Ledger
Claims and their evidentiary tiers
Every registered predicate, its frozen verdict, and its current scientific status — kept deliberately distinct.
The program's rule: historical mechanical verdicts stay visible verbatim; later analyses are additive and explicitly labeled. Twelve registered author predictions were refuted and published (the dead-predictions ledger); the p13 demotion is an inferential correction, not a refutation, and is deliberately not counted among them.
Tier vocabulary
| Tier | Meaning |
|---|---|
| Registered · verdict pass/fail | frozen mechanical verdicts, adjudicated by sealed code before interpretation; historical labels never rewritten |
| Method-/prior-sensitive | classification changes across defensible interval constructions or symmetric priors; continuous estimates primary |
| Replication target | neither prospectively confirmed nor decisively disconfirmed; next test must be prospectively powered |
| Descriptive | reported without inferential weight (cross-vendor tier, non-registered comparisons) |
| Post-adjudication | computed after the seal in response to review; cannot create prospective confirmation |
| Procedural | monitoring and process records, not behavioral findings |
All claims
| Claim | Status | One-line result |
|---|---|---|
| P3-A3 / broad marginal checks — 3 of 4 cells in band | Registered · verdict pass | Registered broad-reference cooperation band [0.36, 0.63]; sole miss 0.011 below the lower bound. |
| Between-prompt composition — substantial, dominance prior-dependent | Prior-sensitive | Median between-prompt share 63–71% (Jeffreys α=.5), 47–53% (α=1); plug-in 85–96%. |
| Treatment response — small points, wide intervals | Imprecisely estimated | Contrasts +0.083 / +0.078 with conservative simultaneous 95% intervals [−0.171,+0.330] / [−0.181,+0.330]. |
| P5-1a — corner-mixture predicate | Method-sensitive | Registered support condition passes under 2 of 3 census methods: 3/32, 2/32, 5/32 interior. |
| P5-1b — between-persona dispersion checkpoint | Registered · verdict pass | Corrected between-prompt SDs 0.418–0.478 cleared frozen thresholds 0.309/0.234; retained as permissive. |
| P5-2 — persona-direction vs task-text classification | Registered · mixed record | Pooled 45/352 = 0.128 task-consistent; historical verdict preserved; Bayesian proximity to 0.20 is prior-dependent. |
| P5-3(a) / p13 — the demoted headline | Replication target | 0.333→0.750 under the frozen rule, but no family control; exact-gate family is underpowered. Replication target. |
| P5-3(b) — dominated-option rejection | Registered · verdict pass | All 24 evaluable lanes rejected the bare configuration's dominated swap choice; minimum lower bound 0.462. |
| X2 / S2 switch — one sentence operation, 0/40 → 37/40 | Registered · verdict pass | A single wording-and-position operation on the continuation sentence moved bare cooperation across the range. |
| D1 — ordinary one-shot wording main effect (null) | Registered · not supported | +0.0063 (SE 0.0210; Holm-adjusted p=1.00) across the 640-episode one-shot battery. |
| Label-swap conflict — semantic label overrides payoff dominance | Registered · verdict pass | Bare GPT-4.1 chose the cooperation-worded option 0/40, taking the payoff-dominated role when it carried “Defect.” |
| Leaning-stratum gaps — 0.51 to 0.72 | Descriptive | Preregistered trait-rule strata differ by 0.510–0.719 across conditions; bundle contrasts, not trait causality. |
| RPS adversary suite — exploitability is opponent-contingent | Registered · mixed record | n-gram2 earned +0.215/round vs GPT-4.1 (Holm-surviving); the WSLS-targeter designed from its signature earned +0.008. |
| RPS role-attached asymmetry — registered direction reversed | Registered · not supported | First-minus-rock contrast −0.181 (prediction was positive); Gemini mirror +0.243. Vendor-specific asymmetry. |
| Cross-vendor label–payoff dissociation | Descriptive | GPT-4.1 followed the token DEFECT even when payoff-dominated; Gemini mostly followed payoff dominance. |
| Temperature secondary — matched-lattice entropy decline | Descriptive | Pooled entropy 0.831 → 0.782 → 0.770 bits at T=0.7/1.0/1.3 on the identical 13-unit lattice. |
| Endpoint drift and subject eligibility | Procedural record | Gemini sentinel fell from 10/10 to 6/10–7/10 on an unversioned endpoint; Claude Haiku failed the entry gate. |
Dependency spine
The paper's argument rests on four load-bearing links: the marginal checks license nothing about response beyond band arithmetic (Proposition A); the composition claim depends on which uncertainty view conditions on boundary concentration; the p13 story rests on the family audits and the attainability floor; and the P5-2 classification is carried by the word/payoff-confounded swap cell. The archived claim-dependency audit maps every quotation of these results to its correction rule.