Triage the data-source queue

curation/source_queue.tsv is the ranked list of data sources this corpus
draws on or might adopt. This skill keeps it honest and answers "what should
we add next?" with evidence rather than enthusiasm. Adapted from the
AntibioticMech skill of the same name; the ranking rule and the adoption gate
are theirs, the gaps and licence rules are this repository's.

Read these first

• curation/source_queue.tsv — the queue; docs/SOURCE_QUEUE.md is its legend.
• just source-queue — the checker's verdict and its "next up" line.
• just report — which fields are actually empty, corpus-wide. A source that
  fills an empty column beats one that thickens a full one.
• docs/CURATION.md — the granularity contract, identifier policy and image
  rules; a source outside scope is a rejection, not a low priority.
• conf/sources.yaml — what the pipeline reads today.
• research/ — the verified assessments behind the rows.

The ranking rule

Rank by what the corpus cannot currently assert, not by how well known the
source is. In order:

1. Does it close a stated gap? Check just report. Today every record
   has components and an image; protein_examples exist on one record,
   causal_graphs on three, physical_properties are thin, and there are
   four records against a backlog of dozens (#12). A source that supplies
   records or components with evidence outranks one that supplies a
   second image.

2. Can we redistribute it? This is a hard gate. The corpus is CC0 and
   hosts image copies under pages/. SEED (copy into records or pages) is
   allowed only under CC0_OK, ATTRIBUTION or SHARE_ALIKE; NONCOMMERCIAL
   is LINK_ONLY at best (the Atlas); RESTRICTED is CURATE_ONLY (EcoCyc:
   cite, never copy); UNVERIFIED cannot be adopted. Record what the licence
   page says, not what it would take to make it work.

   The same tension AntibioticMech carries is open here: CC BY images and CC
   BY vocabularies sit inside a CC0 repository. Attribution is preserved per
   item (images.attribution, evidence notes), but a blanket CC0 dedication
   over CC BY content is not something we can grant. Judge candidates against
   the stricter reading.

3. Does every item carry a citable identifier and a taxon? The project
   brief requires both on images and protein examples. item_id and
   taxon_link say whether the source gives them or the curator must supply
   them; a NO/NONE source is a curation project, not an extraction.

4. Bulk or API access over manual. data/ is committed and just qc
   runs offline; a per-item manual pull can never become ADOPTED (the
   checker requires a script), only EVALUATING.

5. Effort, last.

Two sources that close the same gap: adopt one, measure what it added, then
decide about the second (PMC OA vs Cell Image Library for micrographs).

Verifying a source before it moves to ADOPTED

redistribution starts UNVERIFIED and is checked against the source's own
licence page — not a summary, not a memory, not another database's claim.
Record the date in verified_on. The checker refuses an ADOPTED row with
unverified terms, without a script, or absent from conf/sources.yaml.

Also establish, and write into rationale:

• Coverage against this corpus: which of our records would gain what.
  "6/6 test structures" (UniProt SL) is an answer; "thousands of images" is not.
• Identifier joinability — GO CC ids, NCBI taxids, UniProt accessions,
  DOIs. Name-only joining (Cell Image Library organisms) is curation work.
• Machine-readable licence per item, when the licence varies per item
  (Commons, CIL, BioImage Archive) — where in the API it is, quoted.
• Known traps this corpus has hit: pattern-valid identifiers that do not
  resolve (#2, #4, #34); tracking parameters in API-returned URLs (#34); a
  rendering presented as a micrograph (#35); a retired API still documented
  elsewhere (PMC OA service, Feb 2026); a site that resolves but refuses
  connections (ETDB-Caltech).
• Whether an adopted source already closes it. UniProt SL lists member
  proteins per structure; check it before adding a second components source.

A licence that cannot be reached is a result too: record the URLs tried and
what blocked them (SubtiWiki: nothing on site, bundle, Swagger or paper).

Folding in a research report

When a research note lands, do not paste it into the queue. For each source
it covers: add or update the row; move redistribution off UNVERIFIED only
where the note quotes the licence page; put the specific finding in
rationale; keep the note's own "could not verify" as UNVERIFIED. Add a
Sources section to the note if it lacks one (#32).

What this skill does not do

It does not fetch, seed, or adopt. Moving a source to ADOPTED means a
script under scripts/ reads it, conf/sources.yaml lists it, the terms are
verified and dated, and just qc passes — that is a pull request with the
canary discipline in CLAUDE.md (one item end to end, side effects checked,
just qc run bare), not a row edit. Editing the row to say ADOPTED without
that work makes just source-queue fail, which is intended. Merging and
promoting records to REVIEWED stay the owner's calls.

Output

Report: the top three candidates with the gap each closes and what is
unverified about it; anything whose status should change and why; any source
the corpus has outgrown. If the queue is already accurate, say so — a short
honest answer beats a reshuffle.