Records

The selected pool against the structured prior on held-out setups

Registration c75d346c-11fc-4cb1-af21-d993b87172f3 (study setups/selection-by-self-play), run 2026-09-19 on AMD Ryzen AI Max (Strix Halo), 32 threads, 125 GB unified memory, Radeon 8060S via ROCm, Linux 6.18. Server commit 0eb994c, research commit 19c62a6. The run shared the machine with other agents' jobs, which inflates the per-move clocks but not the deterministic scores.

Command:

cd server && target/release/matchup -- --red 'pool=expectimax:{"samples":16,"depth":2,"setup":"pool:../research/content/assets/setup-pool.json"}' --blue 'prior=expectimax:{"samples":16,"depth":2,"setup":"structured"}' --games 400 --seed 0 --rules training --threads 8 --out ../research/runs/setup-selection/cell-pool-v-prior.json --record-dir ../research/runs/setup-selection/records-pool-v-prior --tag c75d346c-11fc-4cb1-af21-d993b87172f3

Both seats play expectimax over sixteen sampled worlds at depth 2; the red side draws its setup uniformly from the selected 32-entry pool, the blue side from the structured prior. The seeds 0 to 199 draw the prior's setups through the arena's streams, disjoint from the selection's procedure seed 20260919 and its opponent space, so the pool meets opposing setups it never saw.

Player	Wins	Draws	Losses	Effective	95% Wilson	Mean plies	ms/move
pool	260	6	134	0.658	0.610 to 0.702	275	5.083
prior	134	6	260	0.343	0.298 to 0.390	275	4.708
ending: flag_captured	323
ending: no_moves	71
ending: quiet_limit	6

Per-seed paired contrast for the pool side (score as Red plus score as Blue minus one): mean 0.315, 95% normal interval 0.223 to 0.407 over 200 seeds (88 positive, 88 split, 24 negative). The pool side won 131 games as Red and 129 as Blue. Draw share 1.5%. No game ended within 20 plies. Wall time 68.6 s on 8 threads; the budget of 2,000 ms per move was met with a mean of 5.1 ms.

Provenance

Archived JSON: runs/setup-selection/cell-pool-v-prior.json (SHA-256 7b90e6a0946857031811f7347ee6c40abe252131a85c85b0a9657da8b7085b7f); per-game records under runs/setup-selection/records-pool-v-prior/. Games planned 400, attempted 400, completed 400, interrupted 0. The curated replay library holds seed 0 from both seats under records/games/setup-pool/. No deviation from the registration; the study's decision rule for this cell, the interval above zero, is met.