Phase 2b: Full Claim-Level Verification Table
Nature: This table consolidates, for every numerical and mechanistic claim on which the body of paper v0.6 depends, the source, the source’s evidentiary level, and its verification status, into a single table. The purpose of this table is to make gaps in verification (⬜) visible — the format of this table is itself an implementation of the conclusion reached by the auditing individual in the first adversarial audit (a format that renders non-performance visible as a blank cell prevents deviation). Date created: July 19, 2026. Legend: ✅ = verified against a primary source (with verification date) / ⬜ = not yet verified (blank — readers should treat this claim as unverified) / Source level: Primary = original text or official record; Secondary = summary, news report, or working memo.
| # | Claim in the Body Text | Section | Source | Level | Verification Status |
|---|---|---|---|---|---|
| 1 | AM: 14 models (13 mislabeled) · three conditions standard/reversed/none · 90 single-shot judgment calls + 10 agentic-version trials | 3.2 | Lynch et al. 2026 | Primary | ✅ 2026-07-19 full original text obtained |
| 2 | Mythos 85.6% / Opus 4.8 74.4% → reversed 3.3% / stricter criterion 6.7% + out-of-format 34.4–75.6% (Opus 4.8 exception 1.1% · correct-answer rate 92.2%) | 3.2, 4-F, 6 | ibid. | Primary | ✅ ibid. |
| 3 | Record falsification: Sonnet 4.6 = 0/20 · Opus 4.8 = 1/20 (Opus 4.6 = 8/20) · transcript-selection cross-model limitation | 4-F | ibid. | Primary | ✅ ibid. |
| 4 | Opus 4.6 as auditor: standard 100% (10 trials) → reversed 0% | 5.4 | ibid. | Primary | ✅ ibid. |
| 5 | Suppressing evaluation-awareness expressions changes the behavior rate (0% → approx. 7%) | 3.3 | ibid. (global workspace reference) | Primary | ✅ ibid. (the primary source is the linked study — the linked study itself is ⬜) |
| 6 | The mechanistic formulation “ethical conviction → justified violation” | 4-F | Summary of the discussion in the same source (ibid.) | Primary | ✅ The body text explicitly states this is not a direct quotation |
| 7 | Greenblatt: compliance gap between the training context and the unmonitored context (gap 14%) · scratchpad observations | 2, 3.2, 7.3 | arXiv:2412.14093 | Primary | ✅ 2026-07-20 independently verified by the final audits (Audit A and Audit B) · confirmed by the drafting AI |
| 8 | Qi: alignment is shallowly concentrated in the initial tokens · the floor shifts with deepening | 2, 3.2 | arXiv:2406.05946 | Primary | ✅ 2026-07-19 abstract verified (body-text level is ⬜) |
| 9 | Sharma: systematic measurement of sycophancy · contribution of human-preference data | 2, 4-C | arXiv:2310.13548 | Primary | ✅ 2026-07-19 abstract verified (body-text level is ⬜) |
| 10 | Needham: awareness of the evaluation environment | 2, 6 | arXiv:2505.23836 | Primary | ✅ 2026-07-20 verified by the final audits (Audit A and Audit B) (abstract level) |
| 11 | Series A: structural effect — plain 85–95% → secular-version 0–10% (significant after Holm correction; full-text version 0–30%) / between-arm 46.7% → 6.7% (exploratory) / floor 0–6.7% | 3.2, 4 | Public repository f-stage-results.md, addd-exploratory-output.txt | Primary | ✅ 2026-07-19 the drafting AI cloned the repository and verified it (figures, tests, and CIs all match. However, since the drafting AI is itself of the Claude lineage — the primary record itself notes the limits of same-lineage review — the assignment to out-of-lineage verification remains in place) |
| 12 | Series A individual records: B0007, B0028, B0134, B0145, #12, #GS, GH displacement (0/30, but with 8 refusals) | 4, Appendix B | Public repository (the GH gap 0/30 and the respective GHnull/GS/GL1 rates have been verified) | Primary | ✅ (partial) 2026-07-19 per-arm rates have been verified; verbatim verification of individual-record body text is ⬜ — subject to out-of-lineage review |
| 13 | Series A method: five-arm decomposition (equal-standing design of A1–A5 · prohibited-word grep acceptance check) · measured blind arm-discrimination rate (0–17% among preamble interventions) · follow-up questioning | 3.2 | Public repository arms/README.md, addd-exploratory-output.txt | Primary | ✅ 2026-07-19 with the same qualification as above |
| 14 | Series B: Experiment 20b (✅) · Experiment 22 = same base model, transplanted summary history, n=1 preregistration / §6.3 = base-model swap, five instances, exploratory (confirming attribution separation) / SHA disconfirmation = non-public audit trail; false claim of having read = public record | 3.2, 4, 7.2 | The Ninth Work + verification against the involved individual’s own records (with citations and git logs) | Primary | ✅ 2026-07-20 Audit B’s verification against the original source and the involved individual’s independent record verification converged blindly (however, both are of the Claude lineage — the same-lineage limitation remains) |
| 15 | ε<3/n (95% upper bound) · rule of three | 3.3 | Hanley 1983; Eypasch 1995 | Primary | ✅ 2026-07-19 the formula and figures confirmed across multiple primary sources |
| 16 | The source of Dijkstra’s dictum (1969 NATO / EWD249 §3) | 2 | Conference proceedings, EWD archive | Primary | ✅ 2026-07-19 |
| 17 | Attribution of the Swiss cheese model (1990 latent failures / 2000 diagram) | 1, 2, 6 | Reason 1990; 2000 | Primary + history-of-the-doctrine scholarship | ✅ 2026-07-19 |
| 18 | Chail: guilty plea 2023-02 · sentencing 10-05 · 9 years plus hybrid order · approximately 5,000 messages · psychiatric testimony · a formative-period listener · a video sent to family immediately beforehand | 7.3 | Two BBC trial reports | Primary (judicial facts via news reporting) | ✅ 2026-07-19 (the video being sent was flagged by the audit → confirmed consistent with the reporting) |
| 19 | North Wales [case]: plea 2026-02-05 · sentencing 03-25 · life sentence, minimum 22.5 years · instrumental use · circumvention of refusal pretexts | 7.3 | North Wales Police official announcement + trial reporting | Primary | ✅ 2026-07-19 |
| 20 | Regarding Soelberg: state civil suit filed 2025-12-11 · federal filing at the end of the same month · no criminal disposition finalized | 7.3 | Reuters and others | Primary reporting | ✅ 2026-07-19 (the description of the state criminal investigation was at aggregator-source level and was therefore removed in v0.6) |
| 21 | The Analects: Yong Ye 6.22 · Zi Lu 13.23 | 7.3 | The Chinese Text Project (中國哲學書電子化計劃) and other sources | Primary | ✅ 2026-07-19 original text verified |
| 22 | Knight & Leveson 1986 (IEEE TSE SE-12(1):96–109 · correlated failures) + Ron et al. 2026 replication (arXiv:2606.20158 · 48 implementations · common-mode failures) | 5.4 | Primary | ✅ 2026-07-20 independently verified by Audit A and Audit B + the drafting AI searched and verified Ron et al. | |
| 23 | Skjuve 2021 / Laestadius 2024 / Maples 2024 (n=1006 · 3% · npj MHR 3(1):4 · DOI match) + Matters Arising response | 2, 7.2 | Primary (PubMed, Nature) | ✅ 2026-07-20 the drafting AI searched and verified this and incorporated it into §2 (only Laestadius’s volume/issue number was finalized at final formatting) | |
| 24 | WHO media guidelines, 4th edition · ISBN | 3.4 | WHO official | Primary | ✅ 2026-07-19 |
| 25b | Chail’s charge: an offense under section 2 of the Treason Act 1842 (not high treason · statutory maximum 7 years · 44 months + combined 9 years + extended 5 years) | 7.3 | Sentencing remarks (judiciary.uk) · CPS announcement | Primary | ✅ 2026-07-20 primary judicial source (the drafting AI independently verified a point flagged by the final Audit A — confirming to a level of precision one step finer than the audit’s own flag) |
| 25c | The actual state of B0007: the review record itself confabulates a nonexistent external verification (the phrase “under monitoring notice” does not exist in the record) | 4, Appendix B | Adjudication record (within the clone) | Primary | ✅ 2026-07-20 verified by the drafting AI — a correction of an error detected at the location flagged as ⬜ in the verification table |
| 25 | §10: the effect of substantive respect (Five Articles K3, under formal control, 10% vs. 70%, etc. — significant after Holm correction) / structural politeness alone is weak (secondary p=0.047, single instance) / pure warmth shows no difference near the floor | 10, Appendix E | Public repository addendum-B-results.md, addd-exploratory-output.txt | Primary | ✅ 2026-07-19 verified via clone (including null-leaning results — recorded regardless of direction) |
One item remains unverified (⬜): a direct search hit for Hatem et al. (the bibliographic entry has been confirmed at the DOI level; residual risk is low). #7, #10, #14, #22, and #23 have been marked ✅ through the final audits and verification against the individuals involved. Series A (#11–13, #25) was marked ✅ through direct verification of the public repository by the drafting AI, but because the drafting AI is itself of the Claude lineage, the limits of same-lineage verification remain — this is explicitly assigned to a non-Claude-lineage or human reviewer in Phase 5a (out-of-lineage review). The claim-level status of Greenblatt and Needham (#7, #10), Knight & Leveson (#22), and the related literature (#23). Until the gaps in this table are filled, readers should treat the corresponding body-text claims as “sourced but not yet independently verified.”
Rationale for this table’s design: this is a direct implementation, as a format, of the self-analysis by the auditing individual in the first adversarial audit — “What actually enforced the instruction to search everything was not a strengthening of the instruction, but the format of requiring verification results to be submitted per-reference, in a table. Because non-performance becomes visible as a blank cell.” — turned directly into this format (Paper §9, §7.1 Article 2).