The learned leaf through the protocol arena

What was registered

The registered protocol was registered under patterns/ntuple-leaf before the candidate was frozen: the frozen tables at artifacts/ntuple/hex-portfolio-main.bin as the leaf of the baseline's exact search configuration (depth 3, 1.5 s budget, 8 scenarios, 4 sampled worlds) in slot 0, expectimax-v2-plan in slot 1, eta and fast in the other seats, on seeds 320 to 339 in deterministic mode, 80 games. The rule: all games complete validly with no server timeout moves; support only if the 95% interval of the paired per-seed win difference between slots 0 and 1 lies above zero.

80-game protocol cohort played all 80 with twelve matches in flight. The two searches ran as separate remote processes against the server and saw only their own seated view and event history; the tables were loaded from the file named in the policy registry, whose SHA-256 (f8e61e66…) is recorded with the candidate's receipt. Games averaged 73 turns.

The result agrees with the engine-arena screens: at depth 3 against depth 3 on 64 fresh engine seeds the contrast was +0.242, and here through the protocol on 20 seeds it is +0.263.

One game, viewed from above

The first match of the cohort, seed 320 in its first rotation, which the tables' seat won in 65 turns. It is a real recorded game chosen for the replay; it does not by itself establish which policy is stronger.

Limitations

One lineup, one seat rotation scheme, and twenty seeds. The interval is wide because the cohort is small; the engine screens on more seeds carry the sharper estimate. Both searches negotiate only with the fixed builders' willingness model, so nothing here speaks to trade-rich tables with negotiating opponents or to play against humans. The tables decide more slowly than the hand-written leaf under the same budget, which the 1.5 s cap absorbs at depth 3 but which would matter under a tighter deadline.

Run record