What each mechanism is worth

How the arms were built

Every arm seats a search with exactly one change in slot 0 and the unchanged depth-2 control in slot 1, with an ETA builder and a fast builder in the other seats. Games run in deterministic mode, so both searches and both builders face the same board, the same k-th roll, the same k-th theft, and the same deck order on every seed; each seed is played once per rotation of the four slots. The primary outcome is the per-seed difference in wins between slots 0 and 1 with a normal-approximation interval over 64 seeds. Depth 2 is the control because it is representative and four times cheaper than the confirmed depth-3 candidate; a change that helps at depth 2 is a candidate for the deep configuration, not a proof.

The arms

Each row is 256 games. Removing the plan term did not hurt at depth 2, which was unexpected: the card, production, and hand terms already carry most of the acquisition-time signal once the search plans a whole turn, so the term that rescued the first leaf is now nearly redundant. Six arms whose intervals brushed zero (event knowledge, trade margin, expansion terms, risk slider, sampled worlds, and the plan term) are registered again on 192 fresh seeds; their results are added below when they complete.

Confirmation on fresh seeds

The six borderline arms were registered again on 192 fresh seeds (64 to 255), 768 games each, with everything else unchanged.

One mechanism survives: sampling hidden hands from the event history instead of public counts alone is worth about 0.07 wins per game, and its interval excludes zero. The expansion terms, the number of sampled worlds, and the trade margin do nothing measurable at this scale, and the plan term that rescued the first leaf is redundant once turns are planned. The risk slider leans harmful but is not resolved by 768 games.

Four styles at one table

Four depth-2 searches that differ only in their inclination presets (builder, raider, gambler, trader) split a table almost evenly: 66, 59, 69, and 62 wins of 256. The sliders change what a seat builds and how it trades without changing how often it wins against equals, which is what a style control should do.

What this program cannot test

Truthful signals, bluffs, and bounded pliability need messages and a recipient that reads them; the fixed builders ignore conversation. Reputation has one mechanical proxy, a willingness model that learns each partner's acceptance rate from the public trade record. Its arm (run 7c17e2c7-a140-4fe5-b055-68bdea33ea60) changed the contrast by -0.031 wins per game (95% interval -0.125 to +0.063): against builders that accept by a fixed rule there is nothing to learn, so the test needs a reciprocal opponent.