PseudoCAP IBA-seed review protocol

Brief for reviewing one P. aeruginosa PAO1 gene from
iba_seed_worklist.yaml. Read the repository
CLAUDE.md first; everything below is in addition to it.

Context

PseudoCAP annotations on PAO1 genes are seeds for PAINT IBD node annotations.
An error in a seed is copied as an IBA to every leaf below the node. The
worklist entry for the gene lists each PAINT node it seeds, the node term, the
number of seeds (n_seeds; sole_seed: true means the IBD rests only on this
gene), the leaf IBA count, and the experimental annotations that support the
seeding.

Steps

1. just fetch-gene PSEAE <symbol> (re-run if EBI times out). Check that the
   fetched UniProt accession matches the worklist. If a symbol is ambiguous or
   missing, use the locus tag or the accession as the folder name.
2. Read the UniProt file, the GOA file, and the cached publications/PMID_*.md
   for every cited PMID. Many are abstract-only: use the PubMed MCP
   (mcp__PubMed__get_full_text_article, mcp__PubMed__search_articles) for
   full text where possible. Do not write files into publications/ by hand.
3. No deep-research provider keys are available. Do your own literature research
   and write it into <GENE>-notes.md with inline provenance. Never create a
   -deep-research-<provider>.md file.
4. Write the full review: description, references (with reference_review for
   the PseudoCAP-cited papers), every existing annotation with an action,
   core_functions. Follow the CLAUDE.md rules on IBAs, on not overruling
   curators from incomplete evidence (prefer UNDECIDED to REMOVE for a paper
   you cannot read, unless the abstract itself shows the claim is an indirect
   phenotype), on protein binding, and on NEW.
5. Seed assessment. For each PseudoCAP annotation that seeds an IBD, decide
   whether the paper directly supports that term for this gene, and so whether
   the IBD is sound. The typical failure is a mutant phenotype (lost motility,
   reduced enzyme activity in a lysate, less pigment) annotated as the gene's
   own molecular function or process. Also judge whether the node term suits a
   clade of that size and breadth.
6. A wrong PseudoCAP row gets review.source_defects with
   resource: PSEUDOCAP (see docs/source_defects.md).
7. If that leaves the IBA on this gene unsupported, review the IBA row
   accordingly and add a defect with resource: GO_CENTRAL naming the PTN
   node.
8. Add ## PseudoCAP IBA-seed assessment to the notes with a table:
   PTN node | node term | n_seeds | leaf IBAs | seed annotation (term,
   evidence, PMID) | verdict | rationale. Verdicts: SOUND,
   OVER-SPECIFIC, INDIRECT-PHENOTYPE, WRONG, UNVERIFIABLE.
9. Set confidence by what you actually read. Many seed papers are
   abstract-only. An abstract can show what a paper did; it cannot show what
   the paper left out. Absence from an abstract ("only genetics is
   reported", "no assay is described", "it is a review") is never enough for
   PROBABLE: a review from the laboratory that did the work, or a
   genetics-led abstract, can still report the experiment the curator coded.
   • CONFIRMED: you read the full text, or the decisive fact is primary
     data you checked yourself (the sequence, a live resource query, a
     same-day sibling annotation showing an id slip), or the abstract states
     the decisive fact outright (for example, it names a different organism
     or gene).
   • PROBABLE: the abstract (or another paper) positively states facts
     that make the error likely, but the full text could change the call.
   • SUSPECTED: the call rests on what the abstract does not say, or on
     reasoning that the full text could overturn.
     Write the explanation, reason and reference_review to match: never say
     a paper "cannot support" a claim, or mark it MISCITED, unless you have
     read it.
10. just validate PSEAE <GENE> and
    just backfill-record-sections genes/PSEAE/<GENE>; fix all errors and the
    warnings you reasonably can.
11. History record:
    uv run python scripts/new_history.py --kind gene --organism PSEAE --slug <GENE> --event REVIEW --summary "..." --details "..." --sections existing_annotations,core_functions,references --agent-tool claude-code,
    then just validate-history <path>.
12. Edit only genes/PSEAE/<GENE>/, history/, and publications/ (through
    the tooling). Do not commit.

Report back

Under 300 words: per seeding node, the verdict and a one-line reason; action
counts; validation warnings left; anything you could not verify.