Skip to content

workhorse-models.md

Latest commit

 

History

History
442 lines (373 loc) · 27.5 KB

workhorse-models.md

File metadata and controls

442 lines (373 loc) · 27.5 KB
Raw

Workhorse-model verification dossier

Status: literature-sourced, human PDF sign-off pending. This document is the evidence package for promoting the dataset's most-used records from review_status: unverified toward verified. It was assembled by automated literature review against accessible sources (open-access full text, PubMed Central, regulatory reviews, and peer-reviewed reproductions of the original tables). Per CONTRIBUTING.md, an LLM may assemble this evidence but may not set review_status: verified on its own authority — a human must confirm each value against the source PDF and make the flip. Every number below carries its source and an access status so that confirmation is a lookup, not a re-derivation.

Why this matters. The dataset ships every parameter value as illustrative by design (see the CHANGELOG preamble). That is honest, but it also means no value here can be cited in a real analysis yet. This dossier is the first step in closing that gap: it separates what is structurally correct (the model form matches the cited paper) from what is numerically grounded (the value matches a published estimate), and it flags the places where the current illustrative numbers are actively misleading and should be corrected before any promotion.

The pending_human_review state

Records filled from an identified literature source — but not yet PDF-confirmed by a human — now carry review_status: pending_human_review (per-parameter and at the record level). This is the honest bridge between an illustrative placeholder (unverified) and a human-confirmed value (verified), and it lets the dataset carry real, sourced numbers without overclaiming verification. onkos review-queue lists every such record and the source behind each parameter, so a researcher can open the cited table and promote it; the health report summarizes the queue. The lifecycle is defined in CONTRIBUTING.md.

Applied — breast consistency round + the open-data floor

The confirmed breast growth rate (Krishnan 2021, kGROW liver/breast = 0.00917/week, IIV 135%) was propagated to the other breast records for internal consistency: resistance.breast_first_line.two_population (kg) and tgi_metrics.bruno_2020.breast_biexponential (kg) → pending_human_review. The breast context is now consistently grounded across its Claret, two-population, biexponential, and baseline records (growth + resistance from Krishnan; kill/kDRUG kept illustrative per §6).

Negative results — recorded so the search is not repeated. A dedicated pass for per-indication on-treatment exponential-per-week growth constants found that the remaining indications are not cleanly fillable from open sources:

  • Melanoma — Mistry 2018 (the open melanoma TGI paper) uses a piecewise-linear model in mm/day, not an exponential per-week constant, and its numeric estimates live only in a Supplementary Table S4 whose body is missing from the open supplementary file. Not transcribable; stays illustrative.
  • HCC — no open on-treatment clinical TGI growth constant exists; the only open anchor is a natural-history volume-doubling-time meta-analysis (TVDT 4.6 mo → ~0.012/week diameter constant), which would overstate on-treatment growth. Stays illustrative.
  • CRC — no open clinical TGI growth constant; the canonical Claret 2009 capecitabine fit remains paywalled. Stays illustrative.

This marks the practical floor of open-source grounding: what remains unverified is paywalled, non-transcribable (wrong model family or drug-effect-scaling), or illustrative-by-design (the survival-link C-index scheme, generic exposure-response shapes). Further promotion needs a human with institutional PDF access.

Applied — baseline round (per-indication baseline SLD)

The tumor_type_baselines.* baseline tumor burdens (RECIST sum of longest diameters) were grounded in open-access trial/TGI sources and moved to pending_human_review (see §7): breast 55 → 69 mm (Krishnan 2021), CRC 90 → 143 mm (Machida 2008), HCC 110 → 79 mm (Salem 2021 / IMbrave150), melanoma 60 → 72 mm (Mistry 2018 / BRIM-3). NSCLC-1L is kept at 80 mm (round, in-range) to hold the flagship demo stable; NSCLC-2L is kept illustrative (its sourced OAK figure ~67 mm conflicts with the 2L>1L modeling invariant — flagged for human review). The per-context OS-divergence numbers barely moved (baseline scales start size, not model disagreement), so the README divergence column is unchanged.

Applied — model round

  • resistance.breast_first_line.claret → pending_human_review. Growth kL (0.012 → 0.00917/week, IIV 38 → 135%) and resistance-decay lambda (0.05 → 0.126/week, IIV 95 → 63%) grounded in the open-access Krishnan 2021 HER2-negative breast / docetaxel TGI fit (PMID 33818899, same Claret model structure, per-week units). The kill term kD was kept illustrative because Krishnan's kDRUG scales a KPD drug amount, not this record's dimensionless drug-effect E. Context updated to docetaxel / HER2-negative breast (n=183).
  • preclinical_translation.simeoni_2004.xenograft → pending_human_review (values grounded last round from the open reproduction PMC5660732).
  • immuno_oncology.kuznetsov_1994.tumor_immune → pending_human_review for its 7 dynamics parameters (values match the canonical published set; the illustrative initial condition eff0 stays unverified). Record remains tier D / hypothesis.
  • resistance.claret_2009.tgi description updated with the dacomitinib-NSCLC finding (Fostvedt 2022, PMC9893889 — real Claret-form fit, but per-year units and a power-law exposure term, so not transcribable into this record); stays unverified.

How to read the confidence column

  • Structure: confirmed — the model's equations match the cited paper, verified against an accessible source.
  • Values: confirmed (open) — the published estimate was read from an open-access source (or an open reproduction of the original table). High enough to act on after a human cross-check.
  • Values: not located — the original estimate sits behind a paywall and is not quoted in any accessible secondary source. Do not invent it; promotion of these records is blocked on institutional access to the source PDF.
  • Values: illustrative — the current dataset number is a placeholder with no literature grounding (and in one case traces to a known-illustrative third-party default — see Claret below).

Summary

Record Structure Values Status / action
resistance.breast_first_line.claret confirmed confirmed (open) — Krishnan 2021 breast/docetaxel fit ✅ pending_human_review — kL + lambda grounded with IIV; kD kept illustrative (effect-scaling) (see §2)
preclinical_translation.simeoni_2004.xenograft confirmed confirmed (open reproduction) ✅ pending_human_review — real paclitaxel/A2780 set (see §3)
immuno_oncology.kuznetsov_1994.tumor_immune confirmed confirmed (open) — matches the published set ✅ pending_human_review (7 dynamics params); stays tier D / hypothesis (see §5)
tgi_metrics.wang_2009.biexponential biexponential = Stein-Fojo form; Wang's own model was exp-decay + linear not located ✅ Applied — description corrected; kg/ks provenance anchored to Stein 2008 (see §2)
resistance.claret_2009.tgi confirmed illustrative (CRC table paywalled; dacomitinib fit not transcribable) ✅ Provenance documented (Fostvedt 2022); stays unverified; C-index is illustrative-by-design (see §1)
tgi_metrics.*.biexp / tgi_metrics.bruno_2020.* confirmed (biexponential is a real Stein-Fojo form) illustrative Ground magnitudes against Stein 2008 ranges (see §2)
drug_effect.norton_simon.nsclc confirmed illustrative (human Gompertz constants are a distribution) Ground Vmax / growth constant against Norton 1988 (see §4)

§1 — Claret 2009 clinical TGI model

Record: resistance.claret_2009.tgi Source: Claret L, et al. J Clin Oncol 2009;27(25):4103-4108. DOI 10.1200/JCO.2008.21.0807, PMID 19636014.

Structure — confirmed (high). The model is a single ODE on tumor size: dTS/dt = KL·TS − KD(t)·Exposure·TS with KD(t) = KD0·exp(−λ·t) — first-order growth KL, exposure-driven kill KD0, and an exponentially decaying kill effect (rate λ) encoding emergent resistance. This matches the record's kernel. Confirmed via the MonolixSuite TGI library and BioModels MODEL1708310001 (both accessible), which cite Claret 2009.

Provenance of the paper's numbers — colorectal, not NSCLC. Claret 2009 was derived on capecitabine in metastatic colorectal cancer (CRC), with tumor size as RECIST sum of longest diameters and a per-week time base. The record is parameterized for an illustrative dacomitinib / NSCLC context — i.e. it borrows the structure but not the paper's data.

Values — not located + one actively misleading number.

  • The CRC Table 1 estimates (KL, KD0, λ, baseline, and their IIV CV%) are not quoted in any accessible source; the JCO original and ResearchGate full text are paywalled. Promotion of the CRC numbers is blocked on institutional access.

  • ⚠️ Do not trust the circulating "Claret values" KL=0.021, KD=0.025, λ=0.053 (BioModels MODEL1708310001). That entry is non-curated, is in per-day units (implausible for the per-week CRC model — 0.021/day ≈ 0.15/week growth is far too fast for CRC), carries no IIV although the paper reports it, and uses a unitless baseline of 200. These are third-party illustrative defaults. The record's own kL=0.021 / λ=0.061 appear to descend from this same illustrative source and must not be presented as Claret's estimates.

  • ⚠️ The record's predictive_performance C-index is the wrong metric type for this paper. It lists OS_C_index_external = 0.62. Claret 2009 reported no C-index; it reported a median-OS prediction: predicted median OS 431 days (90% PI 362–514) vs observed 401 days on the phase-III capecitabine arm, and a predicted survival improvement of 39 days (90% PI −21 to 110) vs observed 35 days (parametric lognormal survival model; baseline tumor size and week-7 tumor-size change were significant at P < .00001). Source: PMID 19636014 abstract (accessible).

    Important caveat — do not delete this field in isolation. An external C-index labeled (illustrative) is carried on 37 records as the dataset's uniform validation currency, and it is load-bearing: onkos.audit uses its presence as the external-validation gate for tier A/B, onkos.combine reads it for evidence-weighted model averaging, and the discriminability/survival-metric views compare the values across links. It is illustrative by the same design as every parameter value, not a one-off fabrication. The correct fix is therefore not to strip Claret's C-index (that would be inconsistent and would break tested audit/averaging behavior) but, when this record is promoted, to record the metric Claret actually reported (median-OS prediction) and decide dataset-wide whether the illustrative C-index scheme should be replaced with real, per-record validation metrics. That is a maintainer-level design decision, flagged here.

A real NSCLC grounding does exist (biexponential Claret-family). An open-access external-validation study in ALK+ NSCLC (alectinib vs crizotinib; PMC10363035) fits a biexponential TGI model "implemented as in Claret et al." (growth KG + shrinkage KS, no λ term), per-week:

Arm KG (1/wk) KS (1/wk) TS0 (mm)
Alectinib 0.00196 (RSE 13.2%, IIV CV 138%) 0.0342 (RSE 9.8%, IIV CV 97.1%) 57.0 (IIV CV 66.1%)
Crizotinib 0.00438 (RSE 12.1%, IIV CV 105%) 0.0373 (RSE 10.5%, IIV CV 91.3%) 57.0

Recommended actions (maintainer):

  1. Immediately correct or remove the fabricated OS_C_index_external = 0.62; if a validation field is wanted, record the real median-OS prediction (431 vs 401 d) and cite it, or leave the field empty until the paywalled value is confirmed.
  2. Decide the record's intent: either (a) re-anchor it to the actual CRC capecitabine context (and obtain the Table 1 values from the PDF), or (b) keep the NSCLC context but relabel it honestly as a biexponential Claret-family record grounded on PMC10363035 (note: that variant has no λ resistance term, so the resistance axis would need a separate, explicitly-illustrative record).
  3. The IIV CV magnitudes the record carries (kD 89%, λ 96%) are qualitatively well-supported — the NSCLC fits above show 90–140% CV on the kill/growth terms — so the "resistance terms carry ~90% CV" thesis is real even though these specific numbers are illustrative.

§2 — Stein-Fojo growth-rate constants & Wang 2009 NSCLC

Records: tgi_metrics.wang_2009.biexponential, the *.biexp family, and tgi_metrics.bruno_2020.breast_biexponential.

Stein 2008 — confirmed (open), with a context correction. Stein WD, et al. The Oncologist 2008;13(10):1046-1054 (open access). The growth-rate-constant model is f(t) = exp(−d·t) + exp(g·t) − 1 (f = marker normalized to day 0, t in days, g growth-rate constant, d regression-rate constant). ⚠️ Context correction: the original cohort is metastatic castration-resistant prostate cancer measured by PSA (112 patients, two NCI trials; validation 42 patients on ixabepilone), not renal cell carcinoma / sorafenib (those are companion papers). Published magnitudes:

  • g: median 10^−2.5 day⁻¹ ≈ 0.0032/day (≈ 0.022/week), ~1,500-fold spread.
  • d: median 10^−1.7 day⁻¹ ≈ 0.020/day (≈ 0.14/week), ~50-fold spread.
  • Prognostic strength (this is the headline the dataset leans on): survival correlated with log(g) at Pearson r = −0.72 (p < 0.0001) and only weakly with log(d) (r = −0.22); HR for above- vs below-median log(g) was 5.14 (95% CI 3.10–8.52). This strongly supports the dataset's "growth rate is the well-identified, prognostic term" tier rationale.

Wang 2009 — structural mismatch + values not located. Wang Y, et al. Clin Pharmacol Ther 2009;86(2):167-174 (abstract accessible; full text paywalled). ⚠️ The record models a biexponential form, but Wang 2009 used "a mixed exponential decay and linear growth model": TS(t) = BSL·exp(−SR·t) + PR·t, where SR is an exponential shrinkage-rate constant and PR is a linear progression rate (mm/time), not an exponential rate constant. The OS model is parametric with three predictors: ECOG, baseline tumor size, and week-8 tumor-size change (advanced NSCLC, four FDA registration trials). The numeric SR/PR estimates and the week-8 hazard ratio are not quoted in any accessible source — blocked on the paywall.

Recommended actions (maintainer):

  1. ✅ Applied (partial). The record's description now states explicitly that the biexponential form is the Stein-Fojo growth-rate-constant model and that Wang 2009's own NSCLC model was a distinct exp-decay-plus-linear parameterization; the record id and citation are unchanged to avoid breaking the many references to it. A fuller fix — renaming the record id to drop the Wang attribution, or adding a real exp-decay-plus-linear kernel that matches Wang 2009 — is left to the maintainer because the id is referenced across the README, tests, and notebooks.
  2. ✅ Applied to tgi_metrics.wang_2009.biexponential: its kg (0.018/week) and ks (0.055/week) source_locators now cite the open-access Stein 2008 ranges (g median ≈ 0.022/week, ~1500-fold spread; d median ≈ 0.14/week, ~50-fold spread) and confirm both values sit within them. This is a deliberately weaker grounding than Simeoni's: it is a cross-tumor magnitude anchor (Stein's cohort was prostate/PSA), not a same-context fit, and the values were left unchanged because they are already in-range and the record is referenced across 9 test files. The other in-context *.biexp records still carry illustrative magnitudes.
  3. bruno_2020.breast_biexponential cites a review (Bruno 2020 CCR) for specific values; reviews summarize rather than estimate. Note the research confirmed Bruno 2020 reports no specific breast KG/KS and no C-index, so its illustrative numbers are not supported by that citation. A primary breast TGI fit is available (Krishnan 2021, below) — re-anchor to it or keep explicitly illustrative.

§2b — Breast Claret-form TGI (Krishnan 2021) — ✅ grounded

Record: resistance.breast_first_line.claret Source: Krishnan SM, Laarif SS, Bender BC, Quartino AL, Friberg LE. CPT Pharmacometrics Syst Pharmacol 2021;10(5):511-521. DOI 10.1002/psp4.12629, PMID 33818899. Open access. HER2-negative metastatic breast cancer, docetaxel, n=183, organ-specific TGI with the same Claret structure (exponential growth + drug-driven kill with exponential resistance decay). Table 2:

Param Record (was → now) Krishnan 2021 IIV CV% Note
kL growth 0.012 → 0.00917/wk kGROW, liver/breast lesions 135% fastest organ class; per-week
lambda resistance decay 0.05 → 0.126/wk site-independent λ 63% per-week
kD kill 0.28 (unchanged) kDRUG 0.00123/wk 43% not transcribed — Krishnan's kDRUG scales a KPD drug amount, not this record's dimensionless effect E

kL and lambda are now pending_human_review; kD stays unverified with that reason recorded. This is the cleanest same-structure clinical grounding found — an open-access Claret-form breast fit with full IIV.

§3 — Simeoni 2004 preclinical xenograft model

Record: preclinical_translation.simeoni_2004.xenograft Source: Simeoni M, et al. Cancer Research 2004;64(3):1094-1101. DOI 10.1158/0008-5472.CAN-03-2524, PMID 14871843.

Structure — confirmed (high). Exponential-then-linear unperturbed growth dx1/dt = λ0·x1 / [1 + (λ0/λ1·w)^ψ]^(1/ψ); drug damages proliferating cells at rate k2·c(t)·x1; damaged cells traverse a transit chain x2→x3→x4 at rate k1 before dying; observed weight w = x1+x2+x3+x4; ψ fixed at 20. Confirmed via the PubMed abstract, BMC Cancer PMC7076937, and the MonolixSuite TGI library.

Values — confirmed (open reproduction). The original paper is paywalled (AACR 403), but its parameter table is reproduced in Strömberg & Hooker, PMC5660732 (open access), Table I, attributed to Simeoni 2004. Units: λ0 (1/day), λ1 (g/day), w0 (g), k1 (1/day), k2 (ng⁻¹·mL·day⁻¹), ψ fixed = 20.

Compound (cell line) λ0 (1/day) λ1 (g/day) w0 (g) k1 (1/day) k2 (ng⁻¹·mL·day⁻¹)
Paclitaxel (A2780) exp.3 0.311 0.656 0.033 0.968 6.29e-4
Paclitaxel (A2780) exp.4 0.273 0.814 0.055 0.968 6.29e-4
5-FU 0.215 0.412 0.065 0.056 2.021e-3
Compound A 0.349 0.363 0.010 0.405 3.45e-4
Compound B exp.7 0.309 0.796 0.034 0.517 2.89e-4
Compound B exp.8 0.369 0.511 0.016 0.615 2.93e-4

Untreated A2780 growth is characterized by λ0 ≈ 0.27–0.31/day, λ1 ≈ 0.66–0.81 g/day. The threshold ("tumor-static") concentration concept Ct ≈ λ0/k2 is part of the framework (a worked value of ~1100 ng/mL is quoted in the literature); treat the closed form as standard-but-unverified-at-source.

✅ Applied. The record now carries the paclitaxel/A2780 exp.3 values (λ0=0.311/day, λ1=0.656 g/day, k1=0.968/day, k2=6.29e-4 mL·ng⁻¹·day⁻¹, ψ=20 fixed), with each parameter's source_locator citing PMC5660732 Table I and this dossier. The implied tumor-static concentration is now Ct = λ0/k2 ≈ 494 ng/mL; the qualitative tests in tests/test_preclinical.py were re-dosed to physical concentrations around/above that threshold (the old illustrative doses were matched to a potency ~19× too high) and the dataset-health report was regenerated. review_status is now pending_human_review, not verified: the source is a high-quality secondary reproduction, so the verified flip still requires a human reading Simeoni 2004 Table I directly — the values are high-confidence, the sign-off is not an LLM's to give.

§4 — Norton-Simon kill on Gompertz growth

Record: drug_effect.norton_simon.nsclc Sources: Norton L. The Oncologist 2005;10(6):370-381 (open); Norton L. "A Gompertzian Model of Human Breast Cancer Growth." Cancer Res 1988;48:7067-7071.

Structure — confirmed (high). Kill proportional to the unperturbed Gompertzian growth rate: dV/dt = g·V·ln(Vmax/V) − k·E·V·ln(Vmax/V). Confirmed via Mistry et al., PMC3228251 (open), which writes the treated form dN/dt = {αN − βN·ln N}·[1 − D(t)] — exactly the kill-proportional-to-growth-rate structure.

Values — illustrative; the human growth constant is a distribution, not a point. Norton 1988 (read from an open PDF mirror) reports, for human breast cancer: carrying capacity N(∞) = 3.1 × 10¹² cells, lethal burden 10¹² cells, and a growth-rate constant b that is lognormal with mean log_e(b) = −2.9, SD 0.71, with t in months (median b ≈ 0.055/month). The record's Vmax = 200 mm and g = 0.08/week are illustrative NSCLC-scale numbers, not these. For experimental solid tumors, Vaghi et al., PLOS Comput Biol 2020, PMC7059968 (open) gives Gompertz dV/dt = αV·ln(K/V) with breast α = 0.58/day, K = 2,600 mm³ (mouse) — note these are murine, not human.

Recommended action (maintainer): the Norton-Simon record is best kept explicitly illustrative — there is no single published human NSCLC Gompertz constant to anchor it, and Norton's own message is kinetic heterogeneity (a distribution). Document the growth-constant IIV honestly rather than implying a point estimate. The record's purpose (contrast the kill mechanism against log-kill in the divergence view) does not require a verified value.

§5 — Kuznetsov 1994 tumor-immune model

Record: immuno_oncology.kuznetsov_1994.tumor_immune Source: Kuznetsov VA, Makalkin IA, Taylor MA, Perelson AS. Bull Math Biol 1994;56(2):295-321, PMID 8186756.

Structure & values — confirmed (open). This is the one workhorse record whose numbers are already the genuine published values. The original is paywalled, but the canonical nondimensional parameter set is reproduced verbatim in d'Onofrio, arXiv 1309.3337 (peer-reviewed in Math Comput Modelling 2008) and independently corroborated by Marchant et al., arXiv 1909.05203 (J Theor Biol 2020) and EBI BioModels BIOMD0000000762. The dataset's values match:

Symbol Record Published Match
σ (s) source 0.118 0.1181 ✓
ρ (rho) recruitment 1.131 1.131 ✓
η (eta) half-saturation 20.19 20.19 ✓
μ (mu) inactivation 0.00311 0.00311 ✓
δ (delta) effector death 0.374 0.3743 ✓
α (alpha) tumor growth 1.636 1.636 ✓
β (beta) inverse capacity 0.002 0.002 ✓

Scaling (d'Onofrio): t_true = 9.9·t_nondimensional days; one nondimensional cell unit = 10⁶ cells.

Action taken: the seven dynamics parameters are now pending_human_review, with source_locator citing the confirming open reproductions — no value changed (they already matched). The illustrative initial condition eff0 stays unverified. The record correctly stays tier D / hypothesis / non-predictive: the parameters reproduce immune control, dormancy, and escape qualitatively (they were fit to BCL₁ B-lymphoma in mouse spleen, and the immunotherapy effect E is a hypothesis-tier augmentation). Sourced parameter values do not upgrade a model that is non-predictive by construction — exactly the distinction the tier system (separate from review_status) exists to make.

§6 — kD effect-scaling caveat (breast Claret)

Recorded here because it recurs: the dataset's clinical TGI records drive kill from a dimensionless drug-effect E (≈1 at a reference dose), whereas the source fits drive kill from a concentration or KPD drug amount. Growth (kL, kg) and resistance-decay (lambda) constants are independent of that scaling and transfer directly; kill potencies (kD, ks) do not. That is why resistance.breast_first_line.claret grounds kL and lambda from Krishnan 2021 but keeps kD illustrative — the same caveat applies to any future kill-term grounding.

§7 — Per-indication baseline tumor burden

Records: tumor_type_baselines.*. These context baselines supply the starting tumor size (RECIST sum of longest diameters, mm) and a reference growth rate for the divergence view. Baseline SLD is genuinely trial- and enrollment-dependent, so each is anchored to a representative open-access median rather than a universal constant.

Context was → now (mm) Source Access review_status
breast 1L 55 → 69 Krishnan 2021 (PMID 33818899), AVADO docetaxel arm median summed SLD open pending_human_review
CRC 1L 90 → 143 Machida 2008 (PMID 18845522), observed median SLD, mCRC open pending_human_review
HCC 1L 110 → 79 Salem 2021 / IMbrave150 (PMID 34189869), sorafenib-arm median open pending_human_review
melanoma 1L 60 → 72 Mistry 2018 / BRIM-3 (PMID 29222604), vemurafenib cohort median open pending_human_review
NSCLC 1L 80 (kept) round, in published range (ALK+ 57; IMpower150 ITT 70) open unverified (not a specific median; protects flagship demo)
NSCLC 2L 100 (kept) OAK 2L reported ~67 (Shang 2021), but that breaks the 2L>1L invariant open unverified (conflict flagged for human review)

Cross-indication pattern: published median baseline SLD clusters at ~57–80 mm for most solid tumors, with CRC the conspicuous outlier (143 mm) — bulkier abdominal/ hepatic metastatic burden. This real pattern replaced the previous illustrative ordering (which had HCC highest); the test_baseline_drives_context_specific_y0 invariant was updated from HCC>NSCLC to CRC>NSCLC to match it.

Maintainer checklist to set review_status: verified

For each record, per CONTRIBUTING.md:

  1. Open the source PDF (this dossier gives DOI/PMID and the specific table).
  2. Confirm structure, every value + units + IIV, the derivation context, and the transportability boundary.
  3. Where this dossier says "not located," obtain the value from the PDF — do not import third-party defaults (see the Claret BioModels warning).
  4. Flip review_status and update source_locator to the confirmed table/page.
  5. Re-run onkos validate, onkos report (and diff against docs/dataset-health.md), pytest -q, and the affected notebooks.