Paired policy comparisons

Paired mean utility differences · U per 100 policy-role interactions
ComparisonPairsΔ̂95% CI
E20261015-A
R3 − Original
60,000+5.5004[-1.7253, +12.7261]
E20261015-B
R3 − R2
60,000+2.6058[-0.0998, +5.3115]
R3 − OriginalR3 − R20510
Points: paired mean differences. Lines: reported 95% confidence intervals. Dashed line: zero.

Both intervals contain zero; improvement is not established. The comparisons share scenarios and are not independent replications. Intervals are pointwise, without adjustment for multiple comparisons.

Download paired comparison CSV

Design and units

The finite six-agent benchmark uses matched random scenarios, mirrored roles, three agents per policy and zero-sum utility. Each result contains 60,000 independent scenario pairs and 120,000 interactions. Both results use the same scenarios and are dependent. Utility is reported in the original fixed reference unit U per 100 policy-role interactions, without numerical rescaling. These remain observations from the original benchmark; no new domain or cross-domain experiment is claimed.

Completion and integrity

PolicyCompletionsDecisions
E20261015-A · Original696478,759
E20261015-A · R32,003470,336
E20261015-B · R22,193456,082
E20261015-B · R32,130455,556

Both policies use the same deterministic passive completion when supported policy mass is absent. Completion counts are divided by observed policy decisions, not by all possible states. Both completed comparisons passed independent full-record checks of unique pairs, mirrored roles, raw totals and zero-sum settlement; nonzero terminal residuals: 0.

Record identity

E20261015-A
dc3ec12aa60f59431b9ba08f4dd7c9205294b025c581e206ded6dc4e3bea0960
E20261015-B
d9da92785373d9a87b98ffb27c24c7e33a532e2950841feb4302b95eafc525d9

SHA-256 digests identify the completed source result records. Policy assets, individual observations and private execution settings are not included in this aggregate release.

Historical action agreement · August 2026

Reference study A15,169 decisions
Exact action65.76%
Action family73.87%
Reference study B11,000 decisions
Exact action51.29%
Action family62.15%

These earlier replays contain 15,169 and 11,000 evaluated decisions. Exact-action and category agreement retain their published precision and original definitions. Agreement is neither a utility estimate nor a success rate. Record-level observations are absent from this extract; additional confidence intervals are not computed. These studies are not pooled with the October comparisons.

Download historical agreement CSV