The target is met in this evaluation.
Fresh held-out games exceeded 95% wins while averaging less than 100 ms per complete decision.
The objective counts gold, silver and bronze equally. Timing is the average full agent decision, including public features, model inference, search, and explanations; network and browser rendering are excluded.
| Strategy / set | Wins / games | Win rate / 95% CI | Gold / silver / bronze | Mean / p95 ms | Decisions | Search settings |
|---|---|---|---|---|---|---|
| linear8dev · strategy95-screen | 46 / 50 | 92.0%81.2–96.8% | 3 / 30 / 13 | 91.84 / 95.28 | 2,973 | ≤8 actions · 92 msPUCT 0.5 · cap 262,144 |
| linear16dev · strategy95-screen | 46 / 50 | 92.0%81.2–96.8% | 4 / 26 / 16 | 92.11 / 96.29 | 2,995 | ≤16 actions · 92 msPUCT 0.5 · cap 262,144 |
| explore16dev · strategy95-screen | 48 / 50 | 96.0%86.5–98.9% | 3 / 29 / 16 | 92.00 / 96.14 | 3,024 | ≤16 actions · 92 msPUCT 2.0 · cap 262,144 |
| explore2_k16dev · strategy95-refine | 48 / 50 | 96.0%86.5–98.9% | 3 / 28 / 17 | 92.02 / 96.13 | 2,941 | ≤16 actions · 92 msPUCT 2.0 · cap 262,144 |
| explore4_k16dev · strategy95-refine | 49 / 50 | 98.0%89.5–99.6% | 8 / 27 / 14 | 91.97 / 96.11 | 2,939 | ≤16 actions · 92 msPUCT 4.0 · cap 262,144 |
| explore4_k32dev · strategy95-refine | 48 / 50 | 96.0%86.5–98.9% | 5 / 28 / 15 | 92.03 / 96.30 | 2,988 | ≤32 actions · 92 msPUCT 4.0 · cap 262,144 |
| linear95heldout · strategy95-confirmation | 965 / 1000 | 96.5%95.2–97.5% | 105 / 574 / 286 | 90.37 / 96.11 | 59,898 | ≤16 actions · 92 msPUCT 4.0 · cap 262,144 |
What changed
The frozen linear model still supplies action preferences. The optimized native engine can afford more futures per move. We raised the simulation cap to use more of the time allowance, and compared wider action shortlists and stronger exploration on shared development deals. A tested shortcut skips simulation when only one legal move exists. These are approximate searches over sampled hidden worlds; they receive only the player's observation.
Evidence, not a solved game. Error bars describe sampling uncertainty. Development results were used to select settings. A held-out point estimate above 95% is different from a confidence interval entirely above 95%. All intensive runs used ARM Linux on an Apple M5 Max, restricted to two CPU cores and 2 GiB; the eventual deployment CPU may be slower.
Audit and reproducibility
- 1,300 saved games fully replayed through the engine with matching outcomes, enemy counts and Jester usage.
- Confirmed engine: 64 Python tests passed on Linux, 64 on macOS, and 7 Rust tests passed. JavaScript syntax and the live hint API were checked.
- Recomputed every summary from individual game rows; checked all decision-time denominators, finite values, medal totals, no truncated rollouts, and unassisted/unrewound replay lengths.
- Development, confirmation, model-training and model-validation deal sets do not overlap. Development intervals do not account for model selection.
- Measured on local ARM Linux with two CPU cores and 2GiB cap. This is not a latency guarantee for another CPU.
- A sample win rate above 95% does not establish that the underlying win probability exceeds 95%; inspect its uncertainty interval.
Full configurations, model copies, source snapshots, individual timings and all replays are saved in the experiment folders shown above. Failure replays are available in the playground’s Research ledger.