# Adversarial review log - smoking-lung-carcinoma (revised edition)

Canonical file: `consensus_consistency_audit_2026-10-09/revised/smoking-lung-carcinoma/smoking-lung-carcinoma.review.md`.
No propagate_to copies and no divergent copies in the inventory.
Level before -> after: Θ -> Θ (unchanged). Confidence High -> High (unchanged; the rationale already met the lead's confidence-band test).
Findings: 1 Major, 1 Minor, 0 Editorial. Identifiers verified 18/18. Claims spot-checked 11. Validator: PASS (0 fail, 0 warn, 33 pass; was 0 fail, 1 warn, 32 pass).

**Authorship note.** This edition is a cross-model red-team rebuild by Claude under HSEF v5.4, not a Sol 6.1 product; the
score card's provenance header says so. The round-2 banner was adapted rather than asserting Sol 6.1 authorship.

**Regression check (round-2 scope note).** The 2026-10-09 revision introduced no error the earlier edition did not have.
The earlier edition was a thin three-source concise review that asserted the same letter with none of the machinery; the
independence defect corrected here (K1) was introduced by the rebuild's own, much stronger independence basis and is a flaw
in new work rather than an inherited one. It is recorded as such.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** not Sol 6.1. This edition is a cross-model red-team **rebuild** performed by Claude (Anthropic) on
2026-09-30 under HSEF v5.4, replacing a thin three-source concise review authored by a different model family; the score card's
`provenance_header.generated_by` records the same. The standard round-2 banner wording was adapted accordingly rather than
asserting an authorship the document contradicts. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.

**Scope of checks.** All 18 REFERENCES identifiers verified against PubMed `get_article_metadata`: 18/18 resolve to the stated
first author, journal, year and DOI; 0 failures; no retraction or expression-of-concern flag. Eleven load-bearing quantitative
claims were spot-checked against the cited abstracts, and every figure matched: the pooled case-control totals (13,169 cases,
16,010 controls) with the three histology-specific odds ratios and their intervals, and the finding that risk does not return to
never-smoker baseline even 35 years after cessation; the contemporary-cohort lung-cancer death relative risks of 25.66 in women
and 24.97 in men; the 34,439 male British doctors, the roughly ten years of life lost, the tripling of age-specific mortality in
the 1920s birth cohort, and the cessation gains of about 10, 9, 6 and 3 years at ages 30, 40, 50 and 60; the UK cumulative risks
by age 75 of 10, 6, 3 and 2 per cent against 16 per cent for continuing smokers, and the more-than-90-per-cent avoidance when
stopping before middle age; the US figure of about 90 per cent of excess mortality removed by cessation before 40; the mouse
inhalation study's 7.25-fold adenocarcinoma and 4.6-fold adenoma-and-papilloma increases at n = 330 against 326; the 7 of 15
versus 4 of 81 EGFR-mutation split; the 16 per cent per 100 Bq/m3 usual-radon increase; and the 5,243 cancers in the mutational
signature analysis. Scoring logic was re-derived: caps precede aggregation, the prescribed `scope-matched` rule sets the headline
from SC1 (the only subclaim whose scope matches the central claim) and the narrower SC2 and SC3 are reported separately, which is
the correct reading of that rule; weakest-link Ζ and primary-subclaim Θ are both reported; the Β/Γ/Δ boundary and the Ζ-plus
independence basis were checked. Framework validator (v5.4, 2.4) re-run.

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| K1 | Major | `independence_basis`, exposure-accumulation line; AT A GLANCE "Strongest positive evidence" row; SC2 section | Exposure-accumulation line sources [1, 2, 4, 7, 8] to [1, 2, 7, 8]; the headline evidence row re-cited as dose-response [1,2,7,8] and cessation [3,4,5,6]; an explicit shared-cohort sentence added to SC2 | References [3] and [4] are two follow-up reports of one cohort, the British Doctors Study, so [4] was being counted in the exposure-accumulation line while [3] and [4] carried the exposure-withdrawal line. The framework rule is that independence counts studies, not reports, because one cohort's systematic error would defeat all of its reports. This matters at Θ, where the letter rests on an independence basis above Ζ. The cohort is now counted once, on the withdrawal line where it uniquely supplies the cessation gradient; the accumulation line still rests on four genuinely separate investigations and is not weakened. The framework validator moved from WARN to a clean PASS on this check as a result. No level, confidence or subclaim letter changed. |
| K2 | Minor | `citation_ledger` and `evidence_inventory`, all 18 rows | No `study_id` on any row to a distinct `study_id` per study, with rows 3 and 4 sharing `BRITISH-DOCTORS-STUDY` | The validator warned that 18 ledger rows carried no study identity, with the fix "tag rows so a reader does not read N records as N independent investigations". Tagging them is what surfaced K1. |

**Rejected findings (checked, no change).**
(i) Confidence band High: the stated rationale already rests on evidence density and reproducibility — dozens of independent
populations over seven decades, multiple assay modalities, extensive replication — and says nothing about reading access. Under
the lead's two-step test nothing needed striking and the band stands. This review is the model the boilerplate rationales in the
Sol 6.1 batches should have followed.
(ii) Momentum "Mechanistic only": the artifact states that the threshold-crossing causal work was done decades ago and that
current output refines mechanism and the never-smoker remainder rather than re-testing causation. Correct, and it does not treat
a completed publication as work underway.
(iii) Scope-matched aggregation: SC1 is the only subclaim whose scope matches the population-level central claim and it sits at
Θ; SC2 (Η) and SC3 (Ζ) are narrower and reported separately. Applied exactly as the rule is written, and the two-rung spread to
weakest-link Ζ is disclosed with its reason rather than hidden.
(iv) Collective authors: no defect of either known form. Reference 17's first author "Secretan" is a personal surname (Béatrice
Secretan, IARC), not a truncated organisation name; no leading-comma instance and no "Group", "Network" or "Consortium"
fragment standing in for a consortium.
(v) No rendering defects: no glued numerals, no empty DOI field, no "PMIDnot supplied" string.
(vi) No spliced "The earlier adversarial approval refers to its original edition" sentence. The consensus supplement's closing
line is correctly hedged — "any earlier approval" — and this document carried no earlier adversarial approval, so the sentence is
true as written.
(vii) The REVIEWER NOTES do not contradict the consensus-consistency supplement: the notes describe the 2026-09-30 rebuild and
the supplement describes the separate 2026-10-09 consistency check, and `consensus-addendum.md` matches the embedded block.
(viii) Dated facts: the never-smoker fraction, the IARC Group 1 classification and the EGFR-mutant subset are all correct as of
the stated cutoff.
(ix) Levels: SC1 Θ, SC2 Η, SC3 Ζ and the headline Θ follow from the quoted caps, thresholds and aggregation rule. The blocks on Ω
(necessity and sufficiency both fail) and on Η (the rubric's calibration note) are correctly argued. No letter required a change.

**Scorecard change.** `independence_basis` (one line's sources plus two explanatory notes) and `study_id` on the 18
`citation_ledger` and `evidence_inventory` rows. The level, confidence, subclaim levels and every other field are unchanged.
Validator re-run: PASS, 0 fail, 0 warn.

**Residual limitations.** Access depth across the ledger is mixed, with several rows at abstract or metadata depth; the review
states this. The exact smoking-attributable fraction by histology and population remains unresolved, as the knowledge-gaps
section records, and this review does not resolve it.

**Approval:** Following adversarial review and the corrections listed above, the revised document is approved. - Claude Opus 5.5 (Anthropic), 2026-10-14.