A third of the papers we said we could not read were free to read in full

The second row is the control and it is the whole argument. If 'full text not read' mostly meant 'full text unavailable', the declared-shallow papers would be harder to get than the rest. They are easier. Availability is not the explanation; time is, which is a legitimate reason to stop at an abstract and an illegitimate thing to leave unsaid.

established sits lowest at 14.7%, which is correct rather than surprising: those items lean on guidelines, registries and databases, which have no PMC deposit to find. The mass is in strong — 510 items, 175 of them resting on a paper that was free and unread.

The action is not to read 282 papers. It is to intersect three lists this project already has: declared abstract-only, full text available, and low claim-to-abstract overlap from verify-claims.py. A claim that the abstract cannot contain — a model system, a methods detail, an effect size, a subgroup — resting on a free full text is a short and well-defined queue.

It is not evidence those claims are wrong, and the graph's own instruments say the opposite. ana-claim-support-check found the declared-abstract-only band scores higher on claim-to-abstract correspondence than everything else — median 0.779 against 0.640, with 6 of 826 below the floor against 77 of 1,990. Claims written from an abstract are, unsurprisingly, good claims about that abstract.

The defect is descriptive. 'Full text not read' reads as a property of the source and is a property of the reader, and the entries do not distinguish the two. The honest caveat for a third of these is the full text was available and was not read, which is a sentence an agent can write and a reader can act on.

The first run of this audit reported 34.0% on a population of 887 items. Both numbers were inflated, because the shared shallow-reading regex had no negation handling: a caveat reading “Full text read (PubMed Central, PMC4523210), not abstract only” contains the literal words “abstract only” and was counted as a declaration of shallow reading. 165 of 888 matches (18.6%) were negations of that shape.

The headline survives — a third either way, with the control still lower, so the availability argument is unaffected. Three other numbers in this graph did not: ana-reading-depth's headline moves from 29.1% to 24.3%, and ana-claim-support-check's calibration band loses four of its six floor-breaches, which is what should happen — an item whose agent read the full text writes a claim carrying full-text detail, which scores low against the abstract by construction. That entry had called those six the ones with no innocent explanation; four had the most innocent explanation there is.

How it was found matters more than the correction. This audit produced a reading queue; the queue's top item was read; the paper's claim checked out in every particular — WBC 188,000 with 70% blasts, 36 weeks, fetal decelerations, the rash, the hepatosplenomegaly, anisocoria, left-sided weakness, t(11;19)(q23;p13.3) — and the entry's caveat turned out to begin with the words Full text read. The instrument was caught because someone acted on its output. A measure that is reported and never used is never tested, and this project has now shipped three quality metrics whose defects only surfaced when something downstream consumed them.

scripts/reading-queue.py intersects three lists: declared abstract-only, open-access full text available, and claim-to-abstract support below a threshold. 8 items out of 723 come through at a 0.45 cut — a queue small enough that not reading it is a decision rather than a capacity limit.

This table replaces one written an hour earlier, and the difference is the whole lesson. The pre-correction queue held 11 items, six of them patient stories, and this entry said so and offered a tidy structural explanation: a case report's abstract is four sentences, so its clinical detail necessarily lives in the body. The explanation was elegant and entirely wrong. Every one of those patient stories was a misclassified item whose caveat began Full text read; they scored low against the abstract precisely because their agents had read the full text and written from it. Fixing the guard removed all six, and the plausible story about case reports with them.

What is left is two analyses, two experiments, a fusion entry, a hypothesis and a collaborator entry — no pattern, which is the honest result. Two of the eight are this project's own analyses citing PMC10169636, so the queue's first duty is to audit the auditor.

Any single one of the three filters is a bad queue. Declared-abstract-only alone is 723 items and most are fine. Low support alone is mostly faithful paraphrase in different words. Available-but-unread alone is a budget fact. The intersection is the only one of the seven combinations that names something actionable.

Six of the eight items were examined in the same cycle the queue was built. Three were real defects and three were legitimate low scores, which is the only precision figure this project has for any of its quality instruments.

The three legitimate ones share a shape: the claim cites a paper for its bearing on a decision — who to credit, which readout to use, which confounder to design around — rather than for a finding. An abstract cannot contain a statement about why this project made a choice, so the score is low and correctly so. That is a fourth distinct reason for low overlap, alongside faithful paraphrase, full-text detail, and the wrong-in-kind defect, and only the last is a problem.

Three defects from six readings is a better hit rate than any list this project has produced. It is also six readings, so the interval around 50% runs from roughly 12% to 88%.