The frozen candidate through the protocol arena

The lineup was the frozen depth-3 candidate (1.5 s budget, sixteen scenarios, eight sampled worlds), two ETA builders and one fast builder, rotated through every seat over 20 rounds. Twelve matches ran at once against one server, so the candidate's decisions shared the machine with eleven other searches; its budget is a wall clock, so contention makes it search less, never longer.

The other ETA slot won 13 games and the fast builder 11. Two identical ETA policies finishing 12 games apart is a reminder of how far unpaired 80-game cohorts move; the engine arena pairs boards to take that variance out of development comparisons.

This candidate was frozen before a turn-planning defect was found in the search (its inner turns collapsed to ending the turn, and non-ending root actions were valued without the opponents' round). The result stands as recorded for that exact binary and configuration; the corrected search is a new candidate and will have its own cohorts.

One game, viewed from above

The first scheduled game, selected by schedule rather than outcome.