Blood Disorders Claim–Evidence Review (2026-07-25)

Correctness review of 10 haematology entries in kb/disorders/, focused on whether each asserted claim is actually carried by the evidence attached to it.

State this report describes: rebased onto main at e85b40a8. Findings were re-verified against that tree. One finding (#1, ITP) had been fixed upstream in the interim and is marked RESOLVED; the other fourteen still reproduce. just count-verified-snippets over the ten entries at this commit reports 362/372 verified (10 skipped by prefix) — see Mechanical validation for what the skipped ten are and why it matters.

Entries reviewed

Entry	Lines
`Sickle_Cell_Disease`	1118
`Hereditary_Spherocytosis`	1075
`Beta_Thalassemia`	979
`Polycythemia_Vera`	928
`Alpha_Thalassemia`	888
`Hemophilia_A`	650
`Diamond-Blackfan_Anemia`	630
`Immune_Thrombocytopenia`	316
`Hemophilia_B`	267
`Hereditary_von_Willebrand_Disease`	234

Mechanical validation: clean

linkml-validate (Disease target class): no issues across all 10.
linkml-reference-validator: no ERRORs (only 403/404 full-text fetch noise).
Snippet fidelity, via the repo's own tool (uv run python -m dismech.reference_snippet_audit, i.e. just count-verified-snippets) over the ten entries at this tip:

Snippets checked: 362/372 verified against cached references (10 skipped by prefix)

Zero mismatches — but "372" is not "every snippet in these files." Ten are never checked by anything: the DOI:-prefixed items in Polycythemia_Vera.yaml, because DOI sits in skip_prefixes in conf/reference_validator_config.yaml. Neither the reference validator nor the fast audit reads them. Those ten are exactly the deep-research boilerplate stubs that finding #15 flags as degenerate — so the items most at risk of being junk are the ones no tool verifies. One further reference has no cache file at all: url:https://www.fda.gov/... in Alpha_Thalassemia.treatments[6].

Full text is cached for 20 of the cited references; the remainder are abstract_only. Findings were first reached from abstracts, then re-checked against full text (see Full-text re-examination).

No fabricated quotes, no hallucinated PMIDs. Every finding below is semantic — the quote is real, but it does not carry the claim attached to it, or the claim contradicts something else in the same file.

Evidence-coverage profile

Entry	phenotypes w/o evidence	frequency bands w/o evidence	treatments w/o evidence	pathophys nodes w/o evidence	`genetic[]` w/o `gene_term`
Sickle_Cell_Disease	1/21	1	7/9	1/5	3/3
Beta_Thalassemia	17/18	17	6/8	0/6	4/4
Alpha_Thalassemia	0/12	0	2/7	0/5	2/2
Hemophilia_A	0/7	0	0/7	0/5	0/1
Hemophilia_B	0/2	0	0/4	0/1	0/1
Hereditary_von_Willebrand_Disease	0/3	0	0/2	0/1	—
Hereditary_Spherocytosis	0/28	0	3/4	0/3	0/5
Immune_Thrombocytopenia	~~3/4~~ → 0/4	~~3~~ → 0	~~5/5~~ → 0/6	0/4	—
Polycythemia_Vera	0/26	0	5/5	3/4	3/3
Diamond-Blackfan_Anemia	0/8	0	2/6	0/3	10/10

Hemophilia_A and Hemophilia_B are the reference-quality entries: every phenotype, treatment, and pathophysiology node is evidenced.

High-severity findings

1. `Immune_Thrombocytopenia` — the defining mechanism is marked `NO_EVIDENCE` but the explanation claims confirmation

RESOLVED UPSTREAM — fixed on main before this report merged. The contradictory explanation is gone, replaced by an honest one stating the snippet enumerates therapies "rather than directly evidencing autoantibody-mediated platelet destruction; retained for context". A proper supporting item was added: PMID:30801909, a meta-analysis naming "anti-glycoprotein IIbIIIa or anti-glycoprotein IbIX" — exactly the two glycoproteins in the node description. The rest of the entry was filled in too: all 4 phenotypes and all 6 treatments now carry evidence (was 3/4 and 5/5 unevidenced). Recorded for the pattern, not as outstanding work.

pathophysiology[0] "Antiplatelet Antibody Production" (the anti-GPIIb/IIIa → splenic Fc-mediated clearance mechanism, i.e. what makes ITP ITP) carries a single evidence item with supports: NO_EVIDENCE, whose snippet is a list of new targeted therapies, and whose explanation reads "This review confirms that autoantibody-mediated platelet destruction is central to ITP pathogenesis."

The cached PMID:38396839 abstract contains no such statement — it is entirely about management and prediction of treatment response. So the enum is honest and the explanation is not; either way the flagship node has zero supporting evidence. Replace with a mechanism paper (e.g. anti-GPIIb/IIIa autoantibody or splenic clearance literature) and drop the contradictory explanation.

2. `Alpha_Thalassemia` — prevalence record evidenced by a mutation-count sentence

prevalence[0] has no measure_type, no prevalence_class, and no rate_per_100000. Its notes assert "widespread in tropical and subtropical regions… Highest prevalence in Southeast Asia, southern China, the Mediterranean, Middle East, and Africa." The only evidence is PMID:25390741 with the snippet "More than 100 varieties of α-thalassemia have been identified."

That quote supports allelic heterogeneity, not prevalence and not geography. The cached NEJM record is a two-sentence stub, so the geographic claim cannot be sourced from it at all. Either cite a real epidemiology source with structured slots filled, or move the geography to description.

3. Subtype-restricted phenotypes given disease-level `VERY_FREQUENT` bands

Alpha_Thalassemia defines a four-tier severity spectrum whose top tier is explicitly clinically silent, then assigns disease-level VERY_FREQUENT to phenotypes its own notes scope to one subtype:

Hydrops Fetalis — VERY_FREQUENT, notes: Specific to Hb Bart syndrome (four-gene deletion)
Congestive Heart Failure — VERY_FREQUENT, notes: Specific to Hb Bart syndrome; occasional in severe HbH disease
Splenomegaly — VERY_FREQUENT, description says "Present in most individuals with HbH disease"
Hypochromic Microcytic Anemia — VERY_FREQUENT, description says severity "ranges from absent (silent carriers)"

Across all alpha-thalassemia, hydrops fetalis is rare. The schema already provides phenotypes[].subtype as a foreign key into has_subtypes[].name, and the treatments block in this same file uses notes to scope by subtype correctly — but 0 of 12 phenotypes use subtype:. Across all 10 entries reviewed, subtype: scoping is used zero times.

4. `Polycythemia_Vera` — three of four mechanism nodes and all five treatments are unevidenced

Revised after full-text review — see Full-text re-examination below. Downgraded: the evidence exists in PMID:40246933, already cited 11× in this file.

Unevidenced pathophysiology: JAK2 V617F Constitutive Activation, STAT5 Hyperactivation, Erythropoietin-Independent Erythropoiesis. The description asserts "JAK2 V617F… present in approximately 95% of PV cases" with no citation at that point (the claim is evidenced later under genetic[0], so this is a placement problem rather than a sourcing problem).

All five treatments — phlebotomy, aspirin, hydroxyurea, ruxolitinib, interferon-alpha — have no evidence, despite carrying specific claims ("maintain hematocrit below 45%", "approved for PV inadequately controlled by hydroxyurea"). The relevant trials are already sitting in the file's references: block (MAJIC-PV DOI:10.1200/jco.22.01935, RuxoBEAT, ropeginterferon); they need promoting to evidence items.

genetic[1] TET2 and genetic[2] ASXL1 state 18% and 15% with no evidence, though both figures appear verbatim in the Tefferi 2024 review already listed in references:.

5. Uncritical bulk import of Orphanet HPO frequencies

Revised after full-text review — see Full-text re-examination below. Strengthened: PV Splenomegaly is numerically contradicted (~30%) by the entry's own primary source.

Hereditary_Spherocytosis imports 28 phenotypes from ORPHA:822. The snippets are faithful (verified against the cache), but several annotations are not credible for HS, and each has been given an invented mechanistic description that no cited source supports:

Ataxia OCCASIONAL — "a rare neurological finding that may occur in severe cases"
Maculopapular Exanthema OCCASIONAL — "an occasional dermatological manifestation"
Hypofibrinogenemia FREQUENT (30–79%) — "potentially related to chronic hemolysis and coagulation factor consumption"
Hypercoagulability FREQUENT — "potentially related to membrane vesiculation and phosphatidylserine exposure"
Restrictive Cardiomyopathy OCCASIONAL
Muscle Weakness FREQUENT — "likely related to chronic anemia"

Also internally inconsistent: Spherocytosis is FREQUENT (30–79%) while Increased red cell osmotic fragility is VERY_FREQUENT — spherocytes on smear are the defining finding and cannot be rarer than a secondary assay abnormality. The entry's own description calls it "the characteristic finding that distinguishes HS from other hemolytic anemias."

Polycythemia_Vera has the same problem from ORPHA:729, and demonstrates that the curator already knew the source was unreliable: Myelofibrosis and Acute Leukemia were correctly overridden from Orphanet's Very frequent down to OCCASIONAL, with supports: PARTIAL and a note explaining that Orphanet likely encodes lifetime cumulative risk. That same scepticism was not applied to the other twelve Orphanet-derived bands in the file:

Epistaxis, Gingival Bleeding, Bruising Susceptibility at VERY_FREQUENT (80–99%)
Weight Loss, Hypertension, Tinnitus, Vertigo, Abdominal Pain, Hepatomegaly at VERY_FREQUENT
Splenomegaly at VERY_FREQUENT (≈30–40% at diagnosis in practice)

while Thrombocytosis and Leukocytosis sit at OCCASIONAL (5–29%) — contradicting both the entry's own description ("often accompanied by increased white blood cells and platelets") and its second evidence item on those very phenotypes (REVEAL: "characterized by erythrocytosis, thrombocytosis, leukocytosis, and splenomegaly"). The resulting profile ranks nosebleeds above the cardinal features of the disease.

6. `Beta_Thalassemia` — 17 of 18 phenotypes carry frequency bands with no evidence

Revised after full-text review — see Full-text re-examination below. Downgraded: the evidence is in PMID:20492708, already cited 9× — one sentence is already quoted in this file.

Only Microcytic Hypochromic Anemia is evidenced. Everything else — target cells, extramedullary haematopoiesis, splenomegaly, hepatomegaly, cholelithiasis, jaundice, elevated ferritin, cardiomyopathy, pulmonary hypertension, frontal bossing, osteoporosis, short stature, delayed puberty — has a FREQUENT/VERY_FREQUENT/ OCCASIONAL band and no evidence item at all. Per docs/frequency-evidence-guidelines.md these bands should be evidenced or omitted.

Cardiomyopathy additionally asserts "the leading cause of death in transfusion-dependent beta-thalassemia" uncited.

The pathophysiology and biochemical blocks in this same file are excellent (per-edge evidence, readouts with regulatory_endpoint_refs) — the gap is confined to phenotypes.

Medium-severity findings

7. `Diamond-Blackfan_Anemia` — p53/MDM2 node evidenced by a phenotype sentence

pathophysiology[1] "p53-Mediated Erythroid Apoptosis" describes free ribosomal proteins binding MDM2 and stabilising p53. Its only evidence is GeneReviews PMID:20301769 with the snippet "characterized by a profound normochromic and usually macrocytic anemia with normal leukocytes and platelets" — which supports erythroid-restricted failure but says nothing about p53, MDM2, or apoptosis. The canonical mechanism is unevidenced despite abundant literature.

8. `Diamond-Blackfan_Anemia` — `RPL35A` cited to a paper that found no RPL35A mutations

Revised after full-text review — see Full-text re-examination below. Moderated: RPL35A is a genuine DBA gene; the citation is secondary and a better snippet exists.

genetic[3] RPL35A cites PMID:19773262 quoting its BACKGROUND sentence. That paper's own RESULTS state "No mutations were found in RPS14, RPS16, or RPL35A." The quote is accurate, but this is the wrong paper to anchor RPL35A causation. (The neighbouring RPL5/RPL11 claims from the same paper are correct — the cohort is 92 RPS19-negative Italian probands, so the "20% of RPS19-negative" explanation checks out.)

9. `Diamond-Blackfan_Anemia` — "congenital malformations, 50%" bound to `Craniofacial dysostosis`

The phenotype is named Congenital Malformations, described as covering craniofacial, thumb, cardiac, and genitourinary anomalies, evidenced by "congenital malformations in up to 50% of affected individuals" — and bound to HP:0004439 Craniofacial dysostosis. The 50% figure is for all malformations combined; craniofacial dysostosis specifically is not a DBA feature. Split into specific terms (cleft palate, micrognathia, thumb anomalies) or bind to a broader parent.

Related: Triphalangeal Thumb is evidenced by a snippet about "hand malformations and RPL11 mutations" that never mentions thumbs, with the explanation supplying "including thumb anomalies."

10. `Sickle_Cell_Disease` — HBB has no evidence naming HBB

genetic[0] HBB (notes: Glu6Val mutation (rs334)) is evidenced by PMID:18667698 with "Sickle cell disease (SCD) is a debilitating monogenic blood disorder…" — correctly flagged PARTIAL, but the snippet never mentions HBB, β-globin, or Glu6Val. The single most fundamental genetic claim in the entry lacks a quote that names the gene.

11. `Sickle_Cell_Disease` — acute chest syndrome supported only by murine NET data

Acute Chest Syndrome (notes: Leading cause of death) is evidenced by two MODEL_ORGANISM snippets from PMID:24620350 (heme-induced neutrophil extracellular traps in mice), neither of which mentions acute chest syndrome; the explanations supply the extrapolation. CLAUDE.md: "Model organism evidence should not be the only support for human phenotypes." The same pattern affects Pain Crises, though there a second ORPHA:232 item rescues it.

Splenic Sequestration has no evidence at all.

12. `Sickle_Cell_Disease` — unevidenced regulatory-status claims

Seven of nine treatments have no treatment_term and no evidence. Two make specific falsifiable regulatory assertions with nothing attached:

Voxelotor: "voluntarily withdrawn from worldwide markets in 2024 due to postmarketing safety concerns"
Crizanlizumab: "EU authorization revocation was recommended in 2023 when STAND did not confirm benefit"

Both are accurate but need citations. Penicillin Prophylaxis lacks the PROPS trial.

Also prevalence[0] (GBD 2021, 7.74 M living cases) carries no measure_type, prevalence_class, or rate_per_100000 — the sole record in this file that was not migrated to structured slots.

13. `Hereditary_von_Willebrand_Disease` — imprecise disease term and no genetics

disease_term is MONDO:0024574 "von Willebrand disease (hereditary or acquired)" on an entry explicitly scoped to hereditary disease. MONDO:0019565 "hereditary von Willebrand disease" exists and is the precise match.
No genetic: block and no inheritance: block — VWF is never bound to HGNC anywhere in the file.
Subtypes list Type 1, Type 2, Type 2N, Type 3; 2N is a member of the Type 2 group, and 2A/2B/2M are absent, though the cited PMID:21289515 abstract enumerates all four.
prevalence[1] assigns BAND_1_5_PER_10000 to a range spanning <1 to 450 per million (rate_low: 0.1, rate_high: 45.0) — the band matches only the top of a range covering four orders of magnitude.

14. `Hemophilia_A` — same PMID tagged three different `evidence_source` values

PMID:26743572 is an in-vitro iPSC gene-correction study. It appears as HUMAN_CLINICAL (inheritance block), IN_VITRO (twice), and OTHER (genetic features). Per the CLAUDE.md classification rules it is IN_VITRO throughout.

Easy Bruising is VERY_FREQUENT on a PARTIAL generic-bleeding snippet; easy bruising is a platelet/VWD phenotype, not a hallmark of haemophilia A, whose signature is deep bleeding (haemarthrosis, muscle haematoma).

15. `Polycythemia_Vera` — deep-research boilerplate as evidence

The references[].findings[].evidence[] items are DR-generated stubs where snippet == statement == supporting_text and every explanation reads "Deep research cited this publication as relevant literature for Polycythemia Vera." Several are degenerate:

snippet: "The Swedish nationwide study by Leontyeva et al." marked supports: SUPPORT
snippet: "of review Development of hepcidin therapeutics has been…" — a truncated "Purpose of review" fragment

These are non-informative and should be dropped or converted to real evidence items.

Low-severity / conventions

genetic[] HGNC bindings missing: DBA 10/10, Beta_Thalassemia 4/4, Sickle_Cell_Disease 3/3, Polycythemia_Vera 3/3, Alpha_Thalassemia 2/2. In the thalassemias and SCD the genes are bound in pathophysiology[].genes, so the fix is mechanical. Hereditary_Spherocytosis (5/5 bound) is the model.
Subtype naming: Hereditary_Spherocytosis, Beta_Thalassemia, and Alpha_Thalassemia all use long parenthetical subtype name values ("Hemoglobin Bart Hydrops Fetalis Syndrome (Four Alpha-Globin Genes Deleted)") against the CLAUDE.md rule to keep name slug-friendly and put verbose labels in display_name. None carry subtype_term despite MONDO terms existing. Hemophilia_A (Severe/Moderate/Mild + subtype_term) is the model.
Unbound inheritance_term: Beta_Thalassemia (Autosomal recessive, no term, no evidence) and Alpha_Thalassemia (no term). Hemophilia_B has no inheritance: block at all despite being X-linked throughout.
Metadata: Beta_Thalassemia has no creation_date — only the deprecated updated_date. The deprecated percentage field is still populated on recently edited prevalence records in Hemophilia A/B and VWD.
Beta_Thalassemia node genes: Alpha-Globin Chain Excess lists HBA1/HBA2 in genes:. In β-thalassemia those loci are normal — the excess is stoichiometric. Listing them as node genes implies pathogenicity.
Unverifiable reference: Alpha_Thalassemia.treatments[6] cites url:https://www.fda.gov/... with no cache file; url: references sit outside the reference-validation stack entirely.
Thin fragment snippets in Alpha_Thalassemia: "Couples who are members of populations at risk", "hematologic and hemoglobin (Hb) findings", "chelation therapy for iron overload" — sentence stubs that carry no claim standing alone. Iron Overload is banded FREQUENT on the strength of "iron chelation therapy should be instituted."
Upstream data artifact: Sickle_Cell_Disease quotes "1-5 / 10 000 | Europe | Point prevalence | PMID:2019" from ORPHA:232 — a year mis-parsed as a PMID in the Orphanet source. Faithful quote, upstream ingestion bug.
Polycythemia_Vera has no prevalence: block, and its classifications.icdo_morphology: Leukemia is a forced approximation — PV is ICD-O 9950/3 (MPN block), and ICDOMorphologyEnum has no myeloproliferative value. A schema-enum gap rather than a curation error.

Full-text re-examination

The reference cache holds full text for 20 of the cited references (the rest are abstract_only). The findings above were reached from abstracts; re-reading the full texts changes four of them and leaves the rest intact. Sources behind findings #1, #2,

7, #9, #10, and #11 are all still `abstract_only`, so those are unaffected.

#4 downgraded — Polycythemia_Vera's evidence exists, in a paper already cited 11×

PMID:40246933 (Polycythaemia vera, Nat Rev Dis Primers) is cached in full (18,782 words) and supplies clean support for every unevidenced treatment and for the JAK2 node. This is a distribution problem, not a sourcing problem. Ready-to-use snippets:

Target	Snippet from PMID:40246933
`Hydroxyurea`, `Ruxolitinib`, `Interferon-alpha`	"Hydroxyurea or interferons remain the preferred first-line cytoreductive agents, with the JAK1 and JAK2 inhibitor, ruxolitinib, currently approved for the treatment of patients who are resistant to, or intolerant of, hydroxyurea."
`Therapeutic Phlebotomy`, `Low-Dose Aspirin`	"High-risk patients are those aged ≥60 years and/or with a history of thrombosis, and typically are eligible for cytoreductive therapy, in addition to therapeutic phlebotomy and low-dose aspirin."
`JAK2 V617F Constitutive Activation`, and the entry's uncited "~95% of PV cases"	"Case ascertainment in epidemiological studies has been refined by the demonstration in 2005 that most patients (~95%) have the acquired (somatic) mutation JAK2V617F, indicating clonal disease"
the node's pseudokinase description	"the other is a pseudokinase domain located upstream of the kinase domain that binds ATP, but has no or a very weak ability to directly phosphorylate substrates"

One sentence closes three treatments at once, and it matches the entry's ruxolitinib description almost word for word.

Two rows were withdrawn from this table after review, and the reason is the point of the section below. The original JAK2 row proposed "A unique clonal JAK2 mutation leading to constitutive signalling causes polycythaemia vera" — which is a bibliography entry at references_cache/PMID_40246933.md:1876-1877, the title of James et al. 2005 (PMID:15793561), not a claim by the Nat Rev review. Attributing it to PMID:40246933 is a misattribution, and quoting a title as a finding separately violates CLAUDE.md §6 ("A Title Is Not a Finding"). It would have passed count-verified-snippets cleanly. This report identified that exact hazard two sections down and then walked into it — the flattened full-text grep used to find candidates did not distinguish body from back matter. Cite PMID:15793561 directly if the 2005 claim is wanted; the replacement row above is genuine body text and covers the same ground.

A phlebotomy row quoting "Need for phlebotomy to keep haematocrit <45%" was also withdrawn: at PMID_40246933.md:1520 that is a bullet under "Hydroxyurea resistance after 3 months of treatment", so its subject is a resistance criterion, not a treatment goal. The high-risk-patients row already covers phlebotomy and aspirin without the caveat.

#5 strengthened — the Orphanet splenomegaly band is numerically contradicted

The original argument was clinical implausibility. The full text makes it concrete:

"palpable splenomegaly (present in about 30% of patients at diagnosis)"

The entry bands Splenomegaly as VERY_FREQUENT (80–99%) on Orphanet's authority. The review the entry already cites 11 times says ~30% — FREQUENT at best, right at the band boundary. So this is no longer a judgement call about Orphanet's reliability; it is a direct numerical conflict with the entry's own primary source.

#6 downgraded — Beta_Thalassemia's evidence is already in the file

PMID:20492708 (GeneReviews Beta-thalassemia) is cached in full (13,213 words) and already cited 9 times. Two sentences cover roughly ten of the seventeen unevidenced phenotypes:

"extramedullary hematopoiesis and its complications (osteoporosis, masses of erythropoietic tissue that primarily affect the spleen, liver, lymph nodes, chest and spine, and bone deformities and typical facial changes), gallstones, painful leg ulcers and increased predisposition to thrombosis."

covers Extramedullary Hematopoiesis, Osteoporosis, Splenomegaly, Hepatomegaly, Cholelithiasis, and Frontal Bossing. And:

"Findings in untreated or poorly transfused individuals with thalassemia major … are growth retardation, pallor, jaundice, poor musculature, hepatosplenomegaly, leg ulcers, development of masses from extramedullary hematopoiesis, and skeletal changes that result from expansion of the bone marrow."

covers Short Stature, Jaundice, and reinforces the rest — and is already quoted in this very file, under biochemical[4] Indirect Bilirubin. Plus:

"However, cardiac disease remains the main cause of death in patients with iron overload."

which directly evidences the previously uncited "leading cause of death in transfusion-dependent beta-thalassemia" claim on Cardiomyopathy, and:

"Cardiac involvement in thalassemia intermedia results mainly from a high-output state and pulmonary hypertension"

for Pulmonary Hypertension.

Caveat that survives: these sentences establish occurrence, not the frequency bands. Attaching them fixes the "no evidence at all" problem but not the VERY_FREQUENT/FREQUENT justification, which still needs quantitative sources or omission per docs/frequency-evidence-guidelines.md.

#8 moderated — RPL35A is a real DBA gene, the citation is just secondary

The full text of PMID:19773262 shows the paper treats RPL35A as established throughout, not only in its abstract's background. Its Introduction:

"The genetic basis of DBA is heterogeneous. Approximately 40% of patients have mutations in one of the genes for ribosomal proteins (RP): RPS7, RPS17, RPS19, RPS24, RPL5, RPL11, or RPL35A."

and its Discussion notes that genotype–phenotype data are unavailable for RPS24, RPS17, and RPL35A "because the number of subjects studied are too small" — i.e. the negative result is a cohort limitation, not a refutation. So the entry's substance is right; the citation is attributive rather than primary.

Better still, that Introduction sentence is a strictly better snippet than the one in use, and it would simultaneously evidence RPS7, RPS17, and RPS24, which currently have no evidence at all. Downgrade #8 from "wrong paper" to "secondary citation, better snippet available."

#9 — partially improved, sub-point stands

A better snippet exists for the malformation claim:

"Most of the patients with RPL5 (83%) and RPL11 (73%) mutations had physical malformations."

But the paper's own results prose still contains no thumb-specific claim, so binding Triphalangeal Thumb to a hand-malformation quote remains an over-read. Thumb data appear only in per-patient table rows, which are not clean snippet material.

One live example of the full-text quoting risk

The string "Ribosomal protein L5 and L11 mutations are associated with cleft palate and abnormal thumbs in Diamond-Blackfan anemia patients" appears in the PMID:19773262 cache and would validate as a snippet — but it is a title in the bibliography, not a claim by that paper. Worth a lint rule: flag snippets that match only inside a reference list.

Suggested priority

~~Fix the NO_EVIDENCE/explanation contradiction on the ITP antibody node (#1).~~ Done upstream — resolved on main with PMID:30801909 before this report merged.
Re-band or subtype:-scope the Alpha_Thalassemia phenotypes (#3).
Cheap and mechanical, now that full text is cached: attach the Polycythemia_Vera treatment/JAK2 snippets and the Beta_Thalassemia phenotype snippets tabulated above (#4, #6). No new literature search required.
Re-band PV Splenomegaly down from VERY_FREQUENT — contradicted at ~30% by the entry's own primary source (#5) — then audit the remaining Orphanet bands in PV and Hereditary_Spherocytosis and drop the invented mechanistic descriptions.
Swap the DBA RPL35A snippet for the Introduction sentence and reuse it for RPS7/RPS17/RPS24 (#8).
Mechanical sweep: gene_term bindings, inheritance_term bindings, subtype naming, creation_date.

Reproducing

# per-edit loop (seconds, offline)
just validate kb/disorders/<file>.yaml
just count-verified-snippets kb/disorders/<file>.yaml

# pre-PR sweep over all changed files at once — what CI runs
just validate-disorders kb/disorders/<file>.yaml ...

The snippet counts quoted above come from just count-verified-snippets over all ten entries in one invocation. Avoid per-file just validate-references: CLAUDE.md records it at 65 minutes for a single entry, and just validate-terms-file — cited in an earlier draft of this section — is not a recipe at all (just validate-terms <file> is).

Blood Disorders Claim–Evidence Review (2026-07-25)

Entries reviewed

Mechanical validation: clean

Evidence-coverage profile

High-severity findings

1. Immune_Thrombocytopenia — the defining mechanism is marked NO_EVIDENCE but the explanation claims confirmation

2. Alpha_Thalassemia — prevalence record evidenced by a mutation-count sentence

3. Subtype-restricted phenotypes given disease-level VERY_FREQUENT bands

4. Polycythemia_Vera — three of four mechanism nodes and all five treatments are unevidenced