Skip to content

Eye Disorder Claim–Evidence Review (2026-07-25)

Correctness review of 10 eye-disorder entries, focused on claim–evidence match: does each cited snippet actually support the claim it is attached to, at the strength the supports: grade asserts?

Scope

Ten entries spanning the main mechanistic classes of ocular disease (evidence-item counts as of the current branch head, after the fixes described below):

Entry Class Evidence items
Age_Related_Macular_Degeneration Complex/degenerative 68
Achromatopsia Inherited cone dysfunction 105
Fuchs_Endothelial_Corneal_Dystrophy Corneal dystrophy 58
Diabetic_Retinopathy Systemic microvascular 34
RHO-Related_Retinopathy Inherited retinal dystrophy 42
Glaucoma Optic neuropathy 27
Cytomegalovirus_Retinitis Infectious 17
Central_Retinal_Artery_Occlusion Vascular/ischemic 16
Retinopathy_of_Prematurity Developmental/vascular 14
Retinoblastoma Neoplastic 5

Automated validation: all clean

All ten pass the full stack — linkml-validate (schema), linkml-term-validator (ontology IDs + labels), and linkml-reference-validator (snippet-in-source substring matching). The reference validator was control-tested by corrupting a snippet in a scratch copy of Retinoblastoma.yaml; it correctly flagged the mismatch, confirming the "all validations passed" result is meaningful and not a silent no-op.

That control test earned its keep, and the reason is worth recording precisely. On a clean run against these entries the reference validator reports Total checks: 0 / All validations passed — which reads like a no-op and was flagged as one during PR review. It is not: re-running the control test on the current branch (corrupt one snippet in RHO-Related_Retinopathy.yaml, run, revert) produces Total checks: 1 and Text part not found as substring against the right PMID. So detection works; what is broken is the reported count, which appears to tally only failures. The practical consequence is the same either way — Total checks: 0 carries no information about how many snippets were actually compared, so a green line from this validator cannot by itself substantiate a "snippets verified" claim. That asymmetry is a tooling bug worth its own issue, separate from any content here.

Two DOI-based references in this entry set fail to download (403), so they are not checked at all — a second, quieter reason the count is not a coverage measure.

Accordingly, snippet verification behind the later rounds of fixes was done by direct whitespace-normalized substring comparison against references_cache/ (130 snippets across the five edited entries, zero mismatches) in addition to the validator, with the control test re-run each round rather than assumed.

Every finding below is therefore invisible to CI. These are semantic claim–evidence defects: the quote is genuine and correctly transcribed, but it does not establish what the claim asserts. This is the failure mode the validation stack structurally cannot catch.

Tier 1 — Fixed in this branch

Items 1–2 are outright factual errors. Items 3–4 were promoted here from Tier 2 during PR review once it was clear they needed no new source — only a re-quote and a grade correction. Items 5–6 were promoted in a second review round, on the principle that a nav-linked document asserting a grade is wrong should not ship alongside that grade: the regrades need no new literature, so parking them was not defensible.

The Negative electroretinogram waveform phenotype asserted that gain-of-function RHO CSNB produces:

a characteristic negative ERG waveform — preserved a-wave with reduced b-wave — reflecting inner-retinal signal disruption

This inverts the electrophysiology. CSNB splits into two ERG classes by the level of the lesion:

  • Schubert–Bornschein ("negative") ERG — large a-wave, minimal b-wave. Arises from a photoreceptor-to-bipolar-cell transmission defect (NYX, CACNA1F).
  • Riggs-type ERG — loss of the rod a-wave and b-wave. Arises from a phototransduction abnormality.

RHO G90D/T94I CSNB is a phototransduction lesion — the entry's own upstream node says so ("constitutive activity in darkness causing rod desensitization"). It is therefore Riggs-type, and the node contradicted the mechanism it claims to be a readout of.

Per PubMed, Marmor & Zeitz confirm the dichotomy explicitly (DOI, PMID:30051303): "CSNB from abnormalities in phototransduction can be recessive or dominant and is much less common. This produces a Riggs type of ERG with loss of the rod a-wave as well as the b-wave." See also DOI (PMID:33369259), which classes phototransduction-dysfunction CSNB as Riggs type.

Fixed: node renamed to Riggs-type electroretinogram, description and reports_on.interpretation corrected, and PMID:30051303 fetched and cited for the class-level ERG dichotomy. The companion G90D citation (PMID:38743626) originally quoted a generic CSNB definitional sentence; during PR review it was swapped for the RHO-specific bridging sentence already present in the same cached abstract ("One well-studied rhodopsin point mutant, G90D-Rho, is thought to cause CSNB because of its constitutive activity in darkness causing rod desensitization"), so the disease-level step is now sourced rather than inferred. That division of labour matters: PMID:30051303 is a GNAT1 report and establishes only the class dichotomy; without the swap, CSNBAD1's membership in the phototransduction class would have been curator inference carrying a SUPPORT grade.

Note the original snippet — "profound loss of rod sensitivity without severe retinal degeneration" — is a true quote that says nothing about a-wave or b-wave. The wrong claim rode entirely in the explanation field, which no validator reads.

2. Retinoblastoma — pediatric restriction dropped from the source claim

histopathology[0].description read "Retinoblastoma is the most common intraocular malignancy." The source says "the most common intraocular malignancy in children" — and the snippet had been truncated at exactly "...malignancy in", which passes substring validation while cutting the qualifier that makes the claim true. In adults, uveal melanoma is the most common primary intraocular malignancy.

Fixed: description scoped to children, snippet extended to the full sentence (Retinoblastoma is the most common intraocular malignancy in children.), and evidence_source: HUMAN_CLINICAL added. The adult contrast — that uveal melanoma leads in adults — is not in the description: PR review correctly pointed out that the first attempt at this fix stated it there, which made it an uncited claim sitting above an evidence: block that does not cover it. It now lives in the notes: slot on the HistopathologyFinding, explicitly labelled as an uncited orienting note, per the CLAUDE.md "When Evidence Cannot Be Verified" rule.

This is worth naming as a pattern: truncating a snippet mid-clause to make it match is a red flag, because the dropped words are often the ones carrying the scope limit.

pathophysiology[2].downstream[1] claimed "median age to mild visual acuity impairment is 72 years" while quoting only the dangling tail of the sentence: "whereas this could not be computed for lower acuities." The full sentence carrying the figure was already correctly quoted elsewhere in the same file (phenotypes[3], same PMID:32301896), so no new source was needed.

Fixed: snippet replaced with the complete sentence; explanation updated to state what it now quantifies. This is itself an instance of the dangling-clause pattern that recommendation #2 below proposes linting for.

4. Retinopathy_of_Prematurity — evidence graded against its own direction

treatments[1].evidence[0] was graded SUPPORT for preferring anti-VEGF over laser, but the quoted result points the other way (RR 2.14, 95% CI 1.06–4.33 — roughly twice the recurrence risk of laser). The explanation claimed the meta-analysis "quantifies anti-VEGF efficacy"; it quantifies a disadvantage.

Fixed: regraded PARTIAL, explanation rewritten as the trade-off it actually is. The residual unsourced claim in that treatment's description is a separate, still-open item — see Tier 2 below.

Both CSNBAD1 phenotype nodes (Riggs-type electroretinogram, Congenital night blindness) rested solely on PMID:38743626, a G90D knock-in mouse. The companion PMID:30051303 is human but reports a GNAT1 family, and its quoted sentence is a class-level statement about phototransduction-defect CSNB — it establishes the dichotomy, not the RHO patient ERG. CLAUDE.md is explicit that model-organism evidence should not be the only support for a human phenotype, and for the RHO-specific step it was.

Fixed: two human references fetched and added.

  • PMID:33669941 (Kobal 2021; 15 p.G90D patients from three families) states directly that RHO-related CSNB is of the Riggs type with loss of rod-specific ERG activity, a reduced dark-adapted a-wave and low b-wave, and largely preserved cone responses — the human counterpart of the mouse item, on exactly the a-wave/b-wave point.
  • PMID:9888392 (al-Jandal 1999) sources the T94I half of the subtype description from an Irish family segregating that variant, which previously had no human genetic support.

The same cohort also supplied derived counts for the two VERY_FREQUENT bands that were previously unsourced assertions (3/3 CSNB patients with the typical electrophysiology; 15/15 with lifelong non-worsening night blindness) — the "derived counts" pattern in docs/frequency-evidence-guidelines.md rather than a qualitative-term mapping.

And it partly contradicted the entry, which is the more interesting outcome. In that cohort only 20% of p.G90D carriers were classified CSNB while 53.3% developed classic RP, and the authors caution against diagnosing CSNB in p.G90D carriers without long follow-up into adulthood. The Congenital night blindness description no longer asserts a flatly non-degenerative course; the caveat is recorded in notes:, scoped to p.G90D (T94I, A292E, A295V are not associated with progression). The CSNBAD1 subtype description is left as the nosological definition. Going looking for human evidence to satisfy a discipline rule turned up a phenotype-spectrum correction nobody had asked for.

6. Six grade-only corrections applied rather than parked

Every item below was documented in Tier 2 as a wrong supports: grade needing no new source. Each is now regraded with an explanation stating what the snippet does and does not license:

Entry Location Was Now
Glaucoma pathophysiology[3] TM dysfunction (myocilin evidence) SUPPORT PARTIAL
Glaucoma treatments[2] alpha agonist non-difference result SUPPORT PARTIAL
Diabetic_Retinopathy has_subtypes[1] severe-NPDR staging SUPPORT PARTIAL
Diabetic_Retinopathy treatments[1] PRP cost-effectiveness SUPPORT PARTIAL
Diabetic_Retinopathy treatments[2] vitrectomy SUPPORT PARTIAL
Central_Retinal_Artery_Occlusion pathophysiology[1] ischemic-injury mechanism SUPPORT PARTIAL

CRAO is the one where a replacement source is genuinely still needed — the cached abstract contains no mechanistic content at all — so its explanation now says so outright instead of leaving SUPPORT standing while the gap is unfilled.

Also in this round: the Retinopathy_of_Prematurity FREQUENT band resting on a denominator was dropped, with a notes: line recording why so it is not re-added from the same snippet; and the two absent evidence_source values on the Glaucoma items were filled (PMID:10617907 → IN_VITRO, in situ hybridization on ex vivo specimens; PMID:37217093 → OTHER, a narrative review stating consensus rather than primary data).

7. The remaining regrade-only items, applied

A later review round observed that this report was still shipping alongside seven defects it documents as fixable without new literature — the same contradiction item 6 addressed. They are now closed:

Entry Location Was Now
RHO-Related_Retinopathy RP-definition snippet ×4 (night blindness edge, phenotypes[0], [2], [4]) SUPPORT PARTIAL
RHO-Related_Retinopathy phenotypes[4] bone-spicule pigmentation — +SUPPORT from PMID:33669941
RHO-Related_Retinopathy phenotypes[5] rod ERG SUPPORT on a BCVA snippet +SUPPORT from PMID:33669941; BCVA item → PARTIAL
Diabetic_Retinopathy phenotypes[2] retinal hemorrhage (wrong compartment) SUPPORT NO_EVIDENCE
Diabetic_Retinopathy treatments[2] vitrectomy PARTIAL NO_EVIDENCE
Central_Retinal_Artery_Occlusion phenotypes[1] acuity (meta-analysis aim) SUPPORT PARTIAL
Retinopathy_of_Prematurity phenotypes[4] myopia SUPPORT + FREQUENT PARTIAL, band dropped
Glaucoma genetic[1] OPTN typing + 4 missing evidence_source Risk Factor, untagged Causative, tagged, item PARTIAL

One flagged item was not changed after checking it: the fourth PMID:29776671 citation in Diabetic_Retinopathy quotes a different, apt sentence on a Visual Impairment node. See that section for why reference reuse and sentence reuse are different things.

Bulleted items below carry an explicit status marker — [Open], [Fixed], or a split marker where a grade was corrected but the underlying source gap remains (e.g. [Fixed grade / Open source]). Where an item is written as narrative prose rather than a bullet (the Diabetic_Retinopathy and RHO-Related_Retinopathy sections), the status is stated inline in bold instead. Read the status, not the tense: this section drifted out of sync with the KB in four consecutive review rounds while the fixes landed, which is what recommendation 8 is about — and the mixed markup here is itself the residue of that drift, since the narrative sections predate the marker convention.

What remains here needs a replacement source or new literature; the items that needed only a re-quote or a grade change have all been promoted to Tier 1 across two review rounds (items 3–4, then the six regrades in item 6 and the human-evidence gap in item 5). The pattern worth carrying forward: "needs new literature" is a claim to check before it becomes a deferral. In four separate cases here the source was already in the repository — the 72-year sentence quoted elsewhere in the same file, the G90D bridging sentence sitting in the same cached abstract — and in six more the fix needed no source at all. Where a remaining item is similarly cheap, that is noted inline.

Diabetic_Retinopathy — one guideline reference propping up three unrelated claims

PMID:29776671 (ICO Diabetic Eye Care guidelines) is cited three times with generic scope text that supports none of the attached claims:

Location Claim Snippet
has_subtypes[1] Severe NPDR 4-2-1 rule; "~50% progression to PDR within 1 year" "Vision loss from DR can be prevented with broad-level public health strategies…"
treatments[1] PRP is a cost-effective treatment same snippet — never mentions laser
treatments[2] Vitrectomy is among the options "appropriate management of vision-threatening DR…" — never mentions vitrectomy

All three were graded SUPPORT with each explanation asserting guideline content absent from the quoted text. The grades are now PARTIAL (Tier 1 item 6), which is the honest reading of a generic scope sentence attached to a specific claim.

Still open here: the 4-2-1 rule and the ~50%/1-year figure are genuine ETDRS-derived facts but remain unsourced in the entry — the regrade stopped the entry from overstating its evidence, it did not supply the missing citation.

A fourth use of the same reference at phenotypes[0] was reviewed and left as SUPPORT deliberately. It quotes a different sentence — "Diabetic retinopathy (DR) is a major complication of DM and a leading cause of vision loss in working middle-aged adults" — attached to a Visual Impairment phenotype, which it supports directly. Reuse of one reference across many nodes is not the anti-pattern; reuse of one generic sentence to carry specific claims is. Worth stating, because a reviewer reading "fourth use of PMID:29776671" flagged it as part of the same defect.

Also fixed (Tier 1 item 7): phenotypes[2] "Retinal Hemorrhage" described intraretinal dot-blot and flame-shaped hemorrhages in NPDR while citing a snippet about vitreous haemorrhage in PDR — a different compartment at a different stage, and one already modeled by the Vitreous Hemorrhage node immediately below. Regraded NO_EVIDENCE, since the sentence is silent on the claim rather than weakly supportive of it. The NPDR intraretinal pattern still needs a source.

The vitrectomy item was also moved PARTIAL → NO_EVIDENCE on the same reasoning: its explanation already said the snippet "never mentions vitrectomy", which is the definition of NO_EVIDENCE, not PARTIAL.

Also in this entry, both still open: - [Open] treatments[0] claims "18–45% of patients gaining ≥15 ETDRS letters" — not in any of its three snippets. - [Open] classifications.harrisons_chapter[0].evidence[1] uses the article title as its snippet, which is a degenerate citation.

A third bullet here — phenotypes[2] "Retinal Hemorrhage" citing a vitreous haemorrhage snippet — was fixed in Tier 1 item 7 and is described above; it is removed from this list rather than left to contradict it. That duplication is the hazard recommendation 8 below addresses.

Beyond the fixed ERG error, one sentence — "Inherited mutations in the rod visual pigment, rhodopsin, cause the degenerative blinding condition, retinitis pigmentosa (RP)" — is reused as SUPPORT for night blindness, rod-cone dystrophy classification, and bone-spicule pigmentation, none of which it mentions. The explanation fields do the actual work via curator inference ("RP is classically defined by…"), which is reasoning, not evidence.

Fixed (Tier 1 item 7): all four instances — pathophysiology[1].downstream[0], phenotypes[0], phenotypes[2], phenotypes[4] — are now PARTIAL, each with an explanation naming what the sentence does and does not say. The bone-spicule node went further: PMID:33669941 reports the finding directly in RHO patients ("Bone spicule pigmentation was seen in all individuals with RP… and none of the patients diagnosed with CSNB"), so it now carries a SUPPORT item alongside the downgraded one — and the same sentence corroborates the phenotype's RP4 scoping.

Also fixed: phenotypes[5] "Reduced rod electroretinogram" cited a snippet about BCVA and visual-field decline rates with no ERG content. The same Kobal cohort supplies the observation directly ("The function of the rod system, as revealed by dark-adapted (DA) full-field ERG (ffERG), was highly dysfunctional in all patients."), so that is now the SUPPORT item and the natural-history snippet is PARTIAL as corroboration of progressive functional loss.

Both are the third and fourth times in this review that a defect deferred as "needs a replacement source" was closed by a source already in the repository — in these two cases, a cached abstract fetched for an entirely different node.

Two items previously listed here are now fixed: the orphan 72-year fragment at pathophysiology[2].downstream[1] (Tier 1 item 3), and the unsourced VERY_FREQUENT band on the Riggs-type electroretinogram node together with the absence of any human RHO ERG observation — both closed by PMID:33669941 (Tier 1 item 5). The class-level CSNB definitional snippet on Congenital night blindness was also regraded PARTIAL, since the RHO-specific human step is now carried by its own citation rather than by inference from a definition.

Retinopathy_of_Prematurity

  • [Fixed] phenotypes[0] frequency FREQUENT rested on "141 550 infants received ROP screening in Germany" — a denominator, not a rate. Fixed: band dropped, with a notes: line recording why so it is not re-derived from the same snippet (cf. docs/frequency-evidence-guidelines.md, which says omit rather than justify).
  • [Open] treatments[1]'s description still asserts anti-VEGF is "preferred over laser for Zone I and posterior Zone II ROP due to better structural outcomes." Neither remaining evidence item carries that: one is the recurrence-risk trade-off, the other a registry trend in treatment preference. The superlative needs a source or should be softened. (The evidence grade half of this item — SUPPORT on an opposite-direction result — was fixed; see Tier 1 item 4.)
  • [Fixed] phenotypes[4] Myopia FREQUENT cited only "laser photocoagulation can lead to refractive errors" — supporting neither the frequency nor the "regardless of treatment" scope. Fixed (Tier 1 item 7): band dropped with a notes: rationale, description narrowed to what the evidence covers, item regraded PARTIAL. Leaving this while dropping the band on phenotypes[0] eighty lines earlier was the internal inconsistency worth catching.
  • [Open] treatments[2] "fewer than 30% achieving ambulatory vision" in Stage 5 — unsupported by its snippet.

Central_Retinal_Artery_Occlusion

  • [Fixed grade / Open source] pathophysiology[1] "Inner Retinal Ischemic Injury" (retinal edema, neuronal injury) was SUPPORTed by "CRAO has consistently been identified as a serious medical condition that leads to substantial visual impairment" — no mechanism content at all. Grade fixed to PARTIAL (Tier 1 item 6); the explanation now states that a mechanism source is still required. Reading the full cached abstract confirms there is no mechanistic content anywhere in it, so this one does still need a replacement citation — the regrade stops the overstatement, it does not close the gap.
  • [Fixed grade / Open band] phenotypes[1] "Reduced Visual Acuity" VERY_FREQUENT cites a meta-analysis aim statement. Grade fixed to PARTIAL (Tier 1 item 7) — a statement of intent establishes that visual outcomes are the measured endpoint, but evidences neither the phenotype nor the band. The VERY_FREQUENT band remains unsourced.
  • [Open] The THEIA phase 3 trial (PMID:41109232) is cited only for a background definition; its actual result is not recorded anywhere in the entry.

Glaucoma

  • [Fixed grade / Open source] pathophysiology[3] "Trabecular Meshwork Dysfunction" describes age-related change, oxidative stress, and ECM alteration, but both evidence items are about myocilin — a different (Mendelian MYOC) mechanism. Fixed: both are now PARTIAL, and both carry an evidence_source. The node still lacks evidence for the age/oxidative-stress mechanism it actually describes.
  • [Fixed] treatments[2] Alpha Agonists: the snippet "IOP reduction was similar for both groups" is a non-difference result vs timolol (n=16); the explanation read it as evidence of "clinically meaningful intraocular pressure lowering." Fixed: regraded PARTIAL, explanation rewritten as non-inferiority to an established agent rather than a demonstrated absolute effect.
  • [Fixed] genetic[1] OPTN was typed Risk Factor with notes scoping it to normal-tension glaucoma, but the cited snippet says only "associated with primary open angle glaucoma" and OPTN E50K is generally treated as causative-dominant. Fixed (Tier 1 item 7): retyped Causative, the association item regraded PARTIAL since "associated with" settles neither the NTG scope nor the typing, and the notes: now records that the NTG scoping comes from the wider OPTN literature rather than the cited paper. Both OPTN items are tagged IN_VITRO (cultured cells expressing mutant optineurin) and both MYOC items OTHER (review synthesis) — the evidence_source backfill this PR began one file section earlier, now finished.

Cytomegalovirus_Retinitis and Fuchs_Endothelial_Corneal_Dystrophy — minor only

  • [Open] CMV pathophysiology[2]: the characteristic morphology claim (yellow-white opacification, centrifugal advance) is not in its snippet, which covers only rapid progression to blindness.
  • [Open] Fuchs pathophysiology[5].evidence[2]: a textbook statement of what mitochondria do in general, cited as a "mechanistic bridge" to FECD apoptosis.
  • [Open] Fuchs phenotypes[7]: Nyctalopia FREQUENT comes from an Orphanet HPO annotation that primary FECD literature does not support. The curator flagged this transparently in the description — correct handling, but it likely reflects an upstream Orphanet annotation defect worth reporting to ORPHA:98974.

What is working well

Three entries should be treated as reference examples of evidence discipline:

  • Age_Related_Macular_Degeneration — the strongest of the ten. It records a disconfirming caveat as first-class evidence (drusen "might be no more than a biomarker… and not a cause", graded PARTIAL), keeps protective CFB/C2 haplotypes rather than flattening the locus to risk-only, and records the complement-inhibitor negative BCVA result and the pegcetacoplan MNV safety signal alongside the positive lesion-growth result.
  • Achromatopsia — grades all gene-therapy evidence PARTIAL and quotes the investigators' own limitation ("the absence of randomized concurrent control individuals precludes determining a cause-and-effect relationship"). It also holds the progressive-vs-stationary tension openly, citing both the OCT age-dependent thinning and the "predominantly stationary with respect to BCVA" cohort finding.
  • Fuchs_Endothelial_Corneal_Dystrophy — quotes ORPHA rows with their exact frequency bands rather than restating them, so every frequency is traceable.

The common factor: the supports: grade is used as a real signal. Where entries fall down, it is almost always because everything is graded SUPPORT.

Cross-cutting recommendations

  1. description and explanation are unvalidated prose, and that is where wrong claims hide. The two original Tier 1 errors lived there, behind correctly-transcribed snippets; any protocol that checks snippets without reading the surrounding prose will miss them. This cuts both ways, and did: PR review caught an uncited sentence introduced into the Retinoblastoma description by the fix itself (the true-but-unsourced uveal-melanoma contrast), which has since moved to notes:. Unevidenced context belongs in notes:, never in a description sitting above an evidence: block that does not cover it.
  2. Treat mid-clause snippet truncation as a lint target. A snippet ending in a dangling preposition or conjunction ("…malignancy in", "…highly active antiretroviral", "whereas this could not be computed for lower acuities.") is a cheap, mechanically detectable signal of scope-dropping. This would be a useful advisory check in just qc.
  3. One generic snippet reused across many specific nodes is an anti-pattern. The DR/ICO and RHO cases are both this shape. Reusing a definitional sentence as SUPPORT for a downstream specific claim should default to PARTIAL. The DR/ICO instances have been regraded (Tier 1 item 6) and one RHO instance with them; the four-way reuse of the RP-definition sentence in RHO-Related_Retinopathy is the remaining known case.
  4. Frequency bands still need their own evidence. Several FREQUENT / VERY_FREQUENT values here rest on snippets that establish only the association, exactly the failure docs/frequency-evidence-guidelines.md warns about.
  5. A grade correction is not a citation. Regrading SUPPORT → PARTIAL stops an entry overstating what it has, but the underlying claim stays unsourced — Diabetic_Retinopathy's 4-2-1 rule and the CRAO ischemic-injury mechanism are both still gaps after their regrades. Regrade to stop the misrepresentation, then track the missing source separately; do not let the honest grade close the item.
  6. PARTIAL is not the default downgrade — distinguish it from NO_EVIDENCE. The schema defines PARTIAL as "partially or indirectly supports" and NO_EVIDENCE as "does not contain evidence relevant to the claim". Downgrading everything to PARTIAL re-creates the original problem one notch lower, because it still reads as partial support. Working rule used here: if the snippet frames the claim but stops short of establishing it, PARTIAL (a meta-analysis aim statement about visual outcomes; a guideline scope sentence about the treatment area); if the snippet is simply silent on the claim, NO_EVIDENCE (a vitreous haemorrhage sentence under an intraretinal-hemorrhage node; a scope sentence that never mentions the procedure it is attached to). A snippet about a different entity is not weak evidence for this one.
  7. Reference reuse is fine; sentence reuse is the anti-pattern. One paper can legitimately support many nodes when each cites the sentence that speaks to that node — this entry set has several such cases. The defect is one generic definitional sentence carrying specific downstream claims. Audit at the snippet level, not the reference level, or correct citations get regraded along with the bad ones.
  8. A findings document needs an explicit status marker per item, not prose tense. This report drifted out of sync with the KB in four consecutive review rounds, and once contradicted itself inside a single section — a bullet describing a defect in the present tense sat nine lines below a paragraph announcing the same defect fixed. Marking every item Fixed or Open makes a stale entry visually obvious on the next pass, where "defect stated in past tense, then a Fixed note" does not. Any review artifact that ships alongside the data it describes will drift as the data is repaired; the format has to make the drift cheap to spot.

Reproducing

just validate kb/disorders/RHO-Related_Retinopathy.yaml
just validate-references kb/disorders/Retinoblastoma.yaml
just validate-terms-file kb/disorders/RHO-Related_Retinopathy.yaml