Records

The beta-zero ablation against the network alone

Registration 8c2d3e34-5fb9-4892-81a2-00689163a3f9 (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 paper's sign is not reproduced: setting the network KL coefficient beta to zero did not lose to the network. The per-seed contrast is 0.083 with a 95% interval of -0.027 to 0.194, which overlaps zero, so the cell is inconclusive with a positive point estimate, and the beta-zero side lost no game.

Per-seed paired contrast (beta-zero score as Red plus as Blue minus one): mean 0.083, 95% normal interval -0.027 to 0.194 over 12 seeds (2 positive, 10 drawn, 0 negative). Table 17 of the paper has the beta-zero variant 362 Elo below the network; here it is even with the network, which the mechanism explains: the paper's catastrophe is the search overfitting to its own rollouts' quirks without the network anchor, but this checkpoint's value head barely separates the moves (the q spread across legal moves averaged 0.023 over 6,632 search moves, and the calibration measures the head's correlation with realised outcomes at -0.19 with every prediction inside the middle fifth of the outcome scale), so there is no quirk to overfit.

The search's own diagnostics over 6,632 search moves: mean top share of the search distribution 0.218 against 26.7 legal moves (uniform would be 0.037), mean q spread 0.023, 188.6 rollouts per move. With beta at zero the tabular update is e^(q/alpha) rho, so the noise in q is amplified five hundredfold against the magnet policy; the top share 0.218 against cell (a)'s 0.103 shows the amplification, and the score shows it buys nothing decisive either way.

Both contestants are runs/belief-policy (the small preset, seed 0, 30 iterations, belief network 400 steps); the search player uses alpha 0.002, beta 0. Setups are drawn per seed from the setup network with the same two setups across seats. The 251.03 ms per search move and 3.1 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 \
  --blue run:/home/keshav/armies/research/runs/belief-policy \
  --games 24 --seed 0 --rules training --max-seeds 5 --out runs/ue-cell-c.json

Raw output, small enough to sit beside this record: 8c2d3e34-5fb9-4892-81a2-00689163a3f9-match.json (SHA-256 3e67552caed2e86386037d3cca2b059eed50a4070c5aefbdbaaaae5335b17a54, 14 KB, wall 1,528 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 and was written before any game was played.