Ground a claim in the paper that supports it

Every claim-bearing field in this corpus carries evidence, and every evidence
item names a reference. Until this skill existed, nothing read those
references: check_curies.py asks whether a DOI or PMID resolves, never what
it says. That is how the corpus reached 341 evidence items with 3 verbatim
snippets (#133), and how one record cited a review for the opposite of what the
review reports.

scripts/fetch_snippets.py is the tool. This skill is the judgement around it.

The distinction the schema makes

• snippet is a verbatim quotation from the source. Nothing else may go
  here. It is the only field a machine can check, and --verify checks it.
• notes is the curator's paraphrase, gloss, or reason for citing. Anything
  that is not a literal quotation belongs here.

Moving a paraphrase into snippet to satisfy a gate is the worst thing this
skill can lead to. An honest notes with an empty snippet is fine; a
snippet that is not in the paper is a fabricated quotation.

The three modes

just evidence-audit                      # what is readable, corpus-wide
uv run python scripts/fetch_snippets.py --suggest --record data/structures/<cat>/<rec>.yaml
just evidence-verify                     # every snippet occurs in its source

--audit classifies every reference: full_text, abstract, unreadable,
not_literature, or unchecked. It writes reports/evidence_readability.tsv
and prints the one number that matters — claims citing a paper with no reachable
text and no quotation, which nobody can check.

unchecked means no route answered — a network fault, not a fact about the
paper. It is excluded from that count and never fails --verify, because a
question we failed to ask is not an answer. Identities resolved during an outage
are not cached, so the next run retries them.

Two populations are counted and named apart: citations (every field carrying
a reference, including taxon notes, canonical examples, causal-graph edges and
images) and evidence items (entries in an evidence list, the population
#133 counts). Both are checked; only the labels differ.

--suggest prints candidate sentences ranked against the claim. It never
picks. Choosing which sentence supports a claim is the judgement the evidence
field exists to record; a tool that picked would be generating the appearance of
grounding. Read the candidates, read enough of the paper around the one you
choose to know it means what it appears to mean, then paste it verbatim.

--verify is the anti-fabrication gate. Run it before every commit that
touches a snippet. It grades what it finds, because "copied imprecisely" and
"not in the paper" need different responses:

verdict	meaning	action
verbatim	matches exactly, or differs only in whitespace	none
not verbatim	the text is in the paper; the record's copy differs in case, quote or dash style	re-copy the exact characters
ABSENT	not in the source at any tier	the quotation is fabricated or reworded — fix it
not checked	no route answered	re-run when the network is back

Only ABSENT fails the run; add --strict to fail on not verbatim too, when
finishing a record. The report names which of case, quote or dash style actually
differs, not all three. Whitespace differences pass silently: stripping
<xref> tags renders "(Kruse et al., 2005)" with spaces inside the parentheses,
and that is an extraction artefact, not a difference in the text. Hyphens stay
significant — "rod-like" and "rodlike" are different words.

Rules

1. Quote the claim, not the topic. A sentence containing the same words is
   not support. The MreB record cited a review for "MreB is required for rod
   shape"; the quotation that survived says depletion causes rounding and that
   no mutant restores shape without it — the second clause is what makes it
   REQUIRED_FOR rather than MODULATES.
2. Read around the sentence. --suggest returns sentences stripped of
   context, and papers contain sentences stating what the authors go on to
   refute. Full text is fetched to build/literature/; read it.
3. An abstract is a weak source for a specific number. Physical properties,
   subunit counts and rates need the methods or results, not the summary. Four
   such properties were removed in #91 for resting on titles alone.
4. Unreadable is a finding, not a dead end. Say so in notes and leave
   snippet empty rather than paraphrasing a title into something that reads
   like knowledge of the paper's contents. unchecked is not that finding —
   re-run it when the network is back before concluding anything.
5. A quotation contradicting the record outranks the record. File it; do not
   quietly pick a different sentence.

Establishing readability

Probe; never trust a flag. NCBI's ID converter knows only PMC, so a paper
absent from PMC looks absent from PubMed too — that is how PMID 21047262 was
recorded as "no PubMed record" in #133 and the citation removed on a false
premise. Europe PMC's inEPMC: Y is no better: it says so for PMC5433867, whose
fullTextXML returns 404. The script therefore tries NCBI's converter, Europe
PMC search, efetch db=pmc, Europe PMC full text, then efetch db=pubmed, and
reports what actually came back.

If you assert a reference is unreadable, name the routes tried.

A 200 is not text. Publisher text-mining endpoints answer 200 with an
entitlement stub: Crossref's text-mining links for three Elsevier papers each
returned ~1.8 KB of metadata — title, journal, openaccess: false, no body
(#142). Adding that route would have registered them as full text and then
reported their correct quotations as fabricated. Any new full-text route must be
shown to contain body prose, not merely to return 200 — the same rule that
retired Europe PMC's inEPMC flag and Semantic Scholar's isOpenAccess.

Not every reference is a paper

SubtiWiki gene pages, GO terms and deposited datasets are not_literature:
"what does it say" is the wrong question, and check_curies.py already answers
the right one. Counting them among unreadable papers inflated the headline from
3 to 21. --verify refuses a snippet on one of these outright.

Before committing evidence

just evidence-verify   # bare; a pipe hides the exit code (PR #39)
just qc

--verify needs the network. It is not part of just qc, which runs offline —
so it is a discipline, not a gate that will catch you. Run it.

Related

• scripts/check_curies.py — does the identifier resolve (not what it says).
• docs/CURATION.md — the verbatim-snippet rule and the evidence contract.
• source-queue — whether to adopt a source at all.