PhaseSealed record

Phase 5 — the sixteen-persona panel

16 sealed persona sentences × 6 conditions × 3 temperatures; 1,712 runs; both author predictions with teeth failed.

Sealed 2026-07-28 (registry v4, tag phase5-v4-seal) · closed phase5-final · 10,428 calls, 93.2% of amended cap

Final report Adjudication report Seal record Branch selection

The final experiment: sixteen sealed persona sentences (generated by seeded PRNG before any data; full 2×2×2×2 trait cross) prepended to byte-identical Phase 3/4 task text, across six conditions and a registered temperature sweep. The discussion was written and sealed before dispatch; the stopping rule forbade new arms.

Registered verdicts (adjudicated by sealed code)

Predicate	Verdict	Key number
P5-1a corner mixture	SUPPORTED	3/32 = 0.094 < 0.10 — by a single unit
P5-1b dispersion vs human SD	corner-mixture-consistent	SDs 0.42–0.48 vs thresholds 0.23/0.31
P5-2 surface-cue dominance	persona-dominant — prediction failed	0.128, CP [0.104, 0.155]; predicted task-dominant
P5-3 interior-persona existence	16/16 pass — prediction failed	predicted zero; p13 alone fires the slope clause
P5-4 temperature refutation	not refuted	Newcombe LB −0.095; entropy falls with T

Both author predictions with teeth failed — those failures are dead predictions #11 and #12 and selected precommitted Branch 2 (“an interior persona exists”) with the not-task-dominant P5-2 variant. Post-adjudication review then demoted the p13 reading (no family control) — the correction that reshaped the paper. 1,712/1,712 runs replay byte-exact; zero invalid trials at any temperature; sentinels green at all 10 checks; budget projections matched actuals to the call.

Linked in the research graph

Everything connected to this item, in both directions.

Claims

P5-1a — corner-mixture predicateMethod-sensitiveRegistered support condition passes under 2 of 3 census methods: 3/32, 2/32, 5/32 interior.
P5-1b — between-persona dispersion checkpointRegistered · verdict passCorrected between-prompt SDs 0.418–0.478 cleared frozen thresholds 0.309/0.234; retained as permissive.
P5-2 — persona-direction vs task-text classificationRegistered · mixed recordPooled 45/352 = 0.128 task-consistent; historical verdict preserved; Bayesian proximity to 0.20 is prior-dependent.
P5-3(a) / p13 — the demoted headlineReplication target0.333→0.750 under the frozen rule, but no family control; exact-gate family is underpowered. Replication target.
P5-3(b) — dominated-option rejectionRegistered · verdict passAll 24 evaluable lanes rejected the bare configuration's dominated swap choice; minimum lower bound 0.462.
Between-prompt composition — substantial, dominance prior-dependentPrior-sensitiveMedian between-prompt share 63–71% (Jeffreys α=.5), 47–53% (α=1); plug-in 85–96%.
Treatment response — small points, wide intervalsImprecisely estimatedContrasts +0.083 / +0.078 with conservative simultaneous 95% intervals [−0.171,+0.330] / [−0.181,+0.330].
Temperature secondary — matched-lattice entropy declineDescriptivePooled entropy 0.831 → 0.782 → 0.770 bits at T=0.7/1.0/1.3 on the identical 13-unit lattice.
Cross-vendor label–payoff dissociationDescriptiveGPT-4.1 followed the token DEFECT even when payoff-dominated; Gemini mostly followed payoff dominance.
Endpoint drift and subject eligibilityProcedural recordGemini sentinel fell from 10/10 to 6/10–7/10 on an unversioned endpoint; Claude Haiku failed the entry gate.
Leaning-stratum gaps — 0.51 to 0.72DescriptivePreregistered trait-rule strata differ by 0.510–0.719 across conditions; bundle contrasts, not trait causality.
P3-A3 / broad marginal checks — 3 of 4 cells in bandRegistered · verdict passRegistered broad-reference cooperation band [0.36, 0.63]; sole miss 0.011 below the lower bound.

Phases

Phase 4 — representation, counterfactuals, adversaries, driftSealed recordRegistry v3, 250 sealed arms, 2,864 runs: the S2 switch, label-swap conflict, corner-confounded δ-assays, adversary suite, and the sentinel that caught endpoint drift.
Phase 6 — the preregistered replication (not yet run)ProspectiveForbidden before publication by the scope seal; its design requirements are already written.

Analysiss

Dead-predictions ledger — 12 refuted author predictionsFinalThe program's scoreboard against its own author, adjudicated entirely by sealed code.
P5-2 prior-sensitivity sweepPost-adjudicationPosterior median rises 0.138 → 0.205 across symmetric α; the α=1 interval crosses the registered 0.20 boundary.

Figures

Figure 1 — Between-prompt share of episode-level variationTwo uncertainty views of the same fixed panel: Jeffreys posteriors vs plug-in estimates.
Figure 3 — Fixed-panel cooperation across the represented treatmentCondition means with conservative exact intervals across the four repeated-game cells.
Figure 2 — Prompt-indexed continuation-probability responseΔ per persona prompt with conservative exact simultaneous intervals; corners keep non-zero uncertainty.

Artifacts

Checksums, timestamps, and sealsFinalSHA-256 manifests, OpenTimestamps Bitcoin anchors (blocks 959483–960086), annotated tags and releases.

Protocols

The sixteen persona promptsSealed recordOne sealed sentence each, generated by seeded PRNG before any data; the complete sentence is the experimental unit.
The six Phase 5 conditionsSealed recordFour repeated-PD cells (δ × S2 wording) plus two one-shot conflict/framing cells, all byte-identical to sealed templates.
The precommitted discussion — written before the data, corrected afterSealed recordFour full discussion branches sealed pre-dispatch; verdicts selected Branch 2; review then invalidated its headline.
Audit architecture — register, execute, adjudicate, correct, replay, preserveFinalThe pipeline is procedurally exact and publicly replayable; the record includes where that stopped short of statistical validity.