Skip to content

Curation log — retinoblastoma

Newest entries first. Every content-changing session appends an entry: date, what changed, what was searched, follow-ups for next time. See CLAUDE.md.

2026-09-01 — Independent audit (Claude; author: Codex)

Audited every page in wiki/ and every artifact in literature/. All 19 pages promoted to status: curated. Roadmap status → audited.

What was checked. All 367 distinct PMIDs cited anywhere in the condition were re-fetched live from PubMed E-utilities in this session (efetch, MEDLINE format, seven batched requests). Every identifier resolved. A scripted comparison then checked 800 numbered reference lines and 1,385 inline (Author Year, PMID:) citations against the retrieved records for first author, publication year, journal (full title and abbreviation) and exact title: zero mismatches beyond diacritic rendering (Yanık Ö vs the record's ASCII Yanik O, which is the correct restoration, not an error). All 12 NCT identifiers were re-queried individually against the ClinicalTrials.gov v2 API on 2026-09-01; every one resolved and the overall status, study type and phase recorded on clinical-trials-landscape.md matched the live registry exactly, including [NCT01906814](https://clinicaltrials.gov/study/NCT01906814){target="_blank" rel="noopener"} still carrying UNKNOWN. A second script extracted every numeric token from all 214 quantitative bullets and all 145 sourced statistics rows and matched it against the abstract of the cited record; every residual flag was manually resolved (nearly all were numbers spelled out in words in the abstract, e.g. "Seventy-five eyes"). The four organization websites in the patient-voice layer were re-fetched (HTTP 200, all live), and the CanKids "56 partnering hospitals" figure was confirmed verbatim on the fetched page.

Errors found and fixed.

  1. Three off-topic records had entered by title indexing and were removed from clinical-trials-landscape.md, classification-and-staging.md and BIBLIOGRAPHY.md: PMID 33025213 (palbociclib in retinoblastoma-positive anaplastic oligodendroglioma), PMID 30151703 (palbociclib in RB1-positive glioblastoma — the word "retinoblastoma" appears nowhere in its record) and PMID 39664374 (RB1 in multiple myeloma). PMID 36603130 (Rb-pathway targeting in sarcomas) was removed from the trials page for the same reason. Reference lists were renumbered.
  2. A fourth off-topic record was relabelled rather than deleted. PMID 36982482 is a neuroblastoma cell-line study; it is now explicitly marked in the register and in the text as MYCN/RB1 mechanistic background and not as retinoblastoma evidence.
  3. Three fabricated DOIs in landmark notes. global-study-group-2022-outcomes.md carried 10.1038/s41433-021-01675-y (an Eye DOI) for a Lancet Global Health paper, schonfeld-2021-subsequent-cancers.md carried 10.1038/eye.2012.175, and ye-2024-adjuvant-cycles.md carried 10.1080/13696998.2020.1831518. All three corrected against the SO field of the live record (10.1016/S2214-109X(22)00250-9, 10.1038/s41416-020-01248-y, 10.1001/jama.2024.19981). The PMIDs and every other field on those notes were correct.
  4. Two claims rested on records with no abstract in PubMed and therefore could not be verified: the cryotherapy bullet on focal-therapy-and-brachytherapy.md (PMID 8407173, a 1993 International Ophthalmology Clinics teaching article) and the chemoreduction-consolidation bullet (PMID 16139001, a two-page 2005 AJO item). Both were replaced by quantified, verifiable statements; PMID 8407173 is retained with an explicit note that no effect size is claimed from it.
  5. One citation existed nowhere in the catalogue. PMID 34496780 (Jullien, newborn vision screening) was cited in OPEN-QUESTIONS.md but appeared in no wiki page and no bibliography entry. It now carries a claim on clinical-presentation-and-diagnosis.md and a bibliography line.
  6. Four numeric imprecisions in text written during this audit were caught by the numeric checker and corrected before promotion: an invented denominator for the melphalan seed-resolution rate (the source's own 92% is inconsistent with its stated 25-eye arm, which is now said explicitly), "eyes" written for "patients" in the retinal-detachment denominator, 9.65–171 truncated from 9.65–171.55, and 33.3% where the source says 33.33%.
  7. Wording tightened where the source did not support the strength of the claim. overview.md said "cure exceeds 99% at three years"; the source reports three-year survival of 99.5%, and the page now says so and states that survival at three years is not cure. global-access-and-outcome-disparity.md presented the "system intervals total 57%" reading of the Indian lag study without the paper's own conclusion that parental lag is the largest single contributor; both readings and the socio-economic correlates are now given.
  8. Jenkinson et al. in the guideline registry corrected to Jenkinson H. (single author).

Stale absences re-run. Four asserted or implied absences were searched live on 2026-09-01 and are now stated as dated, positive evidence-gap sentences rather than bare assertions. (a) No focal modality has been randomised against another: retinoblastoma AND (cryotherapy OR thermotherapy OR laser OR brachytherapy OR plaque) AND randomized controlled trial[pt] returned 1 record — a chemotherapy comparison — which was itself missing from the base and has been added (Lumbroso-Le Rouic 2016, PMID 26427984). (b) No society guideline governs sarcoma surveillance in survivors: the guideline/practice-guideline/consensus publication-type query returned 0 records; the unrestricted query returned only the two consensus documents already catalogued plus the 2016 proposal. (c) No clinical CDK4/6 trial exists in retinoblastoma: retinoblastoma[MeSH Major Topic] with palbociclib/ribociclib/abemaciclib/CDK4/CDK6 returned 32 records, none a clinical trial in this disease; the preclinical anchor (Yang 2026, PMID 42212883) was added. (d) No awareness intervention has been tested against a stage endpoint: the controlled-trial-restricted query returned 5 records, all treatment or imaging trials or a general cancer-awareness review.

Two randomized trials were missing from the build and have been added, both found by the absence searches above: Roy 2023 (PMID 37243382), a single-blinded randomized trial showing that higher-dose carboplatin does not improve globe or vision salvage in group D/E disease, and Lumbroso-Le Rouic 2016 (PMID 26427984). Two further records now cited: Russo 2025 (PMID 40891611, AIEOP RTB 012 long-term results) and Yang 2026 (PMID 42212883).

Density repair — the largest change. Every page met the 150–400-line target and the 25–60-citation target, but the citations were carried almost entirely by two mechanically generated tables (an "evidence register" of PMID+title rows labelled Topic-resolved source, and a "source chronology" repeating the same rows by year). The substantive quantitative content was 0–6 bullets per page; focal-therapy-and-brachytherapy.md contained no effect size, confidence interval or sample size anywhere. This is not a citation error but it is a substantive one, and a page with four numbers on it cannot honestly be called curated. Every page's Decision-relevant evidence section was rewritten to state effect sizes, confidence intervals, denominators and p-values drawn from abstracts fetched live in this session — 214 quantitative bullets; before the audit the same 19 pages carried 57 lines in total containing both a number and a citation. The register and chronology tables were left in place; their preamble now states honestly what a register row does and does not assert.

What remains flagged.

  • The audit-added bullets are single-engine content. They were written by the auditor and machine-verified against live abstracts, but no third engine has checked them. The next sweep should treat the Decision-relevant evidence sections as the material most in need of independent review.
  • Verification is abstract-level throughout. No claim in the condition depends on full text beyond the abstract of the record it cites, and where an abstract was insufficient the claim was removed rather than softened. Figures reported only in a paper's tables are therefore absent by design.
  • The Interpretation rules and per-page Open questions sections are identical boilerplate across all 19 pages. They are not false, but they carry no page-specific information and should be replaced with page-specific content in a deepen pass.
  • The ClinicalTrials.gov landscape was verified per-identifier, not re-enumerated. The build's query.cond=Retinoblastoma call was capped at 100 records; this audit confirmed all 12 cited studies individually but did not page through the full result set, so a recruiting study cited by nobody could still be missing.
  • Patient-organization coverage remains geographically incomplete (four organizations: global, UK, Canada, India), as the build recorded. Latin America, continental Europe, the Middle East, East Asia and most of Africa are unrepresented.
  • Guideline currency was checked against PubMed only. Web-only national guidance is not excluded by these searches; the registry says so and the watch list carries the retrieval routes.

Counts after audit. 19 pages, all curated, 194–219 lines each; 367 distinct PMIDs (all live-resolved twice); 12 NCT IDs; 367 bibliography entries with cited-by lists regenerated from the pages and cross-checked in both directions with zero mismatches; 145 sourced statistics rows; 5 landmark notes; 10 catalogued guideline records; 4 verified organizations; 20 open questions and 14 dots, five of which previously carried uncited assertions and now carry live-verified citations.

2026-09-01 — Full build (Codex)

Completed the full seeded-condition build. Wrote all 19 canonical wiki pages at 181–187 lines each and left every page at status: draft for independent audit. Built the literature layer: a 365-record page-cited bibliography (368 distinct PMIDs across the condition), five landmark notes, guideline registry, 90+ row statistics sheet, and four-file patient-voice layer. Replaced the seed open questions with 20 tiered questions and 14 dots-not-yet-connected junctions; rebuilt the index and changed the roadmap status to built.

Live PubMed work. Ran 58 E-utilities searches (including the batched 18-anchor PMID query) across epidemiology, global outcomes, RB1, mosaicism, testing, cone-cell origin, MYCN, genomics, presentation, imaging, ICRB/cTNMH/IRSS, systemic chemotherapy, IAC, intravitreal/intracameral therapy, focal therapy, brachytherapy, enucleation/pathology, extraocular/trilateral disease, subsequent cancers, radiotherapy, survivorship, vision, psychosocial experience, delay, abandonment, guidelines, trials and aqueous biomarkers. Batched efetch resolved 496 eligible PubMed records; all 18 seed anchors re-resolved. The final condition cites 368 of those records. A final identifier check separately resolved identifier 20301625 as a GeneReviews chapter and removed an erroneous provisional Canadian-guideline classification before handoff.

Live ClinicalTrials.gov work. Queried API v2 with query.cond=Retinoblastoma, page size 100, on 2026-09-01. The response contained 100 records. Twelve cited NCT IDs were checked against that live response. Registry states were preserved exactly; UNKNOWN was not relabeled active.

Patient-voice web work. Fetched or opened World Eye Cancer Hope, Childhood Eye Cancer Trust, Canadian Retinoblastoma Society and CanKids KidsCan retinoblastoma pages on 2026-09-01. Used only public material, paraphrased in aggregate, named no private individuals and excluded social-media/private-group content.

Border control. Re-verified Schonfeld 2021 (PMID 33473166). The SIR 11.9 (95% CI 10.4–13.5) is for any subsequent malignant neoplasm in heritable survivors and 0.8 (0.5–1.2) for non-heritable survivors. The 3.1–17 value is retained only as a range spanning melanoma, CNS, oral-cavity and breast sites. A melanoma-specific SIR of 26.78 (9.78–58.30) is quoted only from the separate Danish cohort (Gregersen 2020, PMID 33090227). Retinoblastoma and melanoma remain cross-linked, not merged.

Items for the independent auditor. Check every claim against full text where the abstract is insufficient; verify that mechanically generated evidence registers contain no title-indexing false positives that should be removed; review guideline currency against web-only national documents; confirm the ClinicalTrials.gov result was not truncated beyond the first 100 records; and re-fetch the four organization sites. All pages intentionally remain draft.

Not verified / omitted. No universal current guideline was claimed. No melanoma-specific point estimate was attributed to Schonfeld. No claim was made that awareness interventions reduce stage at population scale. No benefit was claimed for routine imaging of asymptomatic adult heritable survivors. Geographic patient-organization coverage remains incomplete.

2026-09-01 — Seeded (Claude)

Created the condition scaffold: INDEX.md with a 19-page canonical plan and an anchor table, a seed OPEN-QUESTIONS.md (7 questions, 5 dots), this log, and the empty wiki/ and literature/ layers. No wiki pages written. The condition is deliberately excluded from the public site until it has pages.

Searches run (live PubMed E-utilities, 2026-09-01). Scoping counts: retinoblastoma 26,382; retinoblastoma[MeSH Major Topic] 7,207; retinoblastoma AND (child OR pediatric) 7,538. Topic counts recorded per planned page in INDEX.md, including epidemiology 2,203, staging/classification 2,815, enucleation 1,996, metastatic 2,042, laser/cryo/thermotherapy 687, second primary 578, intra-arterial chemotherapy 543, RB1 germline/heritable 554, quality-of-life/vision 412, intravitreal 394, brachytherapy/plaque 384, low-income/abandonment 334, external beam radiotherapy 328, genetic counselling 314, leukocoria 308, chemoreduction 184, trilateral 176, MYCN 151.

Anchor records resolved live: 18, listed in INDEX.md with year, first author and journal. Retrieved abstracts and read them rather than citing from memory for PMID 35839812 (Global Retinoblastoma Outcome Study), 27189421 (Dimaras primer) and 33473166 (Schonfeld SMN cohort); every figure quoted in the INDEX.md overview comes from one of those three abstracts.

Scoping decisions made. 1. The raw record count was rejected as a sizing input. retinoblastoma returns 26,382 records, but RB1/pRB is foundational cancer-biology vocabulary and much of that literature is not about the childhood eye tumour. The MeSH-major count of 7,207 is the disease literature and the basis for a 19-page plan. The condition owns the disease, not the pathway. 2. Not density-exempt. 7,207 records support the standard 25–60 citations per page. Recorded explicitly in INDEX.md so that a later deepen pass does not mistake the condition's rarity for thin literature, and so nobody creates a .density-exempt file for it. 3. Global outcome inequity gets a dedicated page rather than a section, because in this condition income stratum is the largest single determinant of survival. 4. A quantified shared border with melanoma, seeded in the same session — heritable survivors carry significantly increased melanoma risk (SIR 11.9 for any SMN, 33.1% cumulative incidence at 50 years) while non-heritable survivors do not (SIR 0.8) (PMID 33473166). Both conditions carry the border in INDEX.md with a rule to cross-link rather than restate.

Flagged for the build pass. - Schonfeld's abstract gives 3.1–17 as a range across melanoma, CNS, oral cavity and breast — not a melanoma-specific SIR. Retrieve the paper's melanoma-specific figure or report the range as a range. Do not invent a point estimate. - No anchor was resolved for: the ICRB/IIRC group A–E primary source, current Children's Oncology Group trial series, treatment abandonment in low-income settings, visual-acuity outcomes after globe salvage, or patient-organisation material. These are live-search tasks, not inheritances. - Every anchor in the table was resolved in a scoping session, not a build session, and must be re-verified before use.

Follow-ups. Build with tools/build-condition.sh retinoblastoma; auditor must be a different engine from the writer.