Most evidence items carry a paraphrase, not a verbatim snippet

[realmarcin]

Raised by a reader asking what the evidence actually was behind the MreB record's trait links. The honest answer for that record was: a paraphrase attributed to papers I had not read. Measuring the rest: 0 of 262 evidence items in the corpus carry a snippet. Every one is notes — my restatement of what a source is taken to say.

docs/CURATION.md already draws the line: "snippet is for verbatim quotes only. When you have not seen the source text, use notes to paraphrase what the source establishes." The corpus has been technically compliant with that rule while making it carry all the weight — a reader cannot distinguish 'paraphrased from a text I read' from 'inferred from a title'.

Two specific failures found while checking one record:

1. A citation nothing could be read from. DOI:10.1146/annurev-genet-102108-134845 (Annual Review of Genetics 2010) has no PubMed and no PMC record — the identifier resolves at Crossref, so the liveness gate passes it, but there is no free abstract, let alone full text. It was cited for a specific claim about MreB and cell shape. Removed in the fix.
2. A claim the source contradicts. The record said shape and mreB track each other. Shi et al. 2018 says: "Many rod-like and curved bacterial species require MreB to elongate, although there are exceptions such as the Actinobacteria and Rhizobiales that lack MreB altogether." Reading the paper bounded the claim; not reading it had let the claim overreach.

Suggested work, roughly in order:

• Prefer sources that can be read. Record whether a reference has an open-access full text (PMC id) at citation time; the liveness checker already resolves DOIs and could resolve to PMC in the same pass.
• Snippet the load-bearing claims first — trait links, causal-graph edges and protein_examples, which are curator-asserted and where a wrong paraphrase is invisible.
• Consider a gate: a claim in one of those fields whose evidence has no snippet and whose reference has no open-access text is a claim nobody can check. That is a reportable state, even if not initially a failing one.

[realmarcin]

Correction, and the countable version of this issue

First, a correction. I claimed here that DOI:10.1146/annurev-genet-102108-134845
has no PubMed and no PMC record. It has PMID 21047262 with an abstract. I
had asked only NCBI's PMC ID converter — which indexes PMC alone — and treated
its empty response as absence from PubMed too. Europe PMC finds it at once.

Second, the number this issue asked for. just evidence-audit now reads
every cited reference and reports what is reachable. Across 341 evidence items
and 86 distinct references:

	count
full text	34
abstract only	31
unreadable	1
not literature (database record, ontology term, dataset)	20

3 claims cite a paper with no reachable text and carry no quotation — all
three cite DOI:10.1128/9781555819323.ch10, an ASM Press book chapter. Those are
the claims nobody can check.

That first run reported 21 unreadable. Eighteen were SubtiWiki gene pages, one a
GO term, two deposited datasets. "What does it say" is the wrong question for a
database record, and check_curies.py already answers the right one; counting
them here inflated the headline sevenfold. They are now classified separately.

Third, the gate. just evidence-verify checks every snippet against its
own source text and fails on any that is not there. Canarying it against this
corpus's three real snippets found three bugs in the checker before it found
anything in the corpus — including a whitespace-sensitive comparison that
rejected a quotation genuinely present in the paper, because tag-stripped JATS
renders (Kruse et al., 2005) with spaces inside the parentheses. Six mutation
tests now hold the loosened comparison to catching reworded, negated, spliced
and fabricated quotations.

just evidence-suggest <record> proposes candidate sentences ranked against the
claim and deliberately never picks one — choosing which sentence supports a
claim is the judgement this field exists to record.

Branch literature-evidence; PR to follow.

[realmarcin]

Retitled from "No evidence item in the corpus carries a verbatim snippet (0 of 262)".

Both halves had drifted:

• 0 → 3. PR MreB: quote the sources instead of paraphrasing them #134 added three verbatim snippets to the MreB record, all
  verified against PMC5846203.
• 262 is confirmed, not changed. The audit initially reported 341, which is
  every field carrying a reference — including taxon notes, canonical examples,
  causal-graph edges and images. Separating those from entries in an evidence
  list (The audit's "341 evidence items" is a different population from #133's 262 #147) gives 262 evidence items, exactly this title's figure.

Current state, from just evidence-audit on 89 distinct references:

	count
full text	37
abstract only	31
unreadable	1
not literature	20

3 claims cite a paper with no reachable text and carry no quotation — all
three the same ASM book chapter.

The issue stays open: the tooling to close it exists, the evidence does not.
259 evidence items still carry only a paraphrase.