# Adversarial review log - mced-cancer-mortality (shipped review)

Canonical file: `mced-cancer-mortality.review.md`. This is an independent second adversarial pass over the shipped review. The
review already carries an `## ADVERSARIAL REVIEW AND APPROVAL` section from a first pass dated 2026-10-14 (findings M1-M4);
that section was read and its claims were re-tested here rather than relied on. Nothing in the review, score card, site pages or
flags was edited in this pass; every correction below is a proposal (exact original text to exact corrected text) for the lead.

Level before -> after: Γ -> Γ (unchanged). Confidence High -> High (unchanged; rationale should be rewritten, see S8).
Findings: 2 Major, 8 Minor, 1 Editorial. PMIDs verified 58/58 (plus 6/6 linked notices). Claims spot-checked 22. Validator: not re-run (no file changed).

**Two claims in the earlier approval section do not survive re-checking.** (a) Its rejected finding (v) says the five screened
discriminating contrasts are "the subset that actually contrast allocated arms". The score card's `evidence_inventory` flags
`discriminating_contrast: true` for [1], [10], [18], [45] and [51]: two of these ([10] PATHFINDER, [18] real-world series) are
single-arm series with no comparator, while the randomized NLST [48] and NELSON [49] contrasts are flagged false (S5).
(b) Its M1 says fused words were repaired in "every source-specific evidence card"; fused strings remain in cards [33]-[58] and
in the reader summary, web-page text and Consensus cross-check (S9).

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1 (as stated in the review header; the score card and site page record "Codex", see S11). **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-11-02.

**Scope of checks.** All 58 REFERENCES PMIDs verified against PubMed `get_article_metadata`: 58/58 resolve to the stated first
author (or collective author), journal, year and DOI; 0 failures; no retraction or expression-of-concern flag. Year offsets of
one at references 9, 39, 42, 44 and 46 are e-pub versus print. The six linked notices named in the evidence cards (36747644
preprint; errata 39228301, 41595793, 41628341, 41812212, 42119586) also resolve. Twenty-two load-bearing claims were checked
against the cited abstracts and, where available, PMC full text: the NHS-Galleri primary report (full author manuscript,
PMC7619610) for randomisation total, age band, primary IRR/CI/P, the stage IV secondary and its testing status, stage III and IV
counts, the fatal-cancer composite and the mortality-power statement; ClinicalTrials.gov NCT05611632 for the registered
mortality outcome; the NHS performance report (PPV, sensitivity, specificity, 937 cancers); CCGA (51.5%, 16.8%, 90.1%,
4077/2823/1254); PATHFINDER (6621, 92, 35, 57, 162 days, 17/57 procedures); DETECT-A (26/24/46, 9 of 15, 0.22%); the
real-world series full text (PMC12578847: 111,080; 1011; 459; 411; 128/259; 49.4%); UKCTOCS Lancet and HTA reports; NLST and its
extended follow-up; NELSON; SHIP biopsy rates; CAP 15-year; NordICC 10- and 13-year; ERSPC 23-year; UK Age; the DTC series; and
the official trial page on mortality follow-up. Every quoted figure matched its source; the defects below are omissions,
framing and internal consistency, not misquoted numbers. Scoring logic was re-derived: caps precede aggregation; the
declared `primary-subclaim` rule returns Γ from SC2; weakest-link returns Α from SC4 and is correctly reported only as a
sensitivity; the Β/Γ/Δ boundary holds (no positive Galleri mortality crossing, so Δ is blocked); Ζ independence is not claimed;
M/S/D carry the `calibration_note` and do not alter the level. Targeted searches for missing literature found two directly
relevant original reports not in the review (PMID42773207, PMID38583868).

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| S1 | Major | AT A GLANCE "Causal role" cell; Final score paragraph; rendered page line "What this mechanism does" (`ROLE_PLAIN` in `tools/build_v5_reviews.py`) | "epiphenomenon" with the site gloss "This travels alongside the process without causing it or being caused by it." to the label retained but qualified: "epiphenomenon (framework step 2: no effect on cancer deaths demonstrated in either direction; not a finding of no effect)", and a per-case gloss "No effect of this screening on cancer deaths has yet been shown in either direction; the label records an unproven link, not a proven absence of effect." | The review itself says "neither proves a mortality benefit or a mortality null", yet the published plain-language line tells a lay reader that Galleri screening does not cause, and is not caused by, the outcome. For a cancer-screening test this is a medical claim stronger than the evidence. Randomised allocation demonstrably changed the diagnostic process (NEJM author manuscript PMC7619610: emergency presentations 225 vs 286; stage IV IRR 0.74, 0.57-0.95, in round 3), so "travels alongside without causing" is also wrong for the screening pathway. The step-2 label follows house convention and does not alter the level; the defect is the unqualified lay gloss. |
| S2 | Major | Final score "Momentum" paragraph; `extended_scoring_MSD.momentum` (value, artifact, justification); SC2 matrix row; Stability trigger | "Momentum: Decisive work underway. The registered NHS-Galleri trial (NCT05611632/ISRCTN91431511) remains in prespecified cancer-mortality follow-up; the official trial site states that cancer-death analyses continue and the primary report states that further follow-up is warranted." to "Momentum: Mechanistic only. NHS-Galleri lists cancer-specific mortality up to 8 years as a prespecified secondary outcome (NCT05611632), and the primary report states that cancer-specific and nested mortality will be analysed at 3 and 6 years after the final trial visit, but also that the trial was not powered for mortality endpoints; after the failed primary endpoint its fixed-sequence plan precludes formal testing of secondary outcomes. This follow-up is informative but cannot by itself decisively cross the SC2 threshold." Add to the SC2 row after "Official trial information states that mortality follow-up continues": "; the trial was not powered for mortality endpoints." | Primary report full text (PMC7619610, DOI 10.1056/NEJMoa2505723): "Cancer-specific and nested mortality will be analyzed at 3 and 6 years after final trial visit as secondary endpoints; the trial was not powered for mortality endpoints", and "Testing stopped at the first non-significant step". ClinicalTrials.gov NCT05611632 lists cancer-specific mortality as SECONDARY. The abstract's "Further follow-up is warranted" refers to stage III/IV and stage IV outcomes, not deaths, so the artifact misattributes it. The official trial page mentions reducing deaths only as a future question and says nothing about power. The review's own deciding experiment requires an "adequately powered" mortality comparison, which no identified study is. M/S/D do not alter the level, but a lay reader is told a decisive answer is coming from a study its authors say cannot provide one. |
| S3 | Minor | AT A GLANCE headline; FULL REVIEW NHS paragraph; inventory [1]; card [1] | "The secondary stage IV ratio was 0.86 (0.74–1.00)." to "The key secondary stage IV ratio, also among 12 prespecified cancer types, was 0.86 (0.74–1.00); because the fixed-sequence plan stopped at the failed primary endpoint, this result was not formally tested and its interval is unadjusted. The stage IV decrease (341 vs 394 participants) was offset by more stage III cancers (360 vs 287), and a prespecified secondary composite of stage III/IV or fatal cancers across routinely staged types was null (IRR 1.04, 0.97–1.12)." Headline "its favorable secondary stage IV result requires clinical follow-up" to "a lower stage IV rate in a secondary endpoint that could not be formally tested requires clinical follow-up". | PMC7619610: "the prespecified hierarchical plan ... precluded any formal testing of secondary outcomes"; CIs "should not be used for hypothesis testing"; stage III 360 vs 287, stage IV 341 vs 394; composite IRR 1.04 (0.97-1.12). The stage III excess and the null fatal-cancer composite are the most direct counter-evidence in the trial to the "favorable signal" framing and bear on SC3. |
| S4 | Minor | Subclaim matrix SC1 row; lay page "Key points"; SC1 literal Γ hard rule "In vitro-only" | SC1 states an existence claim ("Galleri detects cancer-associated signals before clinical diagnosis") but is graded Γ for "the unbridged inference from that diagnostic signal to meaningful clinical benefit", which is SC2's content; the lay page prints "SC1 ... HSEF Γ, High confidence" so readers see demonstrated detection as only "plausible". Corrected: reword SC1 to the proposition actually graded ("Galleri-detected signals translate into clinically meaningful earlier diagnosis") or mark detection as demonstrated and descriptive with no causal letter. Hard-rule line "Applies to the uncontrolled or cell-only positive support class, with ceiling Γ; it does not upgrade that class or describe the whole source set." to "Not triggered: SC1 evidence is human prospective screening and randomized-arm diagnostic data; no in vitro-only class is offered." | NHS performance report (PMID42773209): 937 MCED-detected primary cancers; PATHFINDER 35 confirmed cancers; PATHFINDER 2 (S7) 0.54% detection rate. No in vitro study is assigned to SC1 in the inventory. The review's own pitfall "A phenomenon graded as though it were a claim" applies to its SC1 row. SC1 does not bind the headline under primary-subclaim, so no level changes. |
| S5 | Minor | Decisive-perturbation inventory tally sentence; `causal_tally_audit.screened_discriminating_contrasts`; `evidence_inventory[].discriminating_contrast` | "Tally: 0 positive qualifying study units across scoped subclaims; 5 screened discriminating contrast units." to "Tally: 0 positive qualifying study units across scoped subclaims; 4 screened discriminating contrast units, the randomized allocated-arm contrasts [1], [48], [49] and [51]. [45] is an observational pre/post contrast and [2], [10] and [18] are single-arm descriptive series." | Score card flags `discriminating_contrast: true` for [1], [10], [18], [45], [51] and false for the NLST and NELSON RCTs. PATHFINDER (PMID37805216) and the real-world series (PMID41173830) have no comparator arm. The earlier approval's rejected finding (v) mis-describes this set. The positive tally (0) is unaffected. |
| S6 | Minor | FULL REVIEW comparator paragraph; inventory [51]; SC3 rationale; Rank 1 and Rank 3 models | "UKCTOCS Methods/Results show 47.2% more stage I and 24.5% fewer stage IV ovarian/tubal diagnoses with multimodal screening, while mortality did not fall after 16.3 years" to "... while the combined stage III/IV reduction (10.2%, 95% CI −21.3 to 2.4) was not significant and mortality did not fall after 16.3 years (PMID33991479). Across 41 screening RCTs, reductions in stage III–IV incidence tracked mortality reductions closely for ovarian (ρ 0.99) and lung (ρ 0.92) cancer but weakly for colorectal (ρ 0.39) and prostate (ρ −0.69) cancer (PMID38583868), so stage-surrogate validity is site specific rather than universally absent." | UKCTOCS abstract (PMID33991479) and HTA report (PMID37183782) give the non-significant stage III/IV change. Feng et al., JAMA 2024 (PMID38583868, DOI 10.1001/jama.2024.5814) is the systematic evidence on SC3's exact question and is absent. It supports the review's site-specific verdict but shows UKCTOCS is a stage IV-only dissociation, which is precisely NHS-Galleri's pattern; that sharpens rather than weakens the caution. |
| S7 | Minor | FULL REVIEW, after the PATHFINDER paragraph; REFERENCES | No mention to: "PATHFINDER 2 enrolled 35,878 adults aged 50 or older without clinical suspicion of cancer; among 32,007 performance-analyzable participants, detection rate was 0.54%, PPV 60.3%, episode sensitivity 39.3% (69.8% for 12 prespecified cancers), specificity 99.64%, and 213 (0.6%) had invasive procedures after a positive test. Endpoints were descriptive and the authors state that randomized trials and longer follow-up are needed to assess clinical utility (PMID42773207)." | PMID42773207 (Nat Med, 22 Sep 2026, DOI 10.1038/s41591-026-04618-w), published the same day as the NHS-Galleri report and inside the 2026-10-13 cutoff; the NHS-Galleri report itself cites it as concordant. It is the largest prospective single-arm Galleri screening study and strengthens SC1 and the harms record without touching SC2. |
| S8 | Minor | `final_score.confidence_rationale`; `reader-summary.json` limitations | Rationale "Fifty-eight primary reports include seven large randomized screening trials (NHS-Galleri 142,250; NLST 53,454; NELSON; UKCTOCS 202,562; NordICC 84,585; CAP 415,357; ERSPC 162,236) and multiple independent MCED assay modalities ..." to "High reflects that the ceiling is well characterised: a large randomized Galleri trial with registry outcomes, a concordant large prospective study and a real-world series all leave no Galleri mortality contrast, so the evidence securely does not justify more than Γ. Other-site mortality trials inform SC3 and the tournament, not Galleri replication." Reader summary "density, heterogeneous controls and incomplete access set the confidence, not correctness." to "Confidence reflects evidence density and reproducibility, not access depth or correctness." | Six of the seven trials cited are other modalities the review repeatedly says "do not transfer" to Galleri, so they cannot be the replication base for a Galleri grade. The framework permits (Γ, High) when "we are confident the evidence does not yet justify more than Γ"; that is the defensible basis, so **High stands**. The reader summary still carries the access-depth wording the earlier M4 struck from the score card. |
| S9 | Minor | Evidence cards [33]-[58]; `reader-summary.json`; `web-page.md`; `CONSENSUS_CROSSCHECK.md`; score card `final_score.rationale` and `causal_tally_audit` | e.g. "Referral-enrichedvariantdenominator; does not mean 40% ofall DTCcustomersortestsarewrong" to "Referral-enriched variant denominator; does not mean 40% of all DTC customers or tests are wrong"; "Avoiduncorrectedfinegrainsubgroup CIwhichabstractformatsanegativeupperbound" to "Avoid the uncorrected fine-grain subgroup CI, which the abstract prints with a negative upper bound"; "Gallerioreverycancer", "MCEDmortalityevidence", "UKCTOCSoriginalcomprehensivefollow-upretainsmortalitynull", "Quasiexperimentalnotrandomized", "the2026", "stageIV", "uses142924 enrolled participants aged50–79", "The58 report count" to spaced forms. | Known renderer defect; the earlier M1 claimed every card was repaired. A scan for tokens of 25+ letters finds about 30 remaining in cards [33]-[58]. The [55] note is substantively right: the NordICC 13-year abstract (PMID42102826) prints the men's RR as "0·77 [0·64 to -0·88]". No content change. |
| S10 | Minor | COMPETING MECHANISTIC MODELS | Rank 1 ("Screening benefit is site, assay, risk and pathway specific", Ε) is a framing compatible with the nominal, and its `explains_what_nominal_does_not` repeats its core claim. Add a genuine rival: "Shedding-biased detection: cfDNA tests preferentially detect high-shedding, biologically advanced cancers, so detections concentrate where an earlier diagnosis changes outcomes least." | CCGA sensitivity 16.8% stage I vs 90.1% stage IV (PMID34176681); NHS episode sensitivity 26.7-37.2% (PMID42773209); NHS-Galleri excess stage III 360 vs 287 and the authors' own statement that "cfDNA shedding is a hallmark of more aggressive cancers" (PMC7619610). The tournament informs primacy only, so no level changes, but as printed the field is weaker than the evidence allows. |
| S11 | Editorial | Review header vs score card `review_metadata.reviewer`, `codex-correction-provenance.json` and site page "Review attribution (recorded)"; `reader-summary.json` `redteamReviewed` | Header "Generated by Sol 6.1"; score card and site page "Codex"; provenance `generated_by: "Codex; HSEF v5.3"`. `redteamReviewed: false` although the review carries an approval. To be reconciled by lead adjudication. | This log records the authoring model the review header states, as instructed; the conflict is flagged, not resolved. |

**Rejected findings (checked, no change).**
(i) SC2 at Β rather than Γ: argued. Randomised allocation produced measured system-level changes short of mortality (more
stage I/II, fewer emergency presentations, lower stage IV in incident rounds) and the mechanism has positive precedents in other
screening programmes. "Partial system support with an unverified mortality bridge" is a fair description; Γ holds.
(ii) Aggregation: `primary-subclaim` returns Γ from SC2, matching the headline; weakest-link Α (from SC4) is correctly reported
as a sensitivity only.
(iii) The cancer-stem-cell literal criteria ("Sphere formation alone", "Xenograft tumor-initiation") are disclosed as an
unchanged rubric quotation with an explicit clinical adaptation. This is not undisclosed template bleed; only the one misapplied
SC1 line is raised (S4).
(iv) Every quantitative statement spot-checked matched its source, including the NHS 142,250 versus 142,924 distinction (the
latter is the enrolment-paper figure, PMID39862108; the NEJM manuscript gives 142,826 consented and 142,250 randomised).
(v) NLST collective author renders correctly; no leading-comma or organisation-fragment defect.
(vi) Stability "Uncertain" is defensible: an underpowered secondary mortality analysis could still constrain the boundary.
(vii) The Δ block is correct: no positive exact-assay cancer-death endpoint exists.
(viii) No retraction or expression of concern on any reference; the two errata attached to [55] and [56] are content-unverified in
this pass as in the original.

**Scorecard change.** None applied. Proposed: `role_of_mechanism` gloss qualifier (S1); `extended_scoring_MSD.momentum` value and
text (S2); `causal_tally_audit.screened_discriminating_contrasts` 5 to 4 and the matching inventory flags (S5);
`final_score.confidence_rationale` (S8). Level, confidence and subclaim levels unchanged.

**Residual limitations.** The NEJM report was read from the PMC author manuscript, not the publisher's version of record; figures
match the PubMed abstract. The supplementary appendix (secondary-endpoint tables) was not examined. Scite was not queried.

**Approval:** Not granted in this pass. The level and confidence stand, but the two Major findings concern what a lay reader is
told about whether Galleri affects cancer deaths and whether a decisive answer is coming; approval should follow their
correction. - Claude Opus 5.5 (Anthropic), 2026-11-02.