Records

The prior's exploiter against the prior: the exploitability proxy

Registration 1cea5672-b0f9-48eb-83dc-72f7525e67e2 (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 'prior-exploiter=expectimax:{"samples":16,"depth":2,"setup":"33746364BF/638B72549B/72B222B422/2X55385B16"}' --blue 'prior=expectimax:{"samples":16,"depth":2,"setup":"structured"}' --games 400 --seed 0 --rules training --threads 8 --out ../research/runs/setup-selection/cell-exploiter-v-prior.json --record-dir ../research/runs/setup-selection/records-exploiter-v-prior --tag 1cea5672-b0f9-48eb-83dc-72f7525e67e2

Both seats play expectimax over sixteen sampled worlds at depth 2. The red side always plays the fixed arrangement the exploit procedure found against the structured prior (procedure seed 20260921, 16,368 games); the blue side draws from the structured prior. The evaluation seeds 0 to 199 are disjoint from every procedure seed.

Player	Wins	Draws	Losses	Effective	95% Wilson	Mean plies	ms/move
prior-exploiter	288	7	105	0.729	0.683 to 0.770	281	4.507
prior	105	7	288	0.271	0.230 to 0.317	281	4.017
ending: flag_captured	312
ending: no_moves	81
ending: quiet_limit	7

Per-seed paired contrast for the exploiter side: mean 0.458, 95% normal interval 0.364 to 0.551 over 200 seeds (114 positive, 63 split, 23 negative). The exploiter won 144 games as Red and 144 as Blue. Draw share 1.8%. One game ended within 20 plies and the prior side lost it. Wall time 61.5 s on 8 threads; the budget of 2,000 ms per move was met with a mean of 4.5 ms.

This is the prior's exploitability proxy. The pool's proxy reads 0.150 (0.043 to 0.257) (record): the pool's exploiter scores well below the prior's, so the study's rule, that the pool's proxy is no higher than the prior's by more than its interval, is met with room to spare.

Provenance

Archived JSON: runs/setup-selection/cell-exploiter-v-prior.json (SHA-256 cc2db62130a9d625a0d2d83f5ce372edc5773bf5d326c924f9a986852fdac83d); per-game records under runs/setup-selection/records-exploiter-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.