Records
KLUSS at the 1 s budget against expectimax over sixteen worlds
Registration 05d831b0-d37b-4f03-8930-49fc8541340a (study subgame-solving/kluss-vs-expectimax), run 2026-09-18 on AMD Ryzen AI Max (Strix Halo), 32 threads, 125 GB unified memory, Radeon 8060S via ROCm, Linux 6.18. Server commit 3dcb590, research commit 2b0100a (registration files as committed; this record added afterwards). Other agents shared the machine during the run; per-seed results are deterministic regardless of load.
Command:
cd server && cargo run --release -p armies-arena --bin matchup -- --red 'kluss 1s=kluss:{"samples":64,"shadows":4,"expansions":10000,"cfr_every":64}' --blue 'expectimax:{"samples":16,"depth":2}' --games 200 --seed 0 --rules training --threads 8 --record-dir ../research/runs/05d831b0-d37b-4f03-8930-49fc8541340a/records --out ../research/runs/05d831b0-d37b-4f03-8930-49fc8541340a/matchup.json --tag 05d831b0-d37b-4f03-8930-49fc8541340a
cd research && python3 scripts/paired_contrast.py runs/05d831b0-d37b-4f03-8930-49fc8541340a/matchup.json| Player | Wins | Draws | Losses | Effective | 95% Wilson | Mean plies | ms/move |
|---|---|---|---|---|---|---|---|
| kluss (10,000 expansions) | 67 | 0 | 133 | 0.335 | 0.273 to 0.403 | 181 | 1591.977 |
| expectimax (16 worlds, depth 2) | 133 | 0 | 67 | 0.665 | 0.597 to 0.727 | 181 | 6.036 |
| ending: flag_captured | 191 | ||||||
| ending: no_moves | 9 |
Per-seed paired contrast for kluss (score as Red plus score as Blue minus one): mean -0.330, 95% normal interval -0.467 to -0.193 over 100 seeds (13 positive, 41 split, 46 negative). Kluss won 36 games as Red and 31 as Blue. Wall time 3,647 s on 8 threads. Time per move is as played in the arena under load; the calibration bench (single thread, middlegame positions, other jobs sharing the machine) read 822 ms at 10,000 expansions, 1,024 ms on a quieter repeat.
Decision rule
The hypothesis (kluss beats expectimax at this budget) is not supported: the contrast interval -0.467 to -0.193 lies entirely below zero. Doubling the budget from the 200 ms cell raised the effective score from 0.268 to 0.335 without closing the gap.
Provenance
Archived JSON: runs/05d831b0-d37b-4f03-8930-49fc8541340a/matchup.json (SHA-256 e1c3c5f5c660b35fab131a38e09ae59f786ec916bd9af64bc6bcaa93ba1eacef), with per-game moves in the arena record files under the same directory. Games planned 200, attempted 200, completed 200, interrupted 0. The seed-0 records of both seats are in the replay library under records/games/kluss-vs-expectimax/. The kluss contestant is labelled kluss 1s in the records so the two cells of the study have distinct ids. No deviation from the registration.