishable, because a bulk file arriving under terms we cannot establish is not a file we may pass on. An unestablished permission is not a permission. So the command exists and refuses with the reason, rather than not existing — a missing command reads as an oversight somebody will helpfully add.

B. The cross-check: module ↔ ClinVar ↔ PubMind¶

This runs beside the existing ClinVar clin_sig check rather than replacing it, and it answers a question that check cannot: do two independent authorities agree about this variant, and where they do not, which one is our row standing with?

Superseded 2026-08-28 (PROPOSAL_0_7 § RM134): the outcome table below fails at three authorities. pubmind_only and clinvar_only name the authority inside the vocabulary member, so a third source needs a third member and five need every subset. The cause is that one field carries two axes — concordant is defined as both agree and the authored row agrees with them, with authored_dissents as a sibling. Split into authority_concordance (concordant/discordant/single/none/unchecked) and authored_position (matches_all/matches_some/matches_none/absent/unchecked): five members each at any N, with which authority spoke living in a paired detail table. And nothing resolves a split — five sources ordered E>B>D>C>A with E+A against B/C/D cannot be resolved without a weighting model, which is why authored_position relates to the set rather than to a resolved call. The table below is kept as the reasoning that produced the outcomes, not as the shape that ships.

Per variant, the outcome is a triple, and unknown is withheld rather than negated. Each of the three sides independently answers pathogenic-class / benign-class / uncertain / absent / unreachable, and the combination is classified — Kleene, not withhold-on-any-unknown, because unknown AND false really is false:

Outcome	Meaning
`concordant`	ClinVar and PubMind both spoke and agree, and the authored row agrees with them
`authored_dissents`	the two authorities agree with each other and the module does not
`authorities_differ`	ClinVar and PubMind spoke and disagree — the interesting case, and the one nothing today can report
`pubmind_only`	PubMind has a call where ClinVar has none: literature-derived coverage, the reason to adopt at all
`clinvar_only`	ClinVar has a call where PubMind has none
`neither`	absent from both — evidence of nothing, and it must not be reported as agreement
`unchecked`	a side was unreachable, or nobody asked (`--offline`, no snapshot)

ClinSigConflict.opposed already draws the line the severity turns on: opposed (pathogenic-class vs benign-class) is the finding worth acting on, merely different is worth knowing. Reuse that distinction rather than inventing a second one — a discordance between pathogenic and uncertain_significance is not the same event as one between pathogenic and benign, and our own corpus has 32 of the first and 19 of the second.

Severity is warning-tier in both modes and never escalates under strict. @clinsig-never-escalates is the precedent and it applies with more force here: a disagreement with an LLM's aggregate over the literature is a statement about the extraction's limits at least as often as about the module. 62 % agreement is not a number a gate is built on. authorities_differ in particular is not a defect in the module at all — it is a fact about the field — and reporting it as one would be wrong.

Aggregates land in two places, and only one of them recomputes. Build time stamps corpus-wide concordance into release.json — our own reproduction of their 10.6 % ClinVar-overlap and

80 % concordance claims, against our own denominator and our own normalization, which is the only way those numbers mean anything to us. Enrich time reports per module. The build-time number is computed once and never recomputed by a later pass, because a message embedding a count that runs twice publishes two numbers (@no-rerun-with-counts); the enrich-time check dedupes on message, which is the normal case for a check running on both sides.

Absence is not disagreement, and it is not one thing. A variant missing from PubMind means no paper in the corpus discussed it in text the DistilBERT triage stage kept — not that the literature is silent, and certainly not that the variant is benign. --offline with no snapshot is a third state beside asked-and-failed and asked-and-absent (@unreachable-not-absent), and a check that could not run must not report a zero (@tautology-zero).

C. Drafting from PubMind as an authority¶

draft-panel drafts from ClinVar today. PubMind gets the same treatment through --source pubmind on the existing command rather than a new draft-pubmind — the two produce the same row shapes into the same tables from the same gene argument, and a second command would duplicate the genotype worklist, the placeholder guard, the dedup-against-variants.csv pass and the refusal summary, all of which are the parts that are hard to get right. draft-clinpgx is a separate command because it writes different tables; this does not.

Settled 2026-08-30, while building § C: the section assumes a gene argument works, and the snapshot has no gene column. Every locational thing PubMind publishes is (chrom, start, ref, alt), so --gene BRCA1 needs a gene→locus map — and this workspace deliberately holds none, which is why the compiler's own gene/locus check is chromosome-granular. The map built is ClinVar's own per-record attribution matched at the exact position, taking every position ClinVar records for the gene with no clinical or review filter, because it is a locus universe rather than a selection. A min/max span over those positions was refused: it invents a boundary nobody defined and writes a gene cell that is a false claim wherever two genes overlap. Two consequences, both stated in the drafter and in ENRICHER: --source pubmind needs both snapshots, and a PubMind verdict at a position ClinVar has no record for cannot be reached by gene at all — a class that is not countable, since attributing it to a gene is exactly what there is no map for.

Two smaller corrections from the same build. derivation=indel gets no --include-indels flag: a flag is a permanent name the proposal did not price, and additive means one can be added later if anybody wants those rows. And the drafter does not call check_declared_use — that gate decides whether a fetch may proceed and skips on unknown terms, while nothing here fetches; wiring it in would have made the command skip unconditionally whatever --use said. The reason is reported in the source's own words instead (@acquisition-gate-is-not-a-read-gate).

What changes behind the flag:

--min-confidence is the min_review_stars analogue, defaulting to a floor rather than 0. PubMind's confidence 0–3 counts evidence depth the same way stars count review depth, and a confidence-0 row is a single paragraph in a single paper.
Identity is coordinate-whole or nothing (@identity-whole-or-none). Most PubMind rows carry no rsID at all — the HFE example resolves through one, the BRCA1 example has rsid: "-" — so the provider fills chrom/start/ref/alt together or refuses the row and says which cells it could not establish. Half an identity is worse than none.
A derivation=codon row is drafted with that fact attached, and derivation=indel is not drafted at all by default, since left-normalization is unverified.
Rows append, never mutate (@draft-appends); a <<REPLACE>> placeholder protects genotype as it does for ClinVar (@placeholder-protects-decision); the provider writes its own SourceRow (@write-the-sourcerow), and a test that strips declared_use and asserts the compile refuses is what keeps that row load-bearing.

The self-agreement trap, and the existing answer to it. A module drafted from PubMind and then cross-checked against PubMind agrees with itself, and a check that agrees with itself has told nobody anything. @draft-digest is exactly this problem solved once: the provider hashes the table it wrote projected onto the column its own cross-check later reads, stamps SourceRow.draft_digest, and the check recomputes it and skips when the value has not moved. Add pubmind to DRAFT_PROJECTIONS projected onto clin_sig, and remember the two things that entry warns about — the projection is over raw CSV cells at draft time, when the table is still full of placeholders, and the skip is a conjunction of release and digest, so a matching digest against a newer snapshot is still a real comparison.

The ClinVar half of section B is unaffected by this and is the reason the three-way check is worth more than either two-way: a PubMind-drafted module still gets a genuine independent opinion, from ClinVar.

D. The hint surface¶

Corrected 2026-08-30 while building it: the ANNOVAR channel carries no PMID and no paper count. The section promises "verdict, confidence, paper count and PMIDs"; § A's probe of the real file settled the columns, and they are pvid, clin_sig, clin_sig_raw, pathogenicity_score, confidence and derivation. What the hint surfaces is those, per record. It also answers for a coordinate the caller typed rather than only for a locus an rsID resolved to, which the ClinVar leg does not do — their channel is coordinate-keyed and most of its rows carry no rs-number, so the rsID-only shape would leave the leg unreachable for most of the corpus.

Unchanged from the earlier draft and still the cheapest thing here: hint-variant surfaces the verdict, confidence, paper count and PMIDs beside a cell an author is about to fill. It may not pre-fill clin_sig (@hint-redundancy-bearing) — the cell section B cross-examines cannot be filled by the source doing the cross-examining, which is the same defect @draft-digest handles for the drafting path, one layer up and without a digest to rescue it.

The gate, and what lifts it¶

Sections A and D are buildable now. Sections B and C acquire and carry PubMind values, and both turn on one unanswered question:

The ANNOVAR-distributed table publishes no data terms. LICENSE.md covers the software (academic, non-commercial, CHOP tech-transfer for anything else); the paper is CC BY-NC-ND; the table itself says nothing. Unknown is not permissive (@no-named-licence). The unblock action is to ask — WGLab and CHOP's Office of Technology Transfer, in writing, whether the ANNOVAR-shipped subset may be redistributed and on what terms. Nobody else can answer it, and it will not resolve itself by the file continuing to download without a key.
RM27 still owes the redistribution axis (@redistribution-ungated). redistribution is recorded and not gated, so there is no machinery that would refuse to publish a module carrying bytes we may not pass on. That is a gate on publishing such a module, not on building the snapshot or running the check locally — worth separating, because conflating them is what stalled this whole area in the earlier draft.

PUBMIND_TERMS goes in licensing.py beside the others, recording what is establishable and None for what is not — license=None, license_url pointing at CHOP's LICENSE.md, commercial_use=None, redistribution=None, attribution to CHOP and the paper. The nulls are load-bearing: licensing.py defines null as "the terms could not be established", which is a weaker and more honest statement than either permission or refusal, and it is the state we are actually in.

Client mechanics follow the existing contract without exception: retry then translate, both legs, into a PubMindUnavailable subclass rather than a flat error (@client-exception-contract); the cache path resolves through locations; refresh merges, never clobbers (@sidecar-authoritative). And never persist a URL their API hands back — website_result_url and website_detail_url both currently return http://localhost:5016/..., a deployment leak on their side. Derive links from https://pubmind.wglab.org/ ourselves.

Tests¶

Real fixtures, real paths, values computed at runtime — and specifically no row counts copied out of this document into an assertion. The counts here are a dated reading of a file that will move; a test pinning them fails on the next ANNOVAR release for no reason. Assert the relationships instead:

codon decomposition as a property — a synthetic 3/3 row differing at one base decomposes to the right position and allele; one differing at two is excluded; the exclusion count in release.json plus the kept count equals the input row count, an equality over a walked set rather than a floor (@registry-completeness).
the VALID_CLIN_SIG mapping is total — every distinct pathogenicity_sum in the fixture maps to a vocabulary member, and clin_sig_raw (written here as pubmind_sig_raw) round-trips the composite tokens verbatim.
rebuild is byte-identical from the same input, which is what the fixed column order buys.
all seven concordance outcomes are reachable, each from a constructed three-way case, and authorities_differ is warning-tier under strict — the assertion that would have caught an escalation.
neither never reports as agreement, and unchecked under --offline reports no zero (@tautology-zero).
the multi-PVID case reports every PVID, using the real chr6:26092913 shape: eight records, four distinct verdicts, and the one pairing rs1800562 with TMPRSS6 flagged by the gene/locus check.
the draft digest skips on an unmoved clin_sig and fires once the cell is edited, and does not skip against a newer release with a matching digest.
a PubMind-drafted module still gets its ClinVar opinion — the check that proves the three-way design earns its keep.
PUBMIND_TERMS produces a SourceRow, stripping declared_use makes the compile refuse, and pubmind publish refuses with its reason named.

Open questions¶

~~Are the indel rows left-normalized?~~ Answered 2026-09-27 (RM273): no. 5,158 of 20,006 single-base-anchored indel rows provably shift left, read from the row alone; the rest are undecided without the reference. pubmind_draft states the count for the rows it withholds.
What are the ANNOVAR-distributed table's data terms? Only CHOP can answer, and the answer decides whether sections B and C are ever shippable.

Searched 2026-09-27, still unanswered in writing. No page names terms for hg38_pubmind_db.txt.gz. WGLab's README and pubmind.wglab.org send the full PubMind-DB (references, reasoning, evidence) to LICENSE.md and a CHOP contact, and list the ANNOVAR file as the other tier with no licence pointer. LICENSE.md licenses "this software and associated documentation files" (academic, non-commercial) and never says data. The bioRxiv preprint's Data Availability covers only PubMed/PMC, and it calls PubMind "open-source", which contradicts LICENSE.md. ANNOVAR's EULA covers its software, and its database table flags gated sources (CADD, SPIDEX) but not pubmind_db. The published Nature Communications version (s41467-026-76834-4, itself CC BY-NC-ND 4.0) was then read directly. Its Data availability says only "The PubMind-DB generated in this study can be accessed here: https://pubmind.wglab.org/", with no terms. Its code archive (Zenodo 10.5281/zenodo.20632115, a 26 MB PubMind-v0.1.0.zip, typed software) is labelled CC BY 4.0, which contradicts LICENSE.md's academic, non-commercial grant for the same software. A contradiction between two of the owner's own statements is not terms either. So None on every axis stands. The written question goes to CHOP's Office of Technology Transfer (techtransfer@chop.edu, named in LICENSE.md) or the WGLab/PubMind GitHub repo. Asked 2026-09-27 as WGLab/PubMind#5; the answer goes here. - Does the table get a stable release cadence? One snapshot dated 2026-08-24 is not a cadence, and the refresh design depends on it. - Is the PVID stable across releases? If it is not, then nothing about a PubMind reference is citable, and the pass should record coordinates rather than PVIDs. - ~~Is derivation the right name?~~ Closed 2026-08-28. P3 makes a published name permanent because outside consumers key on it, and pubmind publish is refused — so this column never leaves the machine that built it, is not in an artifact, not attested in artifact.files, not a manifest field and not joinable, and renaming it costs a rebuild the same command performs. The P5 audit belongs on the names that really are one-way doors: --min-confidence, and the DRAFT_PROJECTIONS key pubmind, which is matched against SourceRow.source and so reaches the authored, published sources.csv. - Should build-time concordance be computed over the whole table or only over rows a module could plausibly reach? The whole table reproduces their published claim; a reachable subset is the number an author actually cares about. They are different denominators and only one can be the headline.