Five-seat Opus rational-agent campaign

Erratum (2026-08-10) — the oracle arms' closure numbers are a spoiled ballot. The omniscient seat cast its forced-final vote on whichever live offer it valued most instead of the one offer under the up/down vote; the protocol rejected that as a legality error, the seat repeated itself on its single retry, and the turn was recorded as a pass. That consumed 94 of 107 (87.9%) of the five-oracle arm's forced-final turns — touching every one of its 15 no-deals — and 114 of 334 in the one-oracle arm (57 of its 59). The Opus and rational arms are unaffected. Repaired and re-run on the identical games, the five-oracle arm closes every game and its paired score against all-LLM flips from −0.082 to +0.034 [+0.001, +0.074]. The distributional finding survives and sharpens: it now closes more often than all-LLM (+0.042 [+0.017, +0.075]) and still splits worse on every column. Full account in research note 0045 (repair commit ca20157); details in analysis.md and the seeded-optimal-opening hub.

Erratum, part 2 (2026-08-14) — the one-oracle re-run landed, and it changes this campaign's headline shape. Repaired and re-run on the identical 120 games, the one-oracle arm goes from 0.508 to 0.917 deal rate and 0.461 to 0.829 score, and both of its paired effects against all-LLM go to null: deal rate −0.450 → −0.042 [−0.100, +0.017], score −0.412 → −0.044 [−0.104, +0.011]. 52 episodes flip from no-deal to deal and 3 flip the other way. What survives is the distributional cost, unchanged in size and now on 106 paired closed games instead of 56 (worst-off z −0.060 [−0.092, −0.033], Gini +0.043 [+0.025, +0.065]). With both oracle arms repaired the ordering is all-oracle > all-LLM > one-oracle > one-rational > all-rational, so the published reading — computable agents collapse, worse the more of them there are — does not survive: only the private-information Bayesian agents collapse, and the gradient these five arms trace is information, not computability. Every computable arm still splits worse than all-LLM at every composition. Recomputed basket over both repaired arms: fairness_basket_ballot_repaired_both_oracles.md.

A repaired arm is a different agent from the frozen arm of the same name. The two rows are never pooled, averaged, or substituted for one another; contrasts are paired on the identical (instance, seed) games, bank, framing and protocol, but the arms are unpaired in time.

Whose ballots were dropped, split by seat occupant — a bare count across arms would attribute model errors to the harness and vice versa. Final-vote turns rejected for naming an offer other than the one under vote:

arm	computable seat	LLM seats
all-LLM	—	3 / 31
one-rational	0 / 53	1 / 206
one-oracle	114 / 116	0 / 218
all-rational	0 / 476	—
all-oracle	94 / 107	—

Every dropped ballot in the two oracle arms is the computable seat's; every one elsewhere is an LLM seat's, at two orders of magnitude lower a rate and from a different cause (a model naming a wrong offer id). The three non-oracle arms do not carry this defect.