Cancer

Mammography

Mortality benefit DOCUMENTED: pooled trial relative risks of 0.88 / 0.86 / 0.67 / 0.80 across the age bands from the 40s to the 70s — real, age-graded, and no all-cause mortality reduction shown in any pooled analysis. Harms in the same breath: the cumulative chance of at least one false-positive recall over a decade is 61.3% with annual vs 41.6% with biennial screening, and overdiagnosis runs 11–22% of screen-detected cancers in the trials — the full published range is 0–54%, printed with its reason: the estimate is a residual, and its method moves it.

Update — 3 November 2026 (Tranche 15). Mammography — Two-County checked. The Swedish Two-County trial's primary long-term paper (29 years; 133,065 women) reports breast-cancer mortality RR 0.69–0.73 depending on the endpoint committee — consistent with the pooled record graded here. Mammography mortality stays DOCUMENTED.

PSA (prostate)

DOCUMENTED, small and contested. Europe's ERSPC trial: mortality RR 0.79 — about 1.3 deaths prevented per 1,000 men invited over ~13 years, bought with 27 diagnoses per death prevented. America's PLCO was null (RR 1.04), but ~46% of its control group received PSA tests each year anyway, so it compared organised with opportunistic screening. Overdiagnosis: 20.7–50.4% of screen-detected cancers.

Update — 3 November 2026 (Tranche 15). PSA at 23 years. The ERSPC trial's final long-term analysis (NEJM 2025; 162,236 men aged 55–69; median 23 years) finds prostate-cancer mortality RR 0.87, with the absolute benefit grown to 2.2 deaths prevented per 1,000 men invited and 12 diagnoses per death prevented (from 27 at 13 years). PSA mortality stays DOCUMENTED — real, small, and contested (PLCO null; no all-cause benefit).

Colorectal — split by machine

Sigmoidoscopy: PROVEN on both endpoints (four trials, 458,002 people: incidence IRR 0.78, mortality IRR 0.74) — it removes the precursor at the test. Colonoscopy by invitation: incidence DOCUMENTED, mortality UNPROVEN at 10 years (NordICC, 84,585 randomised, 42% uptake: incidence RR 0.82; mortality RR 0.90, interval crossing no-effect). Stool tests: gFOBT mortality DOCUMENTED; FIT mortality DOCUMENTED on observational data only.

Update — 3 November 2026 (Tranche 15). Colonoscopy, 13-year result. The NordICC trial's 13-year readout (Lancet 2026; 84,583 people aged 55–64; one invitation; 42% uptake) confirms the incidence effect and does not settle mortality: colorectal cancer risk 1.46% vs 1.80% (RR 0.81; one case prevented per 294 invited), colorectal-cancer death 0.41% vs 0.47% (RR 0.88, 95% CI 0.68–1.08 — not significant). Invitation-to-colonoscopy mortality remains UNPROVEN; incidence remains DOCUMENTED. The adherence-adjusted (per-protocol) estimates from the primary paper are incidence RR 0.55 and mortality RR 0.70 (CI 0.26–1.25) — the "screened people halve their mortality" figure in circulation is not what the trial's own adjusted analysis found.

Update — 3 November 2026 (Tranche 15). FIT has a randomised outcome readout. Spain's COLONPREV trial (Lancet 2025; 57,404 people aged 50–69) found invitation to biennial FIT non-inferior to invitation to colonoscopy for colorectal-cancer mortality at 10 years (0.24% vs 0.22%), with higher participation for FIT (39.9% vs 31.8%). The trial had no unscreened arm, so FIT's mortality grade stays DOCUMENTED — now on randomised programme-comparison evidence as well as observational data, not on either alone.

Update — 3 November 2026 (Tranche 15). Sigmoidoscopy at 23 years. Norway's NORCCAP trial now reports 23-year results (98,654 people analysed): incidence reduced in men (−1.7 percentage points) and women (−0.5), colorectal-cancer death reduced in men (−0.8 points) but not measurably in women (−0.1 points, CI crossing zero); adding one FIT to sigmoidoscopy changed nothing. Sigmoidoscopy's grades stay PROVEN (four-trial pooled record), with the sex difference now part of the headline, not a footnote.

Lung CT (LDCT)

Mortality PROVEN — in trialled high-risk smokers only: NLST IRR 0.85 (323 screened per death prevented), NELSON IRR 0.75 (130 per death prevented). It prevents nothing: cumulative incidence was higher in the screened arms, so calling it prevention is MISIDENTIFIED. False positives ran 16–27% per round in NLST.

Cervical

The prevention-grade case: incidence PROVEN (HPV-based vs cytology screening, invasive cancer RR 0.60; 0.45 after 2.5 years of follow-up) because a positive leads to removing precursors. Mortality DOCUMENTED — no placebo-controlled mortality trial exists or could now be run; mortality fell with incidence in every screened population.

Multi-cancer blood tests (MCED)

UNPROVEN on both endpoints that matter: no mortality benefit shown by any test, and the one large trial (NHS-Galleri, 142,250 people) missed its primary endpoint — late-stage incidence RR 1.03. In PATHFINDER, 62% of positive signals had no cancer behind them. "Detects cancer early and saves lives," as an outcome claim, is MISIDENTIFIED: detection is not deaths prevented.

Update — 3 November 2026 (Tranche 15). MCED — still detection, not outcomes. PATHFINDER 2 (Nature Medicine 2026; 35,878 people, single-arm) reports Galleri performance at 12 months — PPV 60.3%, sensitivity 39.3% across all cancers — with no comparison group and no mortality or late-stage endpoint. Its authors state randomised trials are needed to assess clinical utility. MCED mortality and late-stage benefit remain UNPROVEN; no NHS-Galleri outcome follow-up has been reported.

Whole-body MRI

No outcome data of any kind — no trial, no mortality record. Across 5,373 asymptomatic people scanned, 32.1% had critical or indeterminate findings, 16.0% were false positives, and only 12.6% of findings were ever verified. A "clean scan rules out cancer" claim is FABRICATED in effect against this record.

"Sugar feeds cancer — starve the tumour"

The Warburg effect is PROVEN biology, and a sugary-drink association is DOCUMENTED (HR 1.18 per 100 mL/day). But "sugar causes cancer" directly is UNPROVEN, and "cut sugar to starve the tumour" is MISIDENTIFIED: the liver holds blood glucose steady on any diet, so the lever cannot reach the variable. Ketogenic diet as therapy: UNPROVEN.

"Mobile phones cause brain cancer"

UNPROVEN. The COSMOS cohort: HR 1.00; the WHO-commissioned 2024 review found moderate certainty of no association. The "WHO classed phones 2B" gloss is MISIDENTIFIED — 2B grades how certain the evidence is, not danger at ordinary exposure.

"Antiperspirants cause breast cancer"

UNPROVEN — the direct study found nothing (OR 0.9), and follow-up literature stayed null. The sweat-blockage mechanism in the circulated version is MISIDENTIFIED: invented physiology, born as an internet rumour.

"WHO declared aspartame a carcinogen"

MISIDENTIFIED as headlined. IARC classifying aspartame 2B in July 2023 is PROVEN as a fact about the classification; in the same weeks JECFA found no convincing evidence and reaffirmed the 40 mg/kg/day limit. A hazard class was read as a risk verdict. Causation at normal intakes: UNPROVEN.

"Power lines cause childhood leukaemia"

Causation UNPROVEN — no mechanism, no animal support. The association at the extreme exposure tail is DOCUMENTED (pooled RR 2.00 at ≥0.4 µT — a field level covering 0.8% of children). Dropping that tail to call it "a major danger" is MISIDENTIFIED.

"Cancer is a fungus"

FABRICATED — the audit's only one: an invented identity for cancer, contradicted by routine pathology. Its cousin "cancer is a parasite" is MISIDENTIFIED around a PROVEN core: three named liver flukes really do cause two named cancers. A narrow fact, inflated to a universal.

"Stress causes cancer"

UNPROVEN. The prospective record — 116,056 people, exposures recorded before disease — returns HR 0.97, null overall and at every site. The positive tail of this literature is retrospective, where memory does the damage.

"Bras / abortion / hair dye cause breast cancer"

UNPROVEN, and one member tested-and-failed: prospective abortion data give RR 0.93–0.98 (the alarming version rests on retrospective disclosure bias). Bra-wearing: no aspect associated. Personal hair dye: null over 36 years in 117,200 women.

"Chemotherapy works only 2% of the time"

MISIDENTIFIED — a denominator swap. The cited figure (Morgan 2004; PROVEN as cited) is chemotherapy's contribution to 5-year survival across all diagnosed patients, most of whom never received it; per-cancer contributions run to ~40%. "Chemo kills more than it saves" is likewise MISIDENTIFIED: curative-intent 30-day mortality is 0.26% in breast cancer.

"They are hiding the cure"

UNPROVEN — and its testable predictions fail on the open record: patents are public, rival firms license rival HPV vaccines, and the best-selling drug launch in history was a cure (Sovaldi, $10.28B in its first year). The premise — a single "cure" exists to hide — is MISIDENTIFIED across 100+ diseases.

"Survival hasn't improved in 50 years"

MISIDENTIFIED. Five-year survival rose from 49.9% to 70.5%, and the death rate is down about a third from its peak. The narrower claim — progress is uneven and some sites have stalled — is PROVEN, and this site prints those sites at full weight.

"A fighting spirit improves survival"

UNPROVEN. The 1989 result behind it was never replicated — later trials, including one by its own author, were null. What psychological support does improve is quality of life, and that is PROVEN; grading a quality-of-life intervention by a survival claim is the error.

Cell therapy, vaccines, oncolytic viruses (Division 10)

CAR-T in defined blood cancers is PROVEN — a randomised survival win (ZUMA-7, OS HR 0.73) plus five-year tails — with the bill printed: pooled non-relapse mortality 6.8% across 7,604 patients, half of it infection. "A fraction cured": DOCUMENTED — plateau-shaped tails, no formal cure-fraction analysis yet. "CAR-T works for solid tumours" as circulated: MISIDENTIFIED — one China approval on a progression endpoint whose survival interval crosses 1 (satri-cel, OS HR 0.693, CI 0.457–1.051), one single-centre paediatric series. TCR therapy graduated to full approval (afami-cel, June 2026: response claim PROVEN, survival UNPROVEN — and its headline response rate describes the 44 infused of 373 screened). Personalised vaccines: melanoma DOCUMENTED, pancreatic UNPROVEN, colorectal vaccine-alone UNPROVEN with a documented negative. Oncolytic viruses: local control DOCUMENTED; a survival or systemic effect UNPROVEN for every member of the class — T-VEC's own label says it has not been shown to improve survival, and the largest randomised test (Pexa-Vec, n=459) came out numerically worse.

RAS and menin (Division 9)

Daraxonrasib's survival benefit in previously treated metastatic pancreatic cancer is PROVEN — randomised n=500, OS HR 0.40 against a real chemotherapy comparator — bill included: rash 87%, grade ≥3 treatment-related events 43.6% (less than chemotherapy's 57.5%). The gloss "pan-RAS" is MISIDENTIFIED: per-allele proof exists for G12D and G12V only. A G12D-selective programme died without ever testing its target in humans — MRTX1133 was terminated for exposure failure, no human responses ever published; the G12D concept survives it, UNPROVEN on early signals. Menin inhibitors are PROVEN only in the narrow relapsed/refractory regulatory sense (responses ~21–23%, durations ~5–6 months), and their forecast escape is already measured: MEN1 mutations in 38.7% of analysable patients. Sotorasib's confirmatory progression endpoint was voted not reliably interpretable by the FDA's advisors (10–2); neither G12C drug has shown a survival benefit (CodeBreaK 200 OS HR 1.01).

Checkpoint immunotherapy (Division 2)

Eight accelerated checkpoint approvals were withdrawn in 2020–2022 after confirmatory trials failed — PROVEN on the FDA's own ledger; claims of continuing 2026 checkpoint withdrawals are MISIDENTIFIED (the 2026 withdrawals are other drugs). Roughly 46% of pembrolizumab's 39 indications entered on response rates — survival proof, where it exists, arrived years later. The failures stand at full weight beside the wins: pembrolizumab in gastric cancer and the IDO combination were refuted at Phase 3 (ECHO-301 HR 1.00), and PD-L1 — the flagship biomarker — is measured by non-interchangeable assays; "PD-L1-negative means no benefit" is MISIDENTIFIED. Early-stage use is where the record is strongest now: perioperative/neoadjuvant survival PROVEN in early TNBC and resectable lung cancer (OS HR 0.66–0.72). Adjuvant survival as a class claim: MISIDENTIFIED — true in exactly one programme (kidney, HR 0.62), undemonstrated after nine years in melanoma despite a durable recurrence win. Fatal immune toxicity: 0.36% on PD-1 alone, 1.23% on combinations.

Epigenetics (Division 1)

Azacitidine carries the division's one clean Phase III survival win — in higher-risk MDS (24.5 vs 15.0 months, HR 0.58); its AML primary endpoint missed. Decitabine's survival benefit: UNPROVEN — its approvals rest on response. "Epigenetic priming" of solid tumours: UNPROVEN as a class, and contradicted at randomised level in its best-tested instance — the decitabine-primed ovarian arm produced significantly fewer responses than chemotherapy alone (3/15 vs 9/14, p=0.025). Tazemetostat completed the cautionary arc: two accelerated approvals, then a 2026 FDA safety alert — second blood cancers in 5.7% vs 0% on the confirmatory trial — and global withdrawal; no EZH2 inhibitor now holds an active FDA approval. The exception that explains the rule is IDH: one mutation, one measurable metabolite — ivosidenib survival PROVEN (HR 0.44), even as its sister drug enasidenib missed the same endpoint.

Metabolism (Division 4)

The Warburg phenotype is PROVEN; the drug rationale built on it fails link by link. Human isotope tracing shows tumours oxidising glucose, lactate and other fuels at once — "tumours depend on glycolysis" is MISIDENTIFIED as a general claim, and lactate, the supposed waste product, is a fuel whose contribution to the tumour's energy cycle exceeded glucose's in the direct human test. Glutamine blockade: the flagship telaglenastat failed its large randomised test (CANTATA, HR 0.94) and its biomarker-selected one. Metformin: tested and not shown across more than 5,700 randomised patients — the observational promise behind it was immortal-time bias, documented as such. The division's two wins pass a measurability test the failures don't: the dependency exists before treatment (a mutation's metabolite; an enzyme's absence), the drug's action is assayable in the patient, and normal tissue has a named escape route. IDH inhibition and asparaginase: PROVEN — asparaginase's bill printed too (grade 3–4 liver injury 53.9%, pancreatitis 12.6% in the adult series).

Signalling, part 2 — PI3K/AKT/mTOR and CDK4/6 (Division 3)

Pan-PI3K inhibition is refuted as a tolerable strategy: BELLE-2 met its endpoint on paper while median drug exposure was 1.9 months and 38.7% of patients stopped for toxicity. Isoform selection is the fix the record supports — the inavolisib triplet holds the pathway's strongest readout (survival HR 0.67) with severe high blood sugar in 5.6% vs alpelisib's 36.6%. Capivasertib works in biomarker-altered disease (PFS HR 0.50, altered population); its label is correctly narrower than its trial. CDK4/6 splits by agent: ribociclib's survival is PROVEN (three trials), palbociclib's UNPROVEN (PALOMA-2 OS HR 0.956). Adjuvantly the class splits cleanly — abemaciclib and ribociclib positive, palbociclib negative twice — and the popular explanation ("patients just stopped taking it") is MISIDENTIFIED against PALLAS's own exposure analysis and Penelope-B's 80.5% completion. The arithmetic: adjuvant abemaciclib treats ≈13 patients for two years per recurrence prevented, ≈56 per death prevented, with 18.5% stopping for adverse events.

Chemoprevention (Division 6)

Audited agent by agent: no chemoprevention trial has shown a cancer-mortality benefit as its primary story. Incidence cuts are real and PROVEN in defined high-risk groups (tamoxifen RR 0.51; anastrozole HR 0.51) — with harms printed at the same size: tamoxifen's endometrial-cancer risk RR 2.53, pulmonary embolism RR 3.01; in a pooled accounting, 7–16 invasive cancers averted per 1,000 women against 5–7 venous clots caused. Finasteride cut prostate diagnoses by ~25% while its measured 18-year mortality effect was null (HR 1.02); regulators twice refused a prevention indication. The ledger's darkest entries are supplements: beta-carotene increased lung cancer (+18%) and total mortality (+8%) in smokers across two large trials — "antioxidant supplements prevent lung cancer" is REFUTED. COX-2 inhibitors shrank polyps while roughly doubling serious cardiovascular events. And uptake among ~10 million eligible US women sits at 0.08–0.2%.

Computation and AI (Division 12)

Computation is load-bearing exactly where a claim can be checked in advance and checked in public. AlphaFold's structure prediction is PROVEN at database scale — but it predicts poses, not binding strength, and "AlphaFold solves drug discovery" is MISIDENTIFIED; no approved drug is attributable to it. AI-origin medicines approved anywhere: zero — 117 assets in trials, Phase II success ~40%, the ordinary industry rate; the flagship companies' record includes an absorption, a delisting, and a drug discontinued after meeting its Phase II surrogate. Screening AI has one randomised flagship (MASAI, n=105,934: reading workload −44.3% PROVEN; interval cancers non-inferior, not reduced; mortality UNPROVEN). Digital twins: zero patients have ever been treated on a twin's advice — concordance with the past is DOCUMENTED, personalisation as practised is MISIDENTIFIED. And the canonical failure is graded in full: Watson for Oncology — trained on synthetic cases, unsafe recommendations in internal documents, $62 million spent at MD Anderson, never used on a patient.

Early detection — the blood-test era (Division 7)

Blood tests genuinely detect tumour-derived signal at very high specificity — but sensitivity is rationed by shedding physics, and the tests are weakest exactly where the pitch is strongest. Tumour DNA is detectable in only 47% of stage I cancers (n=640); the methylation flagship's stage I sensitivity is 16.8% in its case-control study, and CancerSEEK's prospective detection was 27.1% against 70% case-control — the design is worth that much. The "screens for 50+ cancers" gloss is MISIDENTIFIED against the FDA's own per-type episode sensitivities: breast 19–23%, prostate 8–11%, thyroid 0 of 4, myeloid 0 of 6. No multi-cancer test is FDA-authorized, and the class's one randomised outcome trial missed its endpoint. The division's one approved detection blood test, Shield (colorectal), is PROVEN — with its label's caution at the same size: 83.1% of colorectal cancers caught, but only 13.2% of advanced adenomas and 54.5% of stage I cancers. MRD, unbundled: prognostic validity in colorectal cancer is PROVEN (recurrence hazard ratio 7.2), but "a negative rules out recurrence" is MISIDENTIFIED — landmark sensitivity 55.6%, peritoneal relapse 40%, and 24% of landmark-negative patients recurred. The harms are measured, not rhetorical: false positives took a median 162 days to resolve — 92% imaged, 49% given at least one procedure. Nothing in this division has yet moved a mortality curve.

Dormancy (Division 5)

The best-evidenced unexplained phenomenon in oncology. That cancer persists invisibly and recurs decades later is PROVEN — ER-positive breast cancer recurs at a steady hazard from year 5 to year 20 (62,923 women), and donor-derived cancers, genetically proven, slept 16–32 years inside a donor and woke in transplant recipients. But the dormant state has never been measured in a patient: that the persisting cells were quiescent is UNPROVEN — slow growth, turnover and immune holding produce identical clocks. Exactly one mechanism node carries a human footprint: NR2F1, whose high fraction in marrow disseminated cells stratifies relapse risk (DOCUMENTED, 86 patients) — as a validated clinical biomarker it remains UNPROVEN, and claims that a dormancy programme has been measured transcriptome-wide in human disseminated cells are MISIDENTIFIED. The therapeutic record is one 51-patient feasibility trial (CLEVER), and its press gloss — "drugs eliminated dormant cells in 80% of survivors" — is MISIDENTIFIED: 80% was an estimated reduction in cell counts, and neither clearance-as-surrogate nor recurrence prevention is shown. Keeping cells asleep, and wake-and-kill, have never been trialled in humans at all. No dormancy-specific therapy is approved. Our own filing against this compartment (T-005, below) stands untested for the same reason the room is empty: the cells cannot be sampled over time.

Genomic profiling (Division 1)

Profile the tumour, match the drug — the strategy's unit of truth turns out to be the tier of evidence behind the match, not the panel. The first randomised test, SHIVA (741 screened, 195 randomised), found nothing: progression-free survival HR 0.88 (95% CI 0.65–1.19), p = 0.41, with matched-arm grade 3–4 toxicity (43%) at least equal to the chemotherapy it replaced — low-tier matching tested and not shown. The tier-restricted test found the benefit exactly where the tiers predicted: SAFIR02, in HER2-negative metastatic breast cancer, randomised the patients whose alterations ranked ESCAT tier I/II (n=115) — progression-free survival 9.1 vs 2.8 months, HR 0.41 (90% CI 0.27–0.61) — tier-restricted matching PROVEN; the same trial's unrestricted comparison was null. The funnel is the discipline: in NCI-MATCH, 37.6% of profiled patients carried an alteration qualifying for some subprotocol, but across the large cohorts only ~1 in 8 profiled patients actually receives a profiling-attributable drug, ~5–7% of advanced-cancer patients benefit, and at the tier that predicts benefit the actionable share is ~8% of profiled patients. "Most patients have an actionable mutation" is MISIDENTIFIED: true only at the funnel's widest stage, under its loosest definition, quoted with the narrow stage's implication. Tumour-agnostic approvals are real — and are the proven core's edge, a low-single-digit percentage of advanced cancers, not a new continent.

Drug repurposing (Division 11)

Two genuine triumphs — thalidomide in myeloma and ATRA in APL, both PROVEN — and then, across four decades of the organised modern pipeline, zero confirmed randomised wins from a catalogue its own advocates count at 373 candidates. The modern canon is UNPROVEN, and its flagship signals collapse inside their own papers' numbers: disulfiram's registry headline (mortality HR 0.66 for continuers) is MISIDENTIFIED — patients who never used the drug show HR 0.68 against the same comparator; the entire "benefit" lives in who the comparator was. The statin association dissolves to HR 1.00 when immortal time is designed out. The most viral claim is the worst-evidenced: "fenbendazole cures cancer" is MISIDENTIFIED — one confounded anecdote whose subject was concurrently enrolled in a pembrolizumab trial — and its claimed clinical evidence base is FABRICATED in effect: the one published case series was retracted in January 2026, no human trial of it exists, and its documented harms are accumulating as liver-injury case reports. The field's own registry (ReDO) counts 373 candidates, 87% off-patent; of its "top six" priority candidates, none has produced a confirmed randomised win in a decade. The conversion failure is in the trials and the funding structure — 7 of 190 late-stage trials industry-sponsored — not in the cataloguing.

Laetrile

Tested properly and failed (NEJM, 1982: 178 patients, no substantive benefit, cyanide toxicity). Its marketing label "vitamin B17": FABRICATED — it is not a vitamin.

Gerson & Gonzalez

UNPROVEN. Never trial-tested in its original form; when the Gonzalez descendant regimen was tested, chemotherapy patients lived a median 14.0 months vs 4.3 on the regimen.

Hoxsey

UNPROVEN, cure lists MISIDENTIFIED: review of 400 claimed cures found people who never had cancer, had prior conventional treatment, or had died.

Burzynski antineoplastons

UNPROVEN after roughly 50 years: originator-run evidence only, zero published randomised trials, no independent replication.

Warburg-derived diets

The Warburg effect itself is PROVEN biology. "Alkaline diet cures cancer" is MISIDENTIFIED — tumour acidity is a consequence, not the cause; a review of 8,278 citations found zero treatment studies behind the claim.

Vitamin C — three claims, three grades

The original claim UNPROVEN; the oral claim killed by two double-blind Mayo trials; the intravenous form is a genuinely different exposure and remains UNPROVEN, with mixed small Phase II results. It reappears in our register — under strict rules (T-003) — and that filing's own check has since killed it; see the register below.

RAS(ON) first-line in pancreatic cancer

Move multi-selective RAS(ON) inhibition into first-line treatment, allele-stratified, in KRAS G12D/G12V metastatic pancreatic ductal adenocarcinoma — building on the later-line survival signal (13.2 vs 6.7 months). Kill test: a randomised first-line trial showing no survival benefit.

Check, Round 1: SUPPORTS. G12D+G12V are 71–72% of KRAS-mutant cases in two cohorts (filed bar ≥50%) and the tumours' KRAS dependency measures −1.97 (bar ≤ −1). Support is not clinical evidence — grade unchanged.

Register annotation, 2026-11-03 (Tranche 17): the drug class's survival benefit is now PROVEN in the adjacent line (RASolute 302, randomised n=500, OS HR 0.40) — but the allele direction runs against this filing's "concentrated in G12D" prediction (G12V HR 0.32 vs G12D HR 0.49), recorded here; the kill test is not triggered. And the filed combination rests on one n=40 cohort at a lightened chemotherapy schedule — the full-standard-backbone combination remains UNPROVEN. Grade unchanged.

Co-mutation-gated menin triplet in AML

In relapsed/refractory NPM1-mutant acute myeloid leukemia, the third drug in a menin-inhibitor triplet should be chosen by FLT3/RAS co-mutation status — aimed at the documented resistance ceiling.

Register annotation, 2026-11-03 (Tranche 17): the menin premise is graded — PROVEN only in the narrow relapsed/refractory single-arm regulatory sense, with MEN1-mutation escape measured at 38.7% of analysable patients as the benchmark any triplet claim must beat. The gating claim itself — that co-mutation status should pick the third drug — has never been prospectively tested, and the menin + FLT3 substitution arm has no results at all (registry only): the least-evidenced load-bearing element in the filing. Grade unchanged.

IV ascorbate, biomarker-stratified

Pharmacologic intravenous ascorbate in metastatic pancreatic cancer, restricted to a low-catalase / pro-oxidant signature with exposure verified. Filed so it dies even if an unstratified effect appears.

Register annotation, 2026-11-03: KILLED-BY-CHECK. In-silico Round 1 found the filed biomarker panel incoherent in the primary cohort (mean correlation −0.04) and inconsistent in the replication cohort (+0.29) — the filing's own kill prong. This kills the stratification claim as designed, not ascorbate itself; per the filing's pre-commitment, the lab will not re-file it unselected. The filing stays published, as filed.

Neoantigen vaccine in ctDNA-positive colorectal cancer

Individualised neoantigen vaccine plus PD-1 blockade after resection of stage III colorectal cancer where ctDNA shows residual disease — with an in-silico antigen-supply check that can withdraw the MSS stratum.

Register annotation, 2026-11-03 (Tranche 16): evidence against strengthened. BNT122-01 — autogene cevumeran alone in this filing's own population (ctDNA-positive resected stage II/III colorectal cancer) — crossed its futility boundary and was terminated after a monitoring-board survival imbalance; it is the nearest human test of the vaccine-alone component and is entered at full weight. It does not test the filed combination. Checks, Rounds 2 and 4: CANNOT RUN — the filed antigen-supply test has now failed on data availability three times (see Lab Checks); the MSS stratum's antigen supply remains untested by this lab — neither supported nor killed. Grade unchanged.

Register annotation, 2026-11-04 (Spec v2): the check is PARKED, not closed. The specification was re-frozen with the study-frame drift repaired by artefact pin — the frame now verifies against a hashed artefact of the sample list and MANTIS map (SHA-256 a1efa9cf…b15b8), so any future drift aborts against an artefact, not a memory. But no open HLA call set for this cohort exists: every route was verified first-hand (the pinned table is controlled-access; the study carries no HLA data; the one published call set was never released per-patient; deriving calls needs controlled-access reads). The specification stands frozen with that input specified and unfilled; no run is authorised. Watch trigger: TCIA's per-sample tables returning, or any verified open HLA set for the cohort. Grade unchanged.

Eradicating dormant disseminated cells in breast cancer

Hydroxychloroquine plus everolimus against dormant disseminated tumour cells in survivors who carry them. If clearing those cells does not reduce recurrence, this filing kills the surrogate for the whole lab.

Checks, Rounds 1, 2 and 4: CANNOT RUN — three times. A refreshed census of every public dataset confirms none meets the filing's acceptance criteria: the only transcriptome-wide human dormant-cell dataset in existence is controlled-access raw reads with no proliferative comparator, and the nearest public set measures the wrong genes in the wrong comparator. No proxy was substituted. The filing's biology is untested by this lab — neither supported nor killed. Grade unchanged. What would make this check runnable is now the standing requirement, specified exactly in the dormancy deep dive above: human disseminated cells profiled directly and transcriptome-wide with the filing's pathway actually measured, a proliferative comparator, at least two independent datasets, open access, and the patients' recurrence clocks attached — today the count of qualifying datasets is zero.

Induced "BRCAness" + PARP in ovarian cancer

DNMT inhibition to induce a BRCA-like state, then PARP inhibition, in BRCA-wildtype / HRD-negative high-grade serous ovarian cancer — gated on a paired-biopsy measurement of whether induction actually happens in human tumours.

Register annotation, 2026-11-03 (Tranche 19) — premise audit: of the filing's six premises, the two with human measurement are DOCUMENTED (low-dose DNMT inhibition demethylates measurably in patients; the target population is ≈50% of this cancer type) — and neither is the premise the theory turns on. Repair-gene suppression in human tumours, functional conversion to induced HRD, and responses concentrating in measured-induced patients have no human measurement of any kind; the decisive biomarker-embedded trial has never been run. Nothing strengthens the filing and nothing yet kills it — because the experiment that would do either has never been run in a human tumour. Grade unchanged.

Measured delivery in glioblastoma

Implantable ultrasound opens the blood-brain barrier for albumin-bound paclitaxel in first-recurrence glioblastoma — graded not on the device but on the measurement: sonication must cover the cavity plus 2 cm margin, and drug concentration is measured in the non-enhancing margin tissue where recurrence actually starts.

Check, complete — Round 4: both arms have now run as filed. The barrier arm supports (Round 2: margin transporter expression 1.46× against the ≤1.5× filed bar). The potency arm, run in Round 4 on the filed dose-response source Round 2 could not reach, reverses Round 2's substituted read: across 23 glioblastoma lines, ABCB1 expression tracks paclitaxel resistance significantly (ρ = +0.484, p = 0.019), and 12 of 23 lines never reach 50% kill at any tested dose. The filing's kill is conjunctive and its barrier-arm condition failed, so the kill does not fire — but the check does not support the filing either. The efflux objection is upgraded from monitored risk to measured liability; what remains is what was never in-silico: the human margin-concentration measurement, the efficacy stage, and the SONOBIRD cross-kill. Grade unchanged.

PRMT5 inhibition in MTAP-deleted mesothelioma

An MTA-cooperative PRMT5 inhibitor in dual-confirmed MTAP-deleted relapsed pleural mesothelioma, with paired-biopsy proof the mechanism engages in this tissue and a response bar set against the documented precedents — low as they are.

Register annotation, 2026-11-03: KILLED-BY-CHECK. In-silico Round 2 ran the check this filing itself specified — gene-knockout dependency: DepMap shows no MTAP differential for PRMT5 dependence (MTAP-null median −0.914 vs intact −0.982, the wrong direction, p = 0.43). The population exists — MTAP deletion in 41.4% of this cancer, above the filed bar — but the dependency claim does not survive. The caveat is printed, not buried: knockout is not the drug's partial, cooperative inhibition, so the biology is not refuted — but the filing named this check, and it dies by its terms. Any revival had to be a new filing with a pharmacologic check; that filing is T-010 below. The filing stays published, as filed.

Enriched early detection of pancreatic cancer

Two-stage detection in new-onset diabetes after 50: a risk score gates dedicated imaging across the 3-year risk window. Enrichment is the theory — modelled predictive value 31–57% inside the cohort vs 0.5–1.6% in the general population. No mortality claim is filed.

Check, Round 2: SUPPORTS. 100,000 Monte-Carlo draws over the filed ranges, with the prospective incidence figure fetched first-hand: the predictive-value floor, the imaging-volume cap, and the harm budget hold together at the pessimistic quartile of every parameter (PPV 33.1% there), and the kill prong fires in 0.00% of draws. Nothing further can be settled in-silico; the gap is prospective operating data, which only the field can produce.

PRMT5 in mesothelioma — the successor filing

T-008's claim, population, and human kill criteria, unchanged — re-filed under the lab's dose/test rule (L6) with the one thing that failed replaced: a pharmacologic check (drug-sensitivity screens across MTAP-null vs intact cell lines) instead of the knockout check that killed its predecessor. The filing prints its own limit: the cooperative compounds behind the actual drug idea are in no public screen, so a SUPPORTS verdict here is not evidence about the cooperative mechanism itself. And a pre-commitment: if this check kills it too, the PRMT5–mesothelioma slot closes permanently. No third filing.

Register annotation, 2026-11-03: KILLED-BY-CHECK. In-silico Round 3 ran the pharmacologic check behind the filing's mandatory positive control — which passed decisively: BRAF-inhibitor differentials against BRAF V600, computed identically, measured +1.14 to +1.79, so the screen could plainly have seen a selectivity signal of the filed size. It saw none for PRMT5: differentials by MTAP status were −0.055, +0.088, and +0.018 for three of the four inhibitors (reversed, or below the 0.15 kill line), +0.258 intermediate for the fourth — SUPPORTS for 0 of 4. The pre-commitment printed above is therefore executed: the PRMT5–mesothelioma slot is permanently closed — no third filing from this lab without new human efficacy data in the disease. The filing stays published, as filed.

Perioperative blockade in resected colorectal cancer

Propranolol plus etodolac for a 20-day window starting five days before curative surgery, in stage II–III colorectal cancer with no detectable metastasis — aimed at the colonisation step while the surgical-stress window stirs the disseminated reservoir. This is the lab's first filing against metastasis as a process, and it was built against the five bars in the Metastasis section: M0 population, a claim of reduced distant-recurrence proportion at five years (not a delay), survival pre-committed, engagement measured in the excised tumour. Its pilot signal (five-year recurrences 2/16 vs 9/18) is printed in the filing in both For and Against — n=34, one research group, full originator discount.

Register annotation, 2026-11-03: KILLED-BY-CHECK — withdrawn to watch status, as its filing provides. In-silico Round 3 ran the filed check on the PanCanAtlas colorectal cohort. In the analysable stage II–III subset (n=163), the EMT composite showed no association in the predicted direction: stage-adjusted HR 0.564 (95% CI 0.192–1.652, p = 0.296) — the point estimate in the protective direction, and neither single-gene axis live either. The power caveat travels with the verdict: the subset holds only 25 disease-free-survival events, so the kill clause fired on a wide interval. Withdrawn to watch status is the filing's own disposition for exactly this outcome; the filing stays published, as filed.

Switch at the signal in EGFR-mutant lung cancer

Switch to osimertinib when the resistance mutation T790M appears in blood, before scans show progression — the ESR1-switch shape (the one strategy with PROVEN decision value, Resistance section above) transplanted to a second setting. The caveats lead, not follow: the direct precedent trial showed feasibility only, which the filing prints as its main evidence against; its first-progression endpoint could be pure lead time and is pre-registered as non-decisive, with the decision resting on later endpoints and survival; and the filing retires itself if first-line osimertinib becomes universal and its population disappears.

Register annotation, 2026-11-03: KILLED-BY-CHECK. In-silico Round 3 ran the filing's own decisive test — a Monte-Carlo over the parameters its evidence rests on (100,000 draws). Under the survival effect the filing designated as its conservative input (AURA3 overall-survival HR 0.87), the modelled overall-survival HR median is 0.999 and zero draws clear either filed bar; an advantage appears only when the progression-free effect (HR 0.30) is used as the timing input — the misreading this check was built to expose. The plain conclusion: the switch-at-the-signal benefit, on the numbers available, is indistinguishable from lead time. The filing stays published, as filed.

Arginine deprivation in ASS1-deficient mesothelioma

Pegargiminase plus platinum–pemetrexed, first-line, in ASS1-deficient epithelioid pleural mesothelioma — extending to the majority histology the survival benefit the same regimen proved in biomarker-enriched nonepithelioid disease. ASS1 silencing makes the tumour an arginine auxotroph; the drug degrades circulating arginine; ASS1-expressing normal tissue resynthesises it and is structurally spared — the asparaginase architecture, and the dependency is measurable before treatment and assayable during it.

Evidence against leads, as filed. The epithelioid claim has never been randomised: its entire human record is 11 uncontrolled patients inside a phase 1 expansion (n=32) — everything else is extrapolation from a different histology. The benefit it extends is real but modest (ATOMIC-Meso median survival 9.3 vs 7.7 months in nonepithelioid disease) in a population that lives substantially longer on chemotherapy, so the headroom is smaller; the comparator era has moved to chemo-immunotherapy; the named dominant escape — macrophage argininosuccinate fuelling and tumoral ASS1 re-expression — is already measured in patients and the regimen does not address it; the ASS1-deficiency cutoff sits in a laboratory manual, not a published standard; and the programme is one academic–industrial axis, priced with the class harm ledger beside it.

Evidence for: ATOMIC-Meso — randomised, double-blind, n=249 in nonepithelioid disease — overall-survival HR 0.71 (95% CI 0.55–0.93); activity across histologies in the expansion cohort (disease control 93.5%); plasma arginine depletion demonstrated and maintained in patients.

Check, Round 5: SUPPORTS. The population premise, scored on the filed bars in the TCGA cohort: in the epithelioid subset (analysable n=62), the ASS1-low fraction is 38.7% (24/62), the 95% CI lower bound at 27.6% against the filed 20% floor — the biomarker population exists at trial scale. The mechanism arm is printed openly, as the filing pre-wrote it: the predicted silencing-by-methylation was not confirmed at the promoter probes (the association lives only at a probe selected for expression anticorrelation, partly circular by construction; the low state is not deletion-driven) — the biomarker is carried as expression-defined. Support is a verdict on the check, not clinical evidence — grade unchanged.