Records

Update-equivalence search at 200 rollouts and depth 10 against the network alone

Registration 77aebde0-74c7-4a7a-9835-9c1f853b159e (study ataraxos/update-equivalence), run 2026-09-19 on AMD Ryzen AI Max (Strix Halo), 32 threads shared with three other delegated agents, Radeon 8060S via ROCm. Research commit bd90f70; the armies module is the shared venv build of server main (0.1.0), which this study did not change. The decision rule (95% interval above zero) is not met: the per-seed contrast is 0.208 with a 95% interval of -0.016 to 0.433, which overlaps zero by a hair, so the cell is inconclusive with a positive point estimate, not evidence of no difference.

Per-seed paired contrast (search score as Red plus as Blue minus one): mean 0.208, 95% normal interval -0.016 to 0.433 over 12 seeds (3 positive, 9 drawn, 0 negative). The search side did not lose a game; it won all five decisive games, and every other game drew. Most of the draw share is the checkpoint itself, which drew 79.5% of its null-pair games (the null pair's record), so most seeds have no contrast to contribute.

The search's own diagnostics say why the effect is small: over 6,092 search moves the mean top share of the search distribution was 0.103 against 22.4 legal moves on average (a uniform policy would put 0.045 there, the raw network about the same), the mean spread of the q values across legal moves was 0.019, and each move ran 191 rollouts on average. The value estimates the rollouts average barely separate the moves, so the tabular update is a mildly sharpened copy of the network's own near-uniform policy.

Both contestants are runs/belief-policy (the small preset, seed 0, 30 iterations, belief network 400 steps, from the learned-belief study); the search player samples its worlds from the belief network and plays the magnetic-mirror-descent move with alpha 0.002 and beta 0.02. Setups are drawn per seed from the setup network with the same two setups across seats. The 339.7 ms per search move and the 2.99 ms per plain move were measured while other agents shared the machine, so they are upper bounds, not alone-on-the-machine numbers.

Provenance

Command, from the research worktree root with PYTHONPATH=training, played in three lock chunks of 5, 5 and 2 seeds:

.venv/bin/python -m armies_train match \
  --red run:/home/keshav/armies/research/runs/belief-policy:search=200,depth=10,alpha=0.002,beta=0.02 \
  --blue run:/home/keshav/armies/research/runs/belief-policy \
  --games 24 --seed 0 --rules training --max-seeds 5 --out runs/ue-cell-a.json

Raw output, small enough to sit beside this record: 77aebde0-74c7-4a7a-9835-9c1f853b159e-match.json (SHA-256 c1228631dc79e14e37564d28f167070d0be271ec5212498854ebe805838fa74a, 14 KB, wall 1,921 s across the three chunks). Games planned 24, attempted 24, completed 24, interrupted 0. The count is 24 games on seeds 0 to 11, not the brief's 200 on seeds 0 to 99; the registration records the cost reason (320 to 650 s per paired seed on this machine) and was written before any game was played.