Claim–Evidence Review: 10 Lung Disease Entries (2026-07-25)

Correctness review of ten primarily-pulmonary kb/disorders/ entries, focused on whether each evidence snippet actually supports the claim it is attached to.

Scope

Entry	Lines	Evidence items
`Asthma`	2471	189
`Chronic_Obstructive_Pulmonary_Disease`	1252	111
`Idiopathic_Pulmonary_Fibrosis`	1214	45
`Cystic_Fibrosis`	3261	193
`Bronchiectasis`	630	20
`Silicosis`	730	24
`Asbestosis`	647	23
`Hypersensitivity_Pneumonitis`	686	36
`Hereditary_Pulmonary_Alveolar_Proteinosis`	828	55
`Non-Small_Cell_Lung_Cancer`	1037	118
Total	12,756	814

Method

Verbatim fidelity (mechanical). Every one of the 814 snippets was checked as an exact (whitespace/quote-normalised, ellipsis-part-aware) substring of its references_cache/ body.
Label consistency (mechanical). Scans for claims whose evidence is entirely NO_EVIDENCE/REFUTE, for supports values contradicted by their own explanation, for evidence_source mislabelling, and for quantitative claims whose numbers appear nowhere in the cited snippet.
Semantic support (manual). Read-through of claim ↔ snippet ↔ explanation for every pathophysiology, genetic, treatment, phenotype, and prevalence record.
Full-text re-adjudication. Every flagged item whose reference has (or could be fetched to have) a full-text cache was re-checked against the whole paper, not just the abstract — because a claim unsupported by the quoted sentence may still be supported elsewhere in the source. That distinction separates a bad citation from an under-selected snippet, which need opposite fixes. Searches de-hyphenate OCR line-breaks (paren-\nchyma → parenchyma) before matching.

Headline result

Verbatim snippet fidelity is clean. All 814 snippets verify against their cached sources. 44 initially flagged as mismatches were traced to Unicode/punctuation artifacts in the checker, not the data — e.g. the YAML writes >/= grade 2 where the cached abstract holds ≥ grade 2 (PMID:30537755), and several snippets carry a trailing period absent from the source (PMID:10907591, PMID:3595047). No fabricated quotes were found.

The real defects are in claim–evidence matching, not quoting — and they are concentrated in three of the ten files. A full-text pass (see below) then split those matching defects in two: some citations are genuinely wrong, while others are correct citations with a badly chosen quote. One further defect sits below the entries entirely — a cache file holding the wrong paper's full text (Finding 4b).

The entries split into two clear cohorts:

Cohort	Entries	Character
Well-curated	`Idiopathic_Pulmonary_Fibrosis`, `Cystic_Fibrosis`, `Bronchiectasis`, `Silicosis`, `Asbestosis`, `Hypersensitivity_Pneumonitis`, `Hereditary_Pulmonary_Alveolar_Proteinosis`	100% `evidence_source` coverage; tight claim↔snippet correspondence; deliberate use of `PARTIAL`/`REFUTE`; explanations that name the study design
Legacy / needs repair	`Asthma`, `Chronic_Obstructive_Pulmonary_Disease`, `Non-Small_Cell_Lung_Cancer`	367/418 evidence items missing `evidence_source`; off-topic citations; circular explanations; frequency bands contradicted by their own snippets

evidence_source coverage is the cleanest proxy for the split:

Asbestosis                                  missing    0/ 23
Bronchiectasis                              missing    0/ 20
Cystic_Fibrosis                             missing    0/193
Hereditary_Pulmonary_Alveolar_Proteinosis   missing    0/ 55
Hypersensitivity_Pneumonitis                missing    0/ 36
Idiopathic_Pulmonary_Fibrosis               missing    0/ 45
Silicosis                                   missing    0/ 24
Asthma                                      missing  148/189
Chronic_Obstructive_Pulmonary_Disease       missing  103/111
Non-Small_Cell_Lung_Cancer                  missing  116/118

Full-text coverage. 46 of the 456 unique references cited by these entries have full-text caches (40 already committed; 6 more fetched during this audit). Those back 97 of the 814 evidence items. Re-fetching every abstract-only reference relevant to a finding yielded full text for 6 — the rest are genuinely abstract-only, so those findings could not be re-adjudicated and stand as written.

Meta-finding: abstract-only quoting systematically understates support

Twelve weak-labelled (PARTIAL/NO_EVIDENCE/REFUTE) items are backed by full text. Checking them against the whole paper rescues four outright — the citation is correct and the paper does support the claim; the curator simply quoted a weak sentence from the abstract when an on-point sentence exists in the body.

This reframes part of the audit. Several PARTIAL labels in the legacy cohort are not evidence weaknesses at all — they are artifacts of abstract-only curation, and the fix is to re-quote, not to weaken the label or drop the citation.

Item	Quoted (abstract)	Available in full text
COPD `pathophysiology[3]` Alveolar Destruction / PMID:11993785	"In the lung parenchyma, emphysema defined as alveolar destruction and airspace enlargement is present" (`PARTIAL`)	"The mechanism responsible for the development of emphysematous lesions is thought to be an imbalance in protease and antiprotease enzymes in the lung" and "the loss of elastic recoil induced by the emphysematous lungs…"
COPD `pathophysiology[1]` Chronic Inflammation / PMID:11993785	`PARTIAL`, explanation: "…not specifically detailing the involvement of neutrophils, macrophages and…"	"predominant inflammatory cellular infiltrations … identified as activated CD8+ T-lymphocytes and macrophages … A number of neutrophils have been found in the bronchial lumen"
COPD `environmental[1]` Air Pollution / PMID:25673984	"The major pathogenic factors … infection and inflammation, protease and antiprotease imbalance, and oxidative stress" — says nothing about pollution	"Such exposures can include environmental tobacco smoke, burning of biomass, and air pollution particles. All these particle exposures … can result in oxidative stress"
COPD `pathophysiology[3]` Alveolar Destruction / PMID:32493486	a surfactant-homeostasis sentence (`PARTIAL`)	"The pathophysiology of chronic obstructive pulmonary disease (COPD), which involves emphysematous destruction of alveolar sacs and airway remodeling…"

Notably, the protease–antiprotease mechanism I first recorded as missing from the COPD Alveolar Destruction node is present in a paper the node already cites — and it is the mechanism modelled by emphysema_protease_antiprotease_imbalance, so re-quoting also opens a natural conforms_to edge.

Full text confirms (and strengthens) the bad-citation findings

Where full text was available for a citation I judged off-topic, it removed all doubt:

Citation	Full-text verdict
PMID:37461046 in COPD `prevalence[0]` (Finding 1)	The only occurrence of "COPD" in the entire paper is inside a bibliography entry title in the reference list. Zero COPD content in the body.
PMID:33533174 in NSCLC `pathophysiology[5]` Metastasis (Finding 8)	Exclusively SCLC. Its three NSCLC mentions are explicitly contrastive — "Unlike non-SCLC (NSCLC), which has an intrinsic tendency for CHT resistance, SCLC is tantalizingly chemosensitive". The paper draws the very distinction the citation elides.
PMID:32361678 in COPD `treatments[3]` PDE4 Inhibitors (Finding 9)	Zero occurrences of "PDE4" or "phosphodiesterase" anywhere in the full text. The explanation's conflation of PDE4 inhibitors with kinase inhibitors is a genuine error.
PMID:22550239 in NSCLC `pathophysiology[2]` Angiogenesis	Zero occurrences of angiogenesis / angiogenic / VEGF / neovascular. Confirms the existing `NO_EVIDENCE`; the citation should be deleted, not retained.
PMID:20536932 in NSCLC `phenotypes[6]` Chest Pain	Zero occurrences of "chest pain". Confirms `NO_EVIDENCE`; delete.
PMID:29433833 in COPD `pathophysiology[3]` Alveolar Destruction	About macrophage phagocytosis/efferocytosis; no emphysema or alveolar-destruction content. The existing `PARTIAL` and its candid explanation are correct as written.

High-severity findings

1. COPD global prevalence is asserted with zero supporting evidence, one citation being a silicosis paper

Chronic_Obstructive_Pulmonary_Disease.yaml:51-67

prevalence:
- population: Global
  measure_type: POINT_PREVALENCE
  prevalence_class: ABOVE_1_IN_1000
  rate_per_100000: 11700.0
  percentage: 11.7
  evidence:
  - reference: PMID:35261410
    supports: NO_EVIDENCE
    explanation: ...does not mention a global prevalence rate of 11.7%.
  - reference: PMID:37461046
    supports: NO_EVIDENCE
    explanation: This study focuses on ...silicosis, not COPD.

A structured numeric claim (rate_per_100000: 11700.0) with both evidence items marked NO_EVIDENCE — the entry documents its own lack of support. One citation (PMID:37461046, Global incidence, prevalence and disease burden of silicosis) is a different disease entirely.

Fix: PMID:35261410 reports a pooled prevalence of 11.1% (GOLD fixed criteria) across eight countries — not global. Set rate_per_100000: 11100.0, change population to match the study's actual scope, mark supports: SUPPORT, and delete the silicosis citation. Also drop the deprecated percentage: field (design decision §8).

2. COPD respiratory-failure frequency band is contradicted by its own evidence

Chronic_Obstructive_Pulmonary_Disease.yaml phenotypes[5]

frequency: OCCASIONAL          # HP:0040283 = 5-29% of patients
name: Respiratory Failure
evidence:
- reference: PMID:14621114
  supports: REFUTE
  explanation: ...respiratory failure is an important and common complication of COPD, not an occasional one.
- reference: PMID:38692758
  supports: REFUTE
  explanation: ...hypoventilation, which can lead to respiratory failure, is not uncommon in COPD.

Both evidence items refute the frequency qualifier, not the association. The band is simply wrong.

Fix: raise to FREQUENT (30–79%) and re-label both items SUPPORT.

3. NSCLC `progression[0]` has a structured-only claim and seven off-topic citations

Non-Small_Cell_Lung_Cancer.yaml progression[0:Onset] states its claim only as structured fields — phase: Onset with age_range: 60-80 — with no prose description or notes to say what the evidence is meant to support. Seven evidence items hang off it, all NO_EVIDENCE. Two are grossly off-topic:

PMID:38377969 — multiple sclerosis disease-progression modelling
PMID:34911717 — early-onset colorectal cancer in Lynch syndrome

These are citation-relevance failures, not evidence. An NO_EVIDENCE item whose explanation reads "the study focuses on multiple sclerosis" carries no information and should be deleted, not retained.

Fix: add a prose claim matching the structured age_range (PMID:15477641 — "over 50% of all patients with non-small cell lung cancer (NSCLC) are 65 years of age or older" — is a reasonable partial source, and would be PARTIAL rather than NO_EVIDENCE against a stated claim), and delete the MS and Lynch citations.

4. NSCLC has a duplicated `Bone Pain` phenotype

Non-Small_Cell_Lung_Cancer.yaml:375-378 and :429-433 — two phenotypes entries with identical category: Musculoskeletal, name: Bone Pain, frequency: OCCASIONAL, and notes: May indicate bone metastases, sharing two of the same citations (PMID:20536932, PMID:26690845). One is redundant.

4b. Full-text cache contamination: PMID:31922885 holds a different paper entirely

references_cache/PMID_31922885.md — cited by Idiopathic_Pulmonary_Fibrosis mechanistic_hypotheses[1] as the REFUTE evidence — has a body containing French-language human–computer-interaction ergonomics text:

full_text_provider: openalex
full_text_url: "https://inria.hal.science/inria-00075378"

Les ergonomes réalisent en général des évaluations d'interfaces homme-machine (IHM)
dans 4 contextes principaux… Ce texte propose une revue critique des pratiques
actuelles de l'évaluation d'interface.

The real paper is "Diverse Injury Pathways Induce Alveolar Epithelial Cell CCL2/12, Which Promotes Lung Fibrosis" (Am J Respir Cell Mol Biol, 2020). OpenAlex resolved its full text to an unrelated INRIA HAL technical report and the fetcher ingested it, while content_type still advertises full text.

This is pre-existing in the committed cache (introduced by commit 94e6b830), not by this audit. Severity:

The currently quoted snippet is drawn from the PubMed abstract portion of the file, which is intact — so the IPF evidence item itself is presently valid, and the REFUTE judgement is correct.
But anything quoted from that body would be from an unrelated document and would still pass validate-references, because the validator only checks substring membership in the cache. The anti-hallucination guarantee silently does not hold for this file.

Scope: 1 of the 46 full-text caches used by these ten entries (~2%). The other three institutional-repository resolutions (inserm.hal.science → PMID:38847551, research.rug.nl → PMID:34524912, researchportal.bath.ac.uk → PMID:33197388) were spot-checked and are correct, so this is not a blanket problem with the OpenAlex path — but it is undetectable by current QC.

It is not an isolated file. A corpus-wide sweep (title-word overlap: what fraction of a record's own title: words appear anywhere in its fetched body) surfaces four more caches at 0% overlap, of which a spot-check confirms at least one is the identical failure mode:

DOI_10.1186_s12890-024-03371-5.md
  title:         Evolution of treatment strategies for solid tumors with RET
                 rearrangement … Non-small Cell Lung Cancer (NSCLC)
  full_text_url: https://purehost.bath.ac.uk/…/2014_02_07_Apereo_Europe_2014.pdf
  body:          "Cope, J 2014, 'Training researchers with Sakai',
                  Esup Day 17 & Apereo Europe 2014, Paris…"

Also flagged at 0%: DOI_10.1038_ejhg.2014.61.md, DOI_10.1055_s-0045-1806986.md, DOI_10.11588_heidok.00035789.md. (Four url_…accessdata.fda.gov… hits are false positives — their title is a URL, so title-overlap is meaningless for them.)

Suggested fix: re-fetch the affected files, and add a cheap cross-check to check-reference-cache-frontmatter. Caveat on the detector: title-word overlap did not catch PMID:31922885, because these caches concatenate a correct PubMed abstract with the wrong full-text body — the title words are present in the abstract half. A robust check needs to score the fetched body separately from the abstract, or add a language-ID signal. Both the finding and this blind spot are worth their own issue.

5. NSCLC `ROS1` frequency band contradicts its own snippet

genetic[2:ROS1] is frequency: OCCASIONAL (HP:0040283, 5–29%), while the cited snippet reads "ROS-1 rearrangement is found in 0.9-2.6% of non-small-cell lung cancers" — i.e. VERY_RARE (<5%). ALK at 2–7% straddles the boundary and is defensible; ROS1 is not.

Medium-severity findings

6. `supports` labels contradicted by their own explanation

Asthma has_subtypes[2:Adult-Onset Asthma] / PMID:36833767 — supports: NO_EVIDENCE, but the explanation reads "...which supports that adult-onset asthma could have environmental triggers." One of the two is wrong.

7. Treatment-toxicity evidence used to support disease phenotypes (NSCLC)

phenotypes[4:Fatigue] — four of five citations describe chemotherapy toxicity (PMID:38469616 cancer-related fatigue after chemotherapy; PMID:33755621 "adverse effects of treatment"; PMID:26990789 "toxicities included…"; PMID:30537755 "non-hematological toxicities"). Only PMID:32013812 addresses fatigue as a disease symptom. Treatment toxicity is a distinct claim from disease phenotype — the repo has a myelosuppression module precisely for this distinction.
phenotypes[0:Persistent Cough] — PMID:35224703 and PMID:37920959 both measure postoperative cough after lung-cancer surgery, not cough as a presenting feature.

8. Wrong-disease evidence retained as SUPPORT

NSCLC pathophysiology[5:Metastasis] — the lead citation PMID:33533174 reports metastasis site patterns in SCLC, used to support NSCLC organotropism. The explanation concedes this ("While this article primarily deals with small cell lung cancer…") yet the label is SUPPORT. Full text confirms the paper is entirely SCLC and mentions NSCLC only to contrast the two — see the full-text table above.
Asthma pathophysiology[0:Airway Inflammation] — PMID:23234454 is about cardiac asthma (a heart-failure entity, not asthma). Correctly PARTIAL, but it does not belong on an asthma airway-inflammation node.
COPD biochemical[0:Arterial Blood Gases] — PMID:34756790 describes blood-gas derangement after a detonation/nitrogen-compound inhalation injury; the explanation hand-waves "including COPD contexts". Marked SUPPORT.

9. Claims whose mechanism is not in the snippet

COPD pathophysiology[4:Oxidative Stress and Mitochondrial Dysfunction] — claim is cigarette smoke triggers mitochondrial ROS, impaired mitophagy, reduced antioxidant defences. The sole snippet is "Shared benefits include mitigation of oxidative stress, mitochondrial dysfunction, and extracellular matrix remodeling" — a statement about what SIRT1 activation mitigates, which does not establish the claimed trigger mechanism. Marked SUPPORT.
COPD pathophysiology[3:Alveolar Destruction] — all four items PARTIAL. Partly resolved by full text: PMID:11993785 and PMID:32493486 both support the claim in their bodies (see the rescue table), and PMID:11993785 supplies the protease–antiprotease/elastic-recoil mechanism I first recorded as missing. What remains a defect: PMID:24707174 (a list of eight endothelial microparticle surface markers, abstract-only, unrelated to alveolar destruction) and PMID:29433833 (efferocytosis, full text confirms no alveolar-destruction content).
COPD treatments[3:Phosphodiesterase-4 Inhibitors] — PMID:32361678 is about protein kinases; the explanation conflates them with PDE4 inhibitors ("specifically kinase inhibitors"). Marked SUPPORT. Full text confirms zero mentions of PDE4 or phosphodiesterase — the citation is simply wrong for this node.
~~COPD environmental[1:Air Pollution] — PMID:25673984~~ — withdrawn on full-text review. The paper does support the claim ("…burning of biomass, and air pollution particles. All these particle exposures … can result in oxidative stress"); only the quoted sentence was wrong. Re-quote rather than re-label.
Asthma pathophysiology[2:Bronchoconstriction] — PMID:27603525 (statins in animal models) supports the claim only via the non-sequitur "the use of statins to reduce airway hyper-responsiveness implies the involvement of smooth muscle cells"; also animal data with no evidence_source.
Silicosis pathophysiology[7], [8] and phenotypes[5] (abstract-only sources; not re-adjudicable — refetch yielded no full text) — specific claims (TGF-β/PDGF/TNF-α-driven myofibroblast differentiation; the concentric whorled hyalinised silicotic nodule; silica impairs macrophage killing of M. tuberculosis) rest on generic "…with subsequent fibrosis" / "…associated with silica dust exposure" snippets. The associations hold; the mechanisms are unsupported. PARTIAL would be more accurate than SUPPORT.

10. Quantitative descriptions outrunning their evidence (Cystic Fibrosis)

Several Cystic_Fibrosis nodes carry precise percentages supported only by a generic list-of-complications snippet from PMID:30986316 ("Other complications include sinusitis, diabetes mellitus, bowel obstruction, hepatobiliary disease, hyponatremic dehydration, and infertility"):

Node	Unsupported figure
`pathophysiology[15]` Exocrine Pancreatic Insufficiency	85–90%
`pathophysiology[20]` Intestinal Obstruction	meconium ileus 15–20%
`pathophysiology[21]` Hepatobiliary Obstruction	cirrhosis 5–10%
`pathophysiology[24]` Vas Deferens Agenesis	CBAVD 97–98%
`pathophysiology[25]` Sinonasal Disease	nasal polyposis 10–32%
`treatments[11]` Lung Transplantation	5-yr survival ~50–60%

The existence of each complication is supported; the numbers are not. Per docs/frequency-evidence-guidelines.md, a quantitative claim needs its own citation. This is the least severe class here — the figures are all clinically conventional — but they should either get a source or move to notes.

Process observation: DOI references are skipped by the reference validator

conf/reference_validator_config.yaml:33 lists DOI in skip_prefixes, so linkml-reference-validator does not snippet-check DOI-referenced evidence. On Bronchiectasis, whose evidence is mostly DOI-based, the wrapper reports:

Validation Summary:
  Files validated: 1
  Total checks: 0
  All validations passed!
  Snippets checked: 5/20 verified against cached references (15 skipped by prefix)

Read the last line, not the Total checks: 0. Per CLAUDE.md and issue #7252, Total checks counts issues found, so it is 0 on any clean run; the affirmative signal is the Snippets checked: N/N line that scripts/run_reference_validator.sh appends. Here it says plainly what happened: 5 of 20 items were verified and 15 were skipped by prefix.

So the gap is real but narrower than "nothing was validated": across the ten entries, 20 DOI-referenced items (plus 1 PPR: preprint) are exempt from snippet checking. All 20 have populated references_cache/DOI_*.md files, and independently checking them confirms every snippet is an exact substring — so there is no data defect. Since the caches exist, DOI could be removed from skip_prefixes and validated like PMID.

What is exemplary

Worth preserving as curation patterns:

Idiopathic_Pulmonary_Fibrosis mechanistic_hypotheses[1] — a hypothesis carrying six SUPPORT items, a PARTIAL that qualifies its immune-independence claim (PMID:34524912, cGAS-STING in senescent IPF epithelium), and a REFUTE (PMID:31922885, epithelial CCL12 deletion protects from fibrosis) that argues directly against the model. This is how a contested mechanism should be curated.
Asbestosis genetic[0:GSTM1] — a SUPPORT (PMID:8000297) paired with a REFUTE (PMID:17563610) from a larger nested case-control study, with the description honestly stating "Its role in asbestosis susceptibility remains unsettled".
Silicosis — explanations that name the study design and justify the label ("Evidence source OTHER (review)", "marked PARTIAL"), so a reviewer can audit the reasoning without re-reading the source.
Hereditary_Pulmonary_Alveolar_Proteinosis — disciplined ORPHA row quoting, with every phenotype frequency traceable to a specific Orphanet HPO row.
The newer Asthma nodes (SIRT1/NAD+, early-life rhinovirus–CDHR3, HDM Der p 1 protease–ROS signalling) are as tight as anything in the well-curated cohort — the file's weakness is entirely in its legacy nodes.

Suggested remediation order

Re-fetch references_cache/PMID_31922885.md (contaminated body) and sweep the corpus for the same OpenAlex mis-resolution — this is the only finding that breaks a guarantee rather than a single entry
COPD prevalence[0] — replace the unsupported 11.7% and delete the silicosis citation
COPD phenotypes[5] — fix the respiratory-failure frequency band
NSCLC progression[0] — write the claim; delete the MS and Lynch citations
NSCLC — de-duplicate Bone Pain; fix the ROS1 frequency band; delete the full-text-refuted PMID:22550239 (angiogenesis) and PMID:20536932 (chest pain) citations
Re-quote, don't re-label — the four rescued items above (COPD alveolar destruction ×2, chronic inflammation, air pollution): swap the abstract sentence for the on-point full-text sentence and raise PARTIAL → SUPPORT. While re-quoting alveolar destruction, consider a conforms_to: emphysema_protease_antiprotease_imbalance#… edge.

⚠️ Hyphen-wrap hazard when quoting PDF-derived caches. The validator normalises runs of whitespace, so an ordinary line wrap inside a snippet is safe — but PDF-extracted bodies are hyphen-split at line ends (paren-\nchyma, air-\nways), and those hyphens are real characters in the cache. A snippet spanning one will fail validate-references. PMID_32493486.md has ~280 such lines, PMID_11993785.md ~77, PMID_37461046.md ~10. Either choose a quote that sits within one line, or verify with just validate-references <file> before committing. (This report's own Method §4 de-hyphenates before matching, which is why it could find these sentences in the first place — curators quoting them do not get that for free.)

⚠️ validate-references rewrites the whole cache as a side effect — stage only your YAML. just validate-references <file> runs just fix-references-cache first (project.justfile:734), and that recipe globs the entire references_cache/ directory (project.justfile:815), re-quoting frontmatter wherever it is unquoted:
-reference_id: DOI:10.1001/jama.2008.598
+reference_id: "DOI:10.1001/jama.2008.598"
Validating a single file can therefore leave thousands of unrelated cache files modified — observed runs during this review and its code review ranged from ~3,000 to ~7,800, the count depending on how much of the cache was already quoted when the recipe ran. Pairing that with git add -A would commit all of them, which is exactly what CLAUDE.md's targeted-git add rule exists to prevent.

Recovery, in this order: stage your edited YAML (and only the caches you deliberately re-fetched) first, then git checkout -- references_cache/ to discard the rest. The order is load-bearing — git checkout -- <path> restores from the index, not HEAD, so staging first is what protects your deliberate re-fetches. 7. Replace the genuinely wrong citations in findings 8–9 (PMID:33533174, PMID:32361678, PMID:24707174) 8. NSCLC / Asthma / COPD — backfill evidence_source (367 items) 9. Cystic Fibrosis — source or demote the six quantitative descriptions 10. Consider removing DOI from skip_prefixes so DOI evidence is CI-validated

Process implication. Four of twelve full-text-backed weak labels were curation artifacts rather than evidence problems. Where a full-text cache exists, curators should quote from the body, not the abstract — and a PARTIAL whose explanation reads "the abstract does not mention X" is a prompt to check the full text before accepting the label.