MOD gene descriptions as review input

Bottom line: model organism databases (MODs) publish hand-written gene
descriptions, for example WormBase's paragraph for
rab-11.1. This project
fetches them for every worm, yeast and fly gene we review, saves each as a
deep-research source, scores it against our review, and acts on what it finds.

	Worm (WormBase)	Yeast (SGD)	Fly (FlyBase)
Genes reviewed	432	890	539
With a curated MOD description	374 (87%)	890 (100%)	345 (64%)
Median MOD text length (words)	75	33	37
MOD text correctness, 1-5 (UniProt)	4.2 (4.5)	4.8 (4.6)	4.7 (4.6)
Description-text defects recorded (MOD + UniProt)	100 (73 + 27)	89 (44 + 45)	73 (39 + 34)
Reviews edited in the follow-up	about 60	249	124
NEW annotation rows	0	2	1

What we learned:

• The MOD text is a recall checklist, not an authority. Our review descriptions
  are usually more mechanistic and more current. The MOD text still names roles,
  phenotypes and regulators our reviews had missed. fshr-1's synaptic role is the
  clearest case: the worm review never mentioned it.
• SGD and FlyBase are the most accurate sources, WormBase the least. SGD and
  FlyBase score above UniProt on correctness; WormBase scores below it. UniProt is
  often the more informative source, especially for fly.
• Most errors are stale, transferred or copied text, in both MOD and UniProt
  descriptions. The 262 recorded defects split 156 MOD and 106 UniProt. They are the
  WRONG_STATEMENT, OUTDATED_STATEMENT and MISSING_STATEMENT defects (plus one worm
  WRONG_REFERENCE) on the cited MOD and UniProt files, counted per resource across
  genes/worm, genes/yeast and genes/DROME. Other source defects in those folders,
  mostly on GO annotations, are not included. Typical
  cases are a claim later overturned (FlyBase says amon cleaves Slit; Tolkin is the
  Slit protease), text carried over from another species (UniProt gives fly SCAP the
  mammalian sterol-sensing mechanism) and a copy error (UniProt names Gurken, another
  ligand, as the receptor for Spitz). Each defect quotes the offending sentence, and
  just export-source-defects turns them into per-resource reports. Filing those
  reports is a manual step that has not been done yet.
• Gaps rarely become annotations. 794 gap and discrepancy findings produced 3
  NEW rows. Most gaps are mutant phenotypes, which show that the gene is needed,
  not that it does the work, so they become description sentences instead.
• The Alliance automated text is the weakest source. It is generated from the
  GO annotations under review, so it is never cited as evidence, and it scores lowest
  on every species. Several automated texts have also lost content between releases,
  listed in alliance-automated-regressions.
• No other MOD adds anything. MGI's note is a phenotype summary, RGD's is
  automated text, and ZFIN and Xenbase have none.

Decision (2026-10-30): treat the curated description as a deep-research
source, the same way as Affinage. It is written verbatim to
<gene>-deep-research-<mod>.md (wormbase, sgd, flybase) by
just mod-description-deep-research. It is cited as a file: reference, and the
reviewer's judgment goes in references[].reference_review. The source is sparse,
uncited, undated and only partly trusted, but no new data structures or processes
are needed for it. See Deep-research integration.

The work began with repairs. A fetcher and sidecar format existed
(ai-gene-review fetch-descriptions, <gene>-descriptions.yaml). They had only been
run on 118 yeast genes, nothing in fetch-gene or the reviewer skills read the
files, and the fetcher had stopped working. The Alliance API had changed its response
shape, and the worm cross-reference was parsed wrongly. Both are now fixed.

Data and method

• Fetch: uv run ai-gene-review fetch-descriptions-bulk worm. This writes
  genes/worm/<gene>/<gene>-descriptions.yaml with UniProt FUNCTION, Alliance
  curated (Alliance_Imported, the MOD's own hand-written text), Alliance
  automated, and NCBI RefSeq summaries.
• Alliance now returns synopses as gene.relatedNotes with
  noteType.name set to MOD_provided_gene_description (curated) or
  automated_gene_description. The old geneSynopsis keys are gone, so the
  previous code returned nothing without raising an error.
• Compare: compare_descriptions.py
  writes worm-comparison.yaml, which
  lists, per gene, the WormBase content words that do not appear in our
  description. This is a lexical pointer for choosing genes to read, not a
  quality score.

Measure (worm)	Value
Genes fetched	432
With curated WormBase description	374
Median words, WormBase curated	75
Median words, our description	143 (139 before the follow-up)
Median share of WormBase content words found in ours	0.37 (0.35 before the follow-up)

Acting on the findings (below) raised both numbers: the descriptions gained
sentences for verified gaps, so more of the WormBase content now appears in them.

Hand comparison (16 genes)

There were eight lowest-overlap genes (cep-1, let-60, fshr-1, lgg-2, gst-4,
mpk-1, ced-4, atg-18) and eight random ones (skn-1, sup-17, fasn-1,
hsp-16.41, rsks-1, lag-2, gpr-1, lipl-4). Low lexical overlap mostly reflects
different emphasis, not error.

Gaps in our reviews that WormBase points to

• fshr-1: WormBase describes a role in acetylcholine release at neuromuscular
  synapses (a large-scale synaptic screen; FLP-1 and NLP-12 are possible
  ligands). Our review is entirely about intestinal innate immunity and stress,
  and neither the review YAML nor the notes mention synapses or acetylcholine.
  This needs a follow-up literature check.
• gst-4: WormBase describes a putative prostaglandin D synthase needed for
  sperm guidance towards an oocyte PUFA signal. Our description frames GST-4
  only as a SKN-1-induced detoxifying GST. The review raises the prostaglandin
  question only in a suggested question.
• cep-1: WormBase mentions meiotic chromosome segregation and hypoxia
  sensitivity. The review YAML covers them, but the description does not.
• let-60 and mpk-1: WormBase lists phenotypes our short descriptions omit (male
  spicule, sex myoblast migration, chemotaxis, pachytene progression,
  NOS-3/TRA-1 regulation in the germline). For let-60 none of these appear
  anywhere in the review. For mpk-1 they appear in the annotations but not in
  the description.
• lag-2 and skn-1: WormBase gives detailed expression patterns and their
  regulation (LET-23/SUR-2 regulation of lag-2 in the vulval precursor cells),
  which our descriptions largely leave out.

Where WormBase is stale or weaker

• ced-4 is still a "novel protein" with sequestration by CED-9 to membranes. It
  has no apoptosome and no Apaf-1 homology.
• lgg-2 is said to have "no obvious function" and a speculative MAP1
  microtubule role. Our review describes its established autophagosome
  maturation role through VPS-39/HOPS.
• rsks-1 is a "putative" S6K, and hsp-16.41 is a "passive ligand". The
  WormBase texts were written years ago and, like UniProt FUNCTION text for
  yeast SAN1 in the earlier pilot, are not updated as the field moves.
• Several WormBase texts contain formatting damage (words run together, such as
  "PUFA-basedsignal"), which suggests they were not cleaned when imported.

Deep-research integration

• Emit: just mod-description-deep-research <org> [genes...]
  (src/ai_gene_review/etl/mod_description_deep_research.py). The file holds
  the curated note exactly as served, including WormBase's run-together words,
  with frontmatter recording provider, authorship (MOD curators),
  source_url, MOD gene id, UniProt accession, fetch date, and
  citations_provided: false / authoring_date_provided: false. As with the
  Affinage emitter, it contains no AIGR judgment.
• Only the curated note is used. The Alliance automated description is
  generated from GO annotations, so citing it would cite the annotations under
  review as evidence for themselves.
• Wrong-gene check: the MOD id comes from the UniProt cross-reference.
  Writing is refused (unless --force) when the symbol or systematic name of
  the Alliance record matches neither the folder nor any UniProt gene name.
  Human is out of scope.
• Reviewer guidance (review and annotation-reviewer skills): run the
  emitter for every non-human review, check each claim, and record a verdict in
  reference_review. A claim found only in the MOD text is a lead for a
  literature search.
• The DEEP_RESEARCH publication type now covers external narrative sources
  in general, not only AI reports. The frontmatter records authorship.
• <gene>-descriptions.yaml stays as the input for the description comparison
  (compare_descriptions.py). It is
  not a review source, because it stores the source text together with our AI's
  rubric scores.

Worked example: fshr-1. The WormBase text led to Buckley et al. 2024
(PMID:39561202). That paper shows that intestinal FSHR-1 is necessary and
sufficient for acetylcholine release at the neuromuscular junction, with
GSA-1/Gas and ACY-1 acting downstream and the glycoprotein hormone subunits
GPLA-1 and GPLB-1 upstream. The review now has a NEW
GO:1900075 positive regulation of neuromuscular synaptic transmission (IMP)
row, a corrected description (which had wrongly said FSHR-1 had been
"repurposed" away from hormone signaling), stronger support for the IBA
adenylate-cyclase row, and updated open questions. The WormBase reference's
reference_review records which claims were traced (synaptic role) and which
were judged stale (FLP-1 and NLP-12 as candidate ligands).

Description reviews (all 432 worm genes)

The review of each text is stored in the gene's <gene>-descriptions.yaml,
separate from both the gene review and the deep-research file. Every source
(UniProt FUNCTION, WormBase curated, Alliance automated, RefSeq) gets a 1–5
score on correctness, utility, compactness, completeness and specificity, with a
note. The file also gets an overall_concordance score across sources and
findings (gap, discrepancy, insight). All 432 worm files are COMPLETE
(ai-gene-review descriptions-status worm). Scores are AI judgments made
against each gene's review and the literature it cites. 16 genes were scored by
hand, and parallel agents scored the other 416, calibrated on those 16. Their
output was schema-checked, and several of their
discrepancy calls (sem-5, spg-7, met-2) were spot-checked against the reviews.

Source (genes)	Correctness	Utility	Compactness	Completeness	Specificity
UniProt FUNCTION (315)	4.5	3.7	3.7	3.5	3.7
WormBase curated (374)	4.2	3.4	3.6	3.1	3.3
Alliance automated (432)	3.9	2.7	4.0	2.6	2.7

Concordance across sources: mean 4.25 (5: 187 genes, 4: 177, 3: 57, 2: 10, 1: 1).
RefSeq is left out of the table. It is usually a copy of the Alliance automated
text, but for about a third of genes it comes from an older or newer Alliance
release and was scored separately.

The 313 findings (142 gaps, 105 discrepancies, 66 insights; one lgg-2 discrepancy
was added later) are collected in
worm-findings.yaml, which is a curator
worklist. Recurring patterns:

• Stale WormBase "no obvious function" texts. Twelve WormBase texts score 2
  or less on correctness. Most say "no obvious function in mass RNAi assays"
  for genes with established mutant phenotypes: atg-2, atg-3, atg-5,
  atg-16.2, lgg-2, gst-44 and pef-1. meg-3 is described as a somatic protein,
  but it is the germline P-granule scaffold.
• WormBase or UniProt contradicting the review. Examples: SEM-5 "negatively"
  regulating vulval induction (both sources); spg-7 called the paraplegin
  ortholog (it is AFG3L2; ppgn-1 is SPG7); MET-2 as an H3K36 methyltransferase
  (it is H3K9); SDF-9 as an active phosphatase (it is a pseudophosphatase);
  akt-2 single mutants as dauer-constitutive.
• Mammalian text carried over in UniProt. Examples: "motile cilium" for
  the non-motile worm sensory cilia (dyf and osm genes, ift-20); myelin and
  sebaceous-gland roles for fath-1; SHH signaling for ift-81.
• Automated texts that lead with removed or over-annotated terms. Examples:
  slicer (RNA endonuclease) activity for WAGO Argonautes that lack the catalytic
  residues (wago-1, wago-4, hrde-1); kinase activity for the vps-15 pseudokinase;
  protein tyrosine phosphatase rather than PIP3 phosphatase for daf-18.
• Regressions in the current automated text. For ced-5, che-2, daf-15,
  nud-1, ppgn-1 and nsy-1, the current Alliance text has lost function
  sentences (and, for nsy-1, its core MAP3K activity) that the older RefSeq copy
  still carries. This is worth reporting to the Alliance.
• Gaps in our reviews. Some MOD claims appear nowhere in the review and
  are leads for follow-up. Examples: bbs-1/2/5/7/8/9 dense-core-vesicle
  secretion phenotypes, ddb-1 in LET-23 vulval signaling, lys-1 regulators,
  gsk-3 axon guidance, and the let-60 phenotypes.

Follow-up: acting on the findings

The 244 gap and discrepancy findings (216 genes) were each checked against the
literature, and acted on under the CLAUDE.md curation rules. Parallel agents did
the work; each edited gene was validated and given a history record. Batch
results were spot-checked (pek-1, lrx-1, sem-5).

• Upstream defects: 100 errors in description text across 93 genes (73
  WormBase, 27 UniProt) are now recorded as source_defects on the cited
  reference: 31 WRONG_STATEMENT, 28 OUTDATED_STATEMENT, 40 MISSING_STATEMENT and 1
  WRONG_REFERENCE. Of these, 28 are CONFIRMED by a quoted paper and the rest are
  PROBABLE. just export-source-defects writes them to
  reports/source-defects/WORMBASE.md and UNIPROT.md for filing. Examples:
• UniProt says IRE-1 retains unspliced mRNAs, but its cited paper says the opposite.
• UniProt names nlp-29 as its own receptor; the receptor is NPR-12.
• UniProt's sma-9 FUNCTION text cites an unrelated paper.
• WormBase says the GPR-2 GoLoco motif is a GEF feature.
• WormBase says SOD-2 promotes lifespan, but sod-2 mutants are long-lived.
• WormBase calls ASNA-1 a membrane transporter, but it is the GET3 targeting ATPase.
• Our reviews:
• About 60 descriptions gained sentences for verified gaps. Examples include
  the let-60 phenotypes, the BBS dense-core-vesicle secretion phenotypes and
  the sem-5 PLC-gamma/PI3K outputs.
• No NEW annotation rows were added. Every gap was an indirect or
  pleiotropic phenotype that does not meet the NEW bar.
• Four reviews were corrected where they, not the source, were wrong:
  • pek-1: pek-1;atf-6 silencing has no phenotype (PMID:16184190).
  • lrx-1: its own analysis wrongly rejected the LDL-A repeats that Pfam,
    SMART and CDD detect.
  • lys-1: it is a screened candidate, not an established DAF-16 target.
  • gcy-8: GCY-18 also contributes to the role.
• Unverified: about 30 claims could not be traced to a paper (for example
  gcy-23 lifespan and rps-23 vulval morphogenesis). They are noted in the
  reference's review_notes, with no annotation change.
• Alliance: nine automated descriptions lost content that the RefSeq copy
  of an earlier release still has; see
  alliance-automated-regressions.md.
• Open: the apa-2 review removes "GTPase activating protein binding" because
  RME-6 is a GEF, but RME-6 also has a RasGAP-related domain (PMID:15895077).
  The lrx-1 suggested questions look as if they belong to another gene. Both need
  a human look.

Yeast (SGD)

• fetch-descriptions-bulk yeast refreshed all 890 sidecars, and
  mod-description-deep-research yeast wrote 890 -deep-research-sgd.md
  files. Every yeast gene has an SGD-written description.
• Re-fetching kept the 2026-03 pilot scores wherever a source text was
  unchanged. Where SGD or the Alliance had changed a text since then, that
  source's score and the cross-source review were dropped as stale: 17 files
  stay COMPLETE, 101 are IN_PROGRESS and 772 are new (STUB).
• SGD texts are much shorter than WormBase's (median 33 words, against 75 for
  WormBase and 111 for our yeast reviews). On lexical overlap they are closer to
  ours: a median 0.42 of SGD content words appeared in our descriptions (0.44 after
  the follow-up below), against
  0.37 for worm (0.35 before the worm follow-up). See
  yeast-comparison.yaml.
• Guard against automated text filed as curated (looks_automated): it
  rejects all 12 RGD notes sampled, and flags none of the 902 WormBase and SGD
  curated texts on disk. A first version that matched on the opening alone
  wrongly rejected SGD's CAF120 ("Part of the CCR4-NOT ... complex; involved
  in ..."), and was fixed before any file used it.
• All 890 yeast files are now COMPLETE. The 873 incomplete files were scored by
  16 parallel agents, calibrated on the worm examples and on the yeast pilot. Sources
  that kept a pilot review were not rescored, and each batch was schema-checked.
  Spot-checked findings (GTT2, IRC24, RIB7) hold against the source texts.

Source (yeast genes)	Correctness	Utility	Compactness	Completeness	Specificity
SGD curated (890)	4.8	4.1	4.3	4.0	4.0
UniProt FUNCTION (798)	4.6	3.7	4.0	3.5	3.8
Alliance automated (890)	4.4	2.9	3.8	3.1	2.9

The counts include one protein twice: genes/yeast/ERG19 and genes/yeast/MVD1
are the same entry (P32377; MVD1 is the SGD standard name and ERG19 a synonym). Merging them is
issue 5837. Two genes, HKR1 and
MSB2, have no gene review, so their concordance scores compare the sources with
each other only.

The test the worm run could not make: SGD's maintained text is the best source.
It scores above UniProt on every rubric, whereas WormBase scored below it (4.2
against 4.5). Only two SGD texts score 2 or less on correctness:
- ALG11: Man5 rather than Man3-to-Man5.
- RIB7: called a deaminase; it is the reductase.

Twelve WormBase texts did. Concordance is a mean 4.48 (worm 4.25). The 354
findings (182 gaps, 163 discrepancies, 9 insights) are in
yeast-findings.yaml. Most discrepancies
involve the automated text (108 of 163), which picks GO terms the reviews
reject. Examples:
- kinase activity for the pseudo- or non-kinases KDX1 and CAF120;
- THI3 decarboxylase activity.

SGD text is involved in 61 discrepancies and UniProt in 51 (one discrepancy can
involve several sources). The UniProt cases are often dated or transferred text,
for example:
- ACP1: a Complex I subunit, in a species that has no Complex I.
- ACH1 and VIP1: claims transferred from mammalian or S. pombe homologues.
- ERG12, ERG19 (also filed as MVD1) and ERG20: mevalonate pathway steps placed
in the vacuole; the reviews have these enzymes in the cytosol.

Yeast has only 9 insight findings against 66 for worm, while gaps and
discrepancies rose. That may be because SGD's shorter text yields fewer side
observations, or because the yeast agents used the category more narrowly; the
data cannot tell the two apart.
SGD has two self-referencing copy errors: GTT2 "overlaps with Gtt2p", and IRC24
is "a putative paralog of IRC24".
- Follow-up on the yeast findings. Eight parallel agents worked through the 345 gap
and discrepancy findings (316 genes) under the same rules as for worm. Each edited
gene was validated and given a history record.
- 89 description-text source_defects across 81 genes: 44 SGD and 45 UniProt (one in a UniProt SUBUNIT comment);
49 WRONG_STATEMENT, 25 OUTDATED_STATEMENT, 15 MISSING_STATEMENT; 38 of them
CONFIRMED by a quoted paper.
- 249 reviews edited, mostly with a description sentence for a verified gap.
- Two NEW rows, both checked against the NEW bar:
- MIH1 GO:1905279 regulation of retrograde transport, endosome to Golgi:
purified Mih1 dephosphorylates Vps26.
- HOG1 GO:0071243 cellular response to arsenic-containing substance: Hog1
senses arsenic through conserved cysteines.
- Three action changes. RAS2's duplicate IEA bud-neck row is now ACCEPT, matching
the accepted IGI row. DCI1's two dienoyl-CoA isomerase rows are now UNDECIDED,
because purified Dci1 shows no activity (PMID:19830908), against an earlier
crude-extract result.
- As for worm, lexical overlap rose (recall 0.42 to 0.44, median 111 to 115 words).
- Examples of SGD defects:
- RIB7 called a deaminase;
- ALG11's Man5 substrate;
- PFK1 ribosome association, which belongs to Pfk2;
- KDX1 called a kinase;
- GUT1 regulators swapped.
- Examples of UniProt defects:
- THI3 decarboxylase activity that direct assays do not find;
- BTS1 contradicted by its own kinetic data;
- CPT1 given Ept1's substrate;
- the vacuolar mevalonate module.
- correctness convention, now in the review skill: DISPUTED if any claim is
wrong, otherwise UNVERIFIED if any claim could not be traced, otherwise VERIFIED.
53 earlier worm and yeast MOD and UniProt references were brought into line with it.

Fly (FlyBase)

• mod-description-deep-research DROME wrote 345 -deep-research-flybase.md files
  for the 539 fly reviews; 189 genes have no FlyBase-written description. Every gene
  has a descriptions sidecar.
• Two fetcher fixes came out of the run:
• Unreviewed UniProt entries often lack the DR FlyBase line and carry only
  DR AGR; FB:FBgn.... Both fetchers now fall back to the AGR line for any MOD,
  which recovered 8 genes.
• FlyBase writes Greek letters as characters ("α-Man-Ia"), so the identity check
  now spells them out and ignores case.

Two genes are still not fetched. CG4592's UniProt entry gives only a VectorBase id,
and the Alliance API rejects GatB's FBgn0039153 with HTTP 400.
- FlyBase texts are a median 37 words (SGD 33, WormBase 75), and a median 0.50 of their
content words appear in our descriptions, the highest of the three MODs. See
fly-comparison.yaml.
- 531 files are COMPLETE, scored by 8 parallel agents with the yeast brief adapted for
fly. Eight genes have no description from any source and were left unscored.

Source (fly genes)	Correctness	Utility	Compactness	Completeness	Specificity
FlyBase curated (345)	4.7	3.7	4.8	3.4	3.5
UniProt FUNCTION (319)	4.6	3.8	3.9	3.6	3.8
Alliance automated (529)	4.3	2.9	3.7	3.0	2.8

As with SGD, the curated MOD text is the most accurate source. Unlike SGD, it is not
the most informative: UniProt scores higher on utility, completeness and specificity,
and FlyBase's lead is in compactness. Concordance is a mean 4.34. The 211 findings
(116 discrepancies, 89 gaps, 5 insights, 1 consensus) are in
fly-findings.yaml. The automated text is
involved in 74 of the 116 discrepancies, UniProt in 38 and FlyBase in 37.
- Follow-up on the fly findings. Eight agents worked through the 205 gap and
discrepancy findings (187 genes) under the yeast rules.
- 73 description-text source_defects across 64 genes: 39 FlyBase and 34 UniProt;
34 WRONG_STATEMENT, 13 OUTDATED_STATEMENT, 26 MISSING_STATEMENT; 25 of them
CONFIRMED by a quoted paper.
- 124 reviews edited, mostly with a description sentence or a reference_review.
- One NEW row: Tango5 GO:0000045 autophagosome assembly (IMP, PMID:34019822).
Tango5 is the fly VMP1, the ER scramblase that does part of the work, and human
VMP1 carries the term.
- Lexical overlap did not move (median recall 0.50; review median 125 to 128 words).
- Examples of FlyBase defects:
- Pxd called a catalase-like, glutathione-oxidising enzyme; it is the chorion
peroxidase that makes di-tyrosine cross-links.
- Br140 and Ing5 called histone acetyltransferases; Enok is the enzyme.
- Lis-1 called a PAF-AH subunit; fly Lis-1 does not bind the fly alpha homolog.
- amon said to cleave Slit; Tolkin is the Slit protease.
- SREBP said to respond to cholesterol; fly SREBP responds to phosphatidylethanolamine.
- UniProt defects are mostly mammalian text carried over to the fly entry:
- smo and SCAP: cilium and sterol mechanisms.
- cact and IKKbeta: IKK as the Cactus kinase; Pelle phosphorylates Cactus.
- spi: Gurken named as the EGF receptor.
- Ccs: described as a dismutase; it is the copper chaperone for Sod1.
- borr: copied from the Australin paper and contradicted by the borr paper it cites.
- Alliance automated text: two losses (Mtpalpha, IKKepsilon) and two doubtful human
orthologs (ey given PAX4, tin given NKX2-2), in
alliance-automated-regressions.

Mouse (MGI): not used

MGI's MOD_provided_gene_description is a summary of mutant phenotypes, not a
description of what the gene does. In a random sample of 25 of the 221 mouse
genes, 22 have text and all 22 describe allele phenotypes:

• mouse/Src: "Homozygotes for a targeted null mutation exhibit growth retardation, failure of
  tooth eruption, osteopetrosis ...".
• Ccne1: "Mice homozygous for disruptions in this gene display no abnormal phenotype."

Cd79a and Cd79b, outside the sample, share the same sentence word for word. The sample
and these two notes are in mouse-mgi-sample.yaml,
written by mgi_sample.py. Mutant phenotypes establish
necessity, not participation, so under the NEW bar they almost never support an
annotation, and a worm- or fly-style review would mostly list phenotypes that are
correctly left out. The mouse run was therefore stopped after the sample (2026-11-02),
and mod-description-deep-research now skips MGI, so no -deep-research-mgi.md files
are written. The sample also exposed a latent bug,
now fixed: MGI's own UniProt line carries the prefix (DR MGI; MGI:106913;), so the
descriptions fetcher would have asked the Alliance API for MGI:MGI:106913.

Recommendations

1. Use the curated MOD description as an input, not as an authority. This
   is done through the deep-research file above. Each claim in the curated text
   should be either covered by the review or set aside in reference_review
   (for example, "stale: CED-4 is the Apaf-1-like apoptosome adaptor").
2. Treat a claim that appears only in the MOD text as a literature lead.
   fshr-1 is the model case: the gap is a missed role, not wording.
3. Re-audit existing reviews with the comparison script, starting with the
   lowest-overlap genes. The remaining Alliance MODs add nothing here: MGI's note is a
   phenotype summary, RGD's is automated text, and ZFIN and Xenbase have none.
4. The NCBI step hits rate limits (HTTP 429) on bulk runs. It should honour an
   NCBI_API_KEY, or back off and retry.

Status and next steps

• [x] Fix the Alliance parser and the WormBase gene-id extraction
• [x] Fetch worm descriptions (432 genes) and write the comparison table
• [x] Hand-compare 16 genes
• [x] Decide how to feed the text to reviews: as a deep-research file (2026-10-30)
• [x] Emitter, tests, just recipe and skill guidance
• [x] Write -deep-research-wormbase.md for the worm genes
• [x] Follow up the fshr-1 synaptic role (PMID:39561202) in its review
• [x] Rubric reviews of all description sources for all 432 worm genes, with a findings worklist
• [x] Work through the gap and discrepancy findings in worm-findings.yaml
• [x] Record the contradicted MOD and UniProt statements as source_defects
• [ ] File the WormBase, UniProt and Alliance reports (manual step) and set report_status
• [x] Yeast: fetch descriptions and write SGD deep-research files (890 genes)
• [x] Yeast: score descriptions (890 COMPLETE)
• [x] Yeast: act on gap and discrepancy findings (89 source defects, 249 reviews edited)
• [x] Fly: fetch, score and act on findings (345 FlyBase files, 73 source defects, 124 reviews edited)
• [x] Mouse (MGI): sampled and not used, because the note is a phenotype summary; rat (RGD) has only automated text; zebrafish and frog have none