Claim–Evidence Review: 10 Exposure-Related Disorders (2026-07-25)

Scope

Correctness review of ten exposure/toxicology entries in kb/disorders/, with the review lens set on claim–evidence matching: does each cited snippet actually support the claim it is attached to, is the supports/evidence_source tagging right, and are there claims asserted with no evidence at all?

Entries reviewed (three exposure families):

Family	Entries
Heavy metals	Arsenic_Poisoning, Lead_Poisoning, Mercury_Poisoning, Cadmium_Poisoning, Thallium_Poisoning
Gases / chemicals	Carbon_Monoxide_Poisoning, Methanol_Poisoning, Organophosphate_Poisoning
Pneumoconioses	Silicosis, Asbestosis

Method

1. Schema validation — linkml-validate against the Disease class, all 10 files.
2. Snippet fidelity — every evidence item's snippet normalized (whitespace, Unicode dashes/quotes, case) and checked as an exact substring of the corresponding references_cache/ body. 578 evidence items across all reference types (PMID, DOI, GEO, clinicaltrials).
3. Semantic claim–evidence pass — manual read of all 578 (claim, description, snippet, supports, evidence_source, explanation) tuples, judging whether the snippet supports the specific proposition asserted.
4. Coverage pass — enumeration of top-level items in evidence-bearing sections carrying no evidence block; audit of every frequency: band against its snippets.

Mechanical results: clean

• Schema: 10/10 No issues found.
• Snippet fidelity: 578/578 snippets are exact substrings of their cited cached source. Zero fabricated or paraphrased quotes, zero missing cache entries, including the 14 DOI-keyed deep-research references. No evidence of the classic hallucination modes (invented PMIDs, invented quotes).

The anti-hallucination stack is doing its job on these entries. Every finding below is a semantic issue that snippet-substring validation cannot catch.

Overall assessment

Entry	Claim–evidence quality	Notes
Arsenic_Poisoning	Excellent	Best in the set. Descriptions explicitly bound their own claims ("does not justify a universal sulfhydryl-enzyme blockade model", "cross-sectional associations do not by themselves establish fibrosis"). Same paper correctly split into MODEL_ORGANISM and IN_VITRO items by evidence type. Carries a REFUTE-flavored discussion citing the negative DMSA RCT.
Methanol_Poisoning	Excellent	Tight mechanistic chain, each step its own citation. No unevidenced items except one biochemical readout.
Carbon_Monoxide_Poisoning	Very good	Honest hedging where the snippet supports the consequence rather than the mechanism; discussions carries the negative HBO trial as supports: REFUTE.
Thallium_Poisoning	Very good	Minor review-vs-clinical tagging issue; one missing explanation.
Organophosphate_Poisoning	Very good	One poor snippet choice (below).
Asbestosis	Good	Exemplary handling of conflicting evidence (GSTM1: SUPPORT and REFUTE side by side). Two core mechanism nodes unevidenced.
Silicosis	Good	Explicitly annotates review-derived evidence as OTHER in explanation text. 9 unevidenced items including all three subtypes.
Lead_Poisoning	Mixed	Broad, well-cited backbone, but repeated use of generic background sentences to carry specific clinical claims; frequency bands largely unsupported.
Cadmium_Poisoning	Weakest	Most claim–evidence mismatches; systematic evidence_source misclassification; unevidenced quantitative thresholds; an unevidenced clinical contraindication; one questionable source.
Mercury_Poisoning	Thinnest	9 evidence items from 3 unique references; declared scope not delivered; an unevidenced negative epidemiological claim.

---

Findings

1. Unevidenced clinical contraindication — Cadmium_Poisoning

treatments[Chelation Therapy].description (kb/disorders/Cadmium_Poisoning.yaml:1042):

"BAL (dimercaprol) is contraindicated as it may increase renal cadmium uptake."

This is an actionable, directive clinical claim. Neither attached evidence item mentions dimercaprol or renal redistribution — one covers chelating agents generically, the other covers long-term chelation toxicity. The claim is well-established in the toxicology literature, so it is sourceable; it just is not sourced here. Highest-priority fix in the set, because an unsourced contraindication is the kind of claim a reader will act on.

2. Mechanism claims resting on clinical-observation evidence — Cadmium_Poisoning

pathophysiology[Direct Osteoblast Toxicity] (:403) asserts four mechanistic propositions: cadmium directly inhibits osteoblast differentiation, does so independently of the renal phosphate wasting pathway, inhibits alkaline phosphatase, and promotes osteoclast-mediated resorption. The sole evidence is a case report snippet: "We will present a case of cadmium induced peripheral neuropathy, nephropathy, and decreased bone density."

A single case of decreased bone density in a patient who also had nephropathy cannot establish direct osteoblast toxicity, and specifically cannot establish independence from the renal pathway — the case has renal disease. The node also carries biological_processes: osteoblast differentiation / DECREASED (GO:0001649), a structured assertion with no in-vitro or animal osteoblast evidence behind it. Tagged supports: SUPPORT; should be PARTIAL at most, and the node needs cell-level evidence.

3. Histopathology nodes evidenced by non-histopathological sources — Cadmium_Poisoning

Three of five histopathology nodes assert specific microscopic findings that none of their snippets document:

• Diffuse Alveolar Damage — claims hyaline membrane formation, alveolar edema, type II pneumocyte hyperplasia. Snippets: "multiple organ damage was observed, involving brain, lung, liver, kidney…" and a generic intro sentence ("cadmium is more commonly known to cause acute lung injury"). Neither describes lung histology.
• Renal Tubulointerstitial Disease and Fibrosis — the first snippet is again the generic "multiple organ damage" line. (A second citation does support the claim.)
• Hepatocellular Degeneration — claims degeneration and necrosis; the snippets document cadmium accumulation in liver, plus a NHANES epidemiological finding of hepatic fibrosis. Citing a cross-sectional biomarker study as histopathology evidence is a category error.

4. Unevidenced quantitative specifics carried in description fields — Cadmium_Poisoning

Descriptions assert precise numbers no snippet supports:

Claim	Location	Snippet actually says
"Inhaled cadmium has 25-50% bioavailability; oral absorption 3-8%"	pathophysiology[0]	"long biological half-life"
"renal cortex concentrations above 200 mcg/g"	pathophysiology[3]	nothing quantitative
"Cd-MT… small molecular weight (~7 kDa)"	pathophysiology[2]	nothing quantitative
"Normal levels typically below 5 mcg/L; above 50 mcg/L indicate significant toxicity"	diagnosis[0]	a generic list of diagnostic modalities
"a single cigarette may contain 1-2 mcg cadmium; smokers have levels 4-5× non-smokers"	environmental[2]	"Common risk factors include smoking and alcohol consumption"

The last two matter most: a diagnostic threshold and an exposure magnitude are exactly the numbers a reader would reuse.

5. Systematic evidence_source misclassification — Cadmium_Poisoning, Thallium_Poisoning

Per CLAUDE.md, evidence_source classifies the cited publication's study type. Several mechanistic reviews and in-vitro/analytical papers are tagged HUMAN_CLINICAL:

Reference	What it is	Tagged	Should be
PMID:20204475 (Biometals)	review of cadmium transport in mammalian cells	HUMAN_CLINICAL	IN_VITRO / OTHER
PMID:20354761 (Biometals)	review, cadmium and the kidney	HUMAN_CLINICAL	OTHER
PMID:19106433 (Indian J Med Res)	review, nephrotoxicity of Cd & Pb	HUMAN_CLINICAL	OTHER
PMID:31704329 (Ecotox Environ Saf)	analytical identification of Cd-binding proteins in plasma	HUMAN_CLINICAL	IN_VITRO
PMID:41453694 (Toxicol Lett)	review, heavy-metal neurotoxicity	HUMAN_CLINICAL	OTHER
PMID:14579545 (Toxicol Rev)	review, thallium + Prussian blue	HUMAN_CLINICAL (all ~30 uses)	OTHER

Note the contrast: Silicosis and Asbestosis get this right and even annotate it in prose ("Evidence source OTHER as this is a review article"), and Arsenic_Poisoning correctly splits one paper's animal and enzyme data into separate items. The fix is mechanical and the house style already exists.

6. Questionable source underpinning a diagnostic recommendation — Cadmium_Poisoning

biochemical[Urinary Cadmium Level] and diagnosis[Urine Cadmium Level] both rest on PMID:19364190, "The benefits of pre- and post-challenge urine heavy metal testing: Part 1", published in Alternative Medicine Review, quoted as: "Conducting pre-flush testing is also currently the clinician's only means of identifying cadmium toxicity."

Two problems. The source is not a mainstream clinical-toxicology venue, and provoked/post-challenge urine metal testing is specifically discouraged by mainstream toxicology practice. The diagnosis node's description propagates this ("Post-chelation challenge testing reflects total body stores and helps guide treatment decisions"). The underlying point — that unstimulated urinary cadmium indexes body burden — is uncontroversial and better sourced elsewhere in the same file; the challenge-testing framing should go.

7. Snippet whose sentence points the other way — Organophosphate_Poisoning

phenotypes[Miosis] (:454) asserts miosis as "a classic sign of cholinergic toxicity", tagged supports: SUPPORT. The snippet:

"The classic muscarinic and nicotinic signs of intoxication including increased secretions, bradycardia, fasciculations, and miosis were less common in our patient population."

The sentence names miosis among classic signs but its assertion is a negative finding for that cohort. Using it as unqualified SUPPORT for a phenotype claim inverts the source's point; a better snippet exists in the same corpus (PMID:32626615 lists miosis among OP poisoning signs and is already cited for bradycardia and fasciculations).

8. Description asserts a nerve-fiber pattern its evidence contradicts — Lead_Poisoning

phenotypes[Peripheral neuropathy] (:636) — "Lead neuropathy is classically motor-predominant and may present with wrist drop or radial palsy":

• Evidence 1 (PMID:17405745) describes a "sensitive polyneuropathy to the four limbs" — a sensory neuropathy, i.e. the opposite pattern.
• Evidence 2 (PMID:20142857) is the right paper (radial neuropathy from occupational lead exposure, five patients) but the chosen snippet is the content-free "Neuropathy is one complication of lead poisoning."

The motor-predominance and wrist-drop claims are correct and the second paper supports them — the snippet just needs to be re-drawn from that abstract.

9. Generic background sentences carrying specific clinical claims — Lead_Poisoning

PMID:37478813 is a laboratory ICP-MS methods/trends study. Its introductory sentence — "Exposure to lead may cause severe adverse effects such as anemia, neurologic damage, developmental disorders, and reproductive disorders" — is used as HUMAN_CLINICAL evidence for both phenotypes[Anemia] and phenotypes[Developmental delay]. Likewise PMID:40981357's single sentence about contamination sources ("mining activities, battery manufacturing, electronic waste recycling…") is reused verbatim for three separate environmental nodes. Not false, but a review's throwaway framing sentence is thin support for a curated phenotype, and better primary sources are already in the file.

10. Frequency bands mostly unsupported — 4 entries

Per docs/frequency-evidence-guidelines.md, a frequency: value is a separate quantitative claim needing its own evidence. Audit of all 36 bands in the four entries that use them:

Entry	Bands	Bands whose snippets contain no quantitative or explicit qualitative frequency statement
Cadmium_Poisoning	9	8
Lead_Poisoning	12	9
Carbon_Monoxide_Poisoning	10	5
Thallium_Poisoning	5	2
Total	36	24 (67%)

Worst cases are VERY_FREQUENT (schema: 80-99% of patients) claims in Cadmium_Poisoning supported only by single case reports — Renal Tubular Dysfunction and Low-Molecular-Weight Proteinuria are cited to a one-patient itai-itai report and a one-patient Fanconi report. The domain claim is right (LMW proteinuria is near-universal in chronic cadmium nephropathy); the citation cannot carry the band. Per the guidelines, omit the band or cite a prevalence figure.

The other five entries use no frequency: values at all, which is the compliant choice when a quantitative source is not at hand.

11. Declared scope not delivered; unevidenced negative claim — Mercury_Poisoning

The entry is the thinnest in the set: 228 lines, 9 evidence items, 3 unique references, and only pathophysiology, environmental, and phenotypes sections — no treatments, diagnosis, biochemical, histopathology, prevalence, or genetic. The absence of treatments is conspicuous for a metal poisoning where chelation is a standard intervention.

Two specific claim–evidence problems:

• Scope mismatch. The description commits to "elemental, inorganic, or organic forms, each with a distinct clinical profile" and asserts that "inorganic and elemental mercury, by contrast, classically cause peripheral neuropathy and the neuropsychiatric syndrome of erethism." The pathograph, phenotypes, and environmental sections model only the methylmercury / Minamata arm. The inorganic arm appears solely as one passing snippet about the felt-hat industry. Either the entry should be scoped to methylmercury poisoning or the inorganic arm needs curating.
• Unevidenced negative claim. "Unlike lead, mercury exposure from dietary fish or dental amalgam has not been associated with amyotrophic lateral sclerosis" appears twice (:31, :137) with zero supporting citations anywhere in the file. A negative epidemiological claim needs a source as much as a positive one.

Also, pathophysiology[0] asserts three converging molecular mechanisms (glutathione/NPSH depletion, calcium dyshomeostasis, NMDA-receptor excitotoxicity) on the strength of one in-vitro rat-cortical-neuron memantine study; the explanation acknowledges this ("the same study implicates…"), but two of the three mechanisms have no snippet of their own.

12. Unevidenced items — Silicosis, Asbestosis

Both entries leave 9 top-level items without any evidence:

• Silicosis — all three has_subtypes (Chronic, Accelerated, Acute), which is where the latency and exposure-intensity claims live ("developing after 10-30 years of relatively low-level exposure", nodule size cutoffs); phenotypes Dyspnea, Cough, Respiratory Insufficiency; treatments Silica Exposure Cessation and Dust Control, Whole-Lung Lavage, Lung Transplantation.
• Asbestosis — two core mechanism nodes, Pro-fibrotic mediator release (asserts TGF-beta, PDGF, IGF-1, fibronectin as the inflammation→fibrosis bridge) and Excessive extracellular matrix deposition; phenotypes Chronic cough, Bibasilar inspiratory crackles, Abnormal pulmonary interstitial morphology, Digital clubbing; treatments Supplemental oxygen therapy, Lung transplantation, Vaccination.

Unevidenced mechanism nodes are the more serious of the two, since the pathograph edges through them are load-bearing. For the clinical items, per CLAUDE.md the options are a quotable source or moving the claim to notes.

13. Minor

• Thallium_Poisoning pathophysiology[3].downstream[0] has no explanation (every other evidence item in the file has one).
• Cadmium_Poisoning pathophysiology[Hepatic Glutathione Depletion] asserts glutathione depletion, but its snippets report reduced GSH-Px enzyme activity (not glutathione), and one (PMID:41188353) states cadmium "triggered a hepatic antioxidant response" — arguably the opposite direction. A second citation (PMID:40164036) reports the exercise arm's effect and is being read backwards to infer cadmium's effect.
• Cadmium_Poisoning biochemical[Hepatic Transaminases (ALT/AST)] is a human laboratory readout sourced only to a mouse exercise study.
• Cadmium_Poisoning reuses one systematic-review snippet (a generic list of diagnostic modalities) across five separate diagnosis nodes.

Recommended fixes, in priority order

1. Source or remove the BAL contraindication in Cadmium_Poisoning (finding 1).
2. Drop the PMID:19364190 challenge-testing framing from Cadmium_Poisoning diagnosis/biochemical; reanchor on the existing mainstream citations (finding 6).
3. Re-tag the misclassified evidence_source values (finding 5) — mechanical, and house style already exists in Silicosis/Asbestosis.
4. Downgrade Cadmium_Poisoning Direct Osteoblast Toxicity to PARTIAL and either add in-vitro osteoblast evidence or drop the independence-from-renal claim and the DECREASED process modifier (finding 2).
5. Strip or source the unevidenced quantitative thresholds in Cadmium_Poisoning descriptions (finding 4).
6. Re-draw the Lead_Poisoning neuropathy snippet from PMID:20142857 (finding 8); swap the Organophosphate_Poisoning miosis snippet to PMID:32626615 (finding 7).
7. Audit the 24 unsupported frequency: bands; omit where no quantitative source exists (finding 10).
8. Fix Cadmium_Poisoning histopathology nodes — either add real histology citations or relax the claims to what the autopsy snippets say (finding 3).
9. Decide Mercury_Poisoning's scope, remove or source the ALS negative claim, and add the missing treatments/diagnosis sections (finding 11).
10. Add evidence to the two Asbestosis mechanism nodes and the three Silicosis subtypes (finding 12).

Reproducing this review

# schema
uv run linkml-validate --schema src/dismech/schema/dismech.yaml \
  --target-class Disease kb/disorders/Cadmium_Poisoning.yaml

# snippet fidelity + coverage: see scripts described in Method above
just validate-terms-file kb/disorders/Cadmium_Poisoning.yaml
just validate-references kb/disorders/Cadmium_Poisoning.yaml

Note: just validate-references reported Total checks: 0 on these files in this environment (several publisher PDF fetches returned HTTP 403 behind the proxy), so snippet fidelity here was established by direct normalized-substring comparison against references_cache/ bodies rather than by the validator.

---

Resolution (applied 2026-07-25)

All ten recommended fixes were applied in the same branch. Post-fix state:

• Schema: 10/10 No issues found.
• Snippet fidelity: 602/602 snippets exact substrings (578 before; net +24 evidence items added, none removed except the two retired PMID:19364190 items).
• Frequency bands: 36 → 7. A follow-up audit of the 13 bands left by the first pass found that only some were genuinely supported, so a second sweep dropped 6 more (Lead Anemia; CO Syncope, Coma, Chest pain; Cadmium Acute Respiratory Distress Syndrome; Thallium Constipation) whose snippets carried no frequency signal, and re-banded 3 that contradicted their own snippet (Lead Hypertension and Gout OCCASIONAL → FREQUENT, per the source's "frequently presents"; CO Headache VERY_FREQUENT → FREQUENT). Of the 7 bands that remain: 4 are quantitatively supported (Pattern A — CO Cognitive impairment 15–40%, CO Myocardial injury ~one-third of moderate-to-severe cases, Thallium Alopecia 68%, Thallium Abdominal pain 51%), 2 rest on a qualitative literature term mapped to a band (Pattern C — Lead Hypertension, Gout), and 1 is a documented clinical estimate (Pattern D — CO Headache). Every remaining band names its pattern in the evidence explanation.
• Unevidenced top-level items: 21 → 12, with all mechanism nodes and all disease subtypes now evidenced.

Four new references were fetched via the cache pipeline (never hand-written): PMID:6734559, PMID:25191413, PMID:30053129, PMID:30558238.

#	Finding	Resolution
1	BAL contraindication unevidenced	Rewritten and sourced to PMID:6734559. The blanket "is contraindicated" is replaced with what the evidence actually shows: BAL given ~30 min after cadmium exposure increased renal cadmium deposition in rats, but the same study found no renal increase when BAL was given at 24 h after metallothionein induction. Both directions cited, supports: PARTIAL, evidence_source: MODEL_ORGANISM, and the text now states plainly this is an animal-derived caution rather than a demonstrated human contraindication.
2	Challenge-testing source	PMID:19364190 (Altern Med Rev) removed entirely from both the biochemical and diagnosis nodes. Reanchored on PMID:20354761 (LMW proteinuria as the sensitive screening measure) and PMID:41000307. Both nodes now state explicitly that provoked/post-chelation urine testing is not accepted practice and has no validated thresholds.
3	evidence_source misclassification	49 values re-tagged — 17 in Cadmium (PMID:20204475, 20354761, 19106433, 31704329, 41453694), 32 in Thallium (PMID:14579545). Reviews → OTHER, cell-transport/analytical papers → IN_VITRO.
4	Direct Osteoblast Toxicity	Downgraded to supports: PARTIAL; the unsupported biological_processes: osteoblast differentiation / DECREASED assertion removed; description rewritten to say the direct arm is not established by the curated evidence and that the independence-from-renal claim is unsupported. New KNOWLEDGE_GAP discussion (gap_cadmium_direct_osteoblast_toxicity) with two proposed experiments — osteoblast/osteoclast culture, and an animal comparison with vs without renal phosphate wasting.
5	Unevidenced quantitative specifics	All five removed (bioavailability percentages, 200 mcg/g cortex threshold, ~7 kDa, the 5/50 mcg/L blood cut-offs in both biochemical and diagnosis, and the per-cigarette / smoker-ratio figures). Replaced with qualitative statements plus an explicit note that numeric cut-offs are laboratory- and population-specific and are deliberately not asserted.
6	Two misleading snippets	Lead phenotypes[Peripheral neuropathy]: snippet re-drawn from PMID:20142857 to the sentence describing five battery-factory workers with electrophysiologically characterised radial neuropathy; the description now says motor-predominance is a tendency not an exclusive pattern, and the sensory-polyneuropathy case's explanation flags the discrepancy rather than hiding it. OP phenotypes[Miosis]: the "were less common in our patient population" snippet replaced with the positive sign list from PMID:32626615.
7	Unsupported frequency bands	23 bands dropped in the first pass (9 Lead, 8 Cadmium, 4 CO, 2 Thallium), including both Cadmium VERY_FREQUENT bands that rested on single case reports. A second sweep (see the Resolution note above) dropped 6 more and re-banded 3, leaving 7.
8	Histopathology nodes	Claims relaxed to what the sources show and the three "multiple organ damage" citations downgraded to PARTIAL with corrected explanations. Descriptions now state which features (hyaline membranes, type II pneumocyte hyperplasia, hepatocellular necrosis) are the expected pattern rather than an observed finding in the curated sources.
9	Mercury scope, ALS claim, missing sections	ALS negative claim sourced to PMID:30558238, with the explanation noting the design limit (online self-report case-control, 401 vs 452) — absence of an observed association, not proof of no effect. Added diagnosis (blood/hair/toenail mercury, PARTIAL, carrying the source's own caveat that these matrices may not reflect remote injury; plus umbilical-cord methylmercury for prenatal exposure) and treatments (exposure cessation; supportive care, cited to "There is no effective treatment."). Added a SCOPE NOTE to the description and a KNOWLEDGE_GAP (gap_mercury_inorganic_arm_not_curated) recording that the elemental/inorganic arm is named but not modelled. Mercury evidence items 9 → 15.
10	Silicosis subtypes / Asbestosis mechanism nodes	All three Silicosis subtypes evidenced with PMID:25191413 (the three-way acute/chronic/accelerated classification by exposure intensity and symptom onset) and PMID:30053129 (accelerated silicosis after ~2 years of intense sandblasting). Both Asbestosis mechanism nodes evidenced: Pro-fibrotic mediator release with PMID:12444030 (pulmonary TGF-beta1 alone is sufficient to produce fibroproliferative disease) and PMID:37569765; Excessive extracellular matrix deposition with PMID:38192052 and PMID:32553000. Both descriptions now flag which sub-claims come from the wider literature rather than the curated sources.

Minor findings also fixed: the missing Thallium explanation; the Cadmium hepatic-glutathione node's direction-of-effect overreach (now states the readout is antioxidant enzyme activity, not glutathione stores, and that one source reports an antioxidant response); and the mouse-only hepatic transaminase readout downgraded to PARTIAL.

Deliberately not fixed (recorded so the remaining gap is visible, not silent): 12 clinical phenotypes/treatments items still carry no evidence — 3 in Silicosis (Respiratory Insufficiency, Silica Exposure Cessation and Dust Control, Whole-Lung Lavage), 6 in Asbestosis (Chronic cough, Bibasilar inspiratory crackles, Abnormal pulmonary interstitial morphology, Digital clubbing, Lung transplantation, Vaccination), 2 in Carbon Monoxide, 1 in Methanol. These are uncontroversial clinical facts with no quotable sentence in the currently cached sources; per CLAUDE.md the options are a new citation or a move to notes, and neither was in scope for this pass.

Finding 9 in the list above (Lead's use of generic background sentences from a laboratory methods paper to carry the Anemia and Developmental delay phenotype claims) was also left in place: the claims are correct and the snippets are honest, they are simply thinner support than better primary sources already present in the file would give.