The evaluated file is the public v2 Baseline training.bin checkpoint:
No training update occurred during evaluation.
For each opponent, the current policy played 5,000 games: 2,500 as first player and 2,500 as second player. Policy actions were selected greedily from legal actions. The engine used the same rules and hidden-information observation boundary as training and the desktop UI.
| Opponent | First-player wins | Second-player wins | Total | Win rate |
|---|---|---|---|---|
| Uniform random | 2,140 / 2,500 | 2,124 / 2,500 | 4,264 / 5,000 | 85.28% |
| Handwritten heuristic | 2,012 / 2,500 | 1,976 / 2,500 | 3,988 / 5,000 | 79.76% |
| Saved champion | 1,339 / 2,500 | 1,313 / 2,500 | 2,652 / 5,000 | 53.04% |
The v2 checkpoint materially exceeds the included random and heuristic baselines, performs similarly from either seat, and slightly exceeds its saved champion snapshot in this sample.
The v3.1 Windows package also contains a validated recurrent-belief checkpoint:
AAIV3C01, inner version 31The current v3 policy was evaluated without training updates, using a fixed seed and 5,000 games per opponent: 2,500 as first player and 2,500 as second player. Actions were selected greedily from legal actions.
| Opponent | Wins | Total | Win rate |
|---|---|---|---|
| Uniform random | 4,294 | 5,000 | 85.88% |
| Handwritten heuristic | 4,195 | 5,000 | 83.90% |
| Fixed public v2 Baseline | 2,728 | 5,000 | 54.56% |
| Saved v3 Champion | 2,439 | 5,000 | 48.78% |
For an additional promotion check, the saved v3 Champion scored 2,632 / 5,000
against the fixed v2 Baseline (52.64%). The last recorded GUI evaluation in
training-evaluations.csv was 123 / 200 (61.5%) against v2; that smaller
sample is retained as training history, not presented as the release result.
These measurements establish performance for this exact checkpoint and seed. They do not establish expert-human strength, statistical significance across multiple seeds, or a general solution to the game.
The automated suite separately checks 14,848 enumerated outcome combinations and their player-swapped mirrors, 5,000 complete random games (80,638 actions in the recorded run), hidden-information isolation, exact checkpoint resume, checksum/backup recovery, evaluation-state isolation, selected numerical gradients, and 1,000 complete games through the UI transition wrapper.
The originally reported outcome was also reproduced. In that screenshot, the AI had two, not three, Daredevils after resolution and won by catching the human player. A separate test asserts that receiving a true third Daredevil loses unless a simultaneous-result tie rule applies.