Run the experiment yourself.
38,199labelled positions
$0.77Jev calls for training data
~1 minutelaptop training time
What was held out, and what was measured?
The score comparison uses seeds 100–119, untouched by training or tuning. Each game runs until a crash or the configured 2,000-move limit. All players receive the same board information, represented differently for Jev and Echo.
Scores are averages over 20 games. Timing is the median across games of each game's median move time, measured one player at a time on an Apple M4 CPU. Jev includes the API request; Echo and Echo + search run locally. The near-equal 7.5 and 7.9 means do not establish exact equivalence.
Echo matched Jev’s top move 93% of the time in the held-out imitation evaluation. That is a different measurement from game score or overlap between full probability distributions.
What does the search actually do?
It expands every offered move breadth-first and evaluates the living boards at each level in one Echo batch. A leaf is valued by food gained, or −(score + 10) on death. Each move is weighted by Echo’s preference multiplied by exp(child value), then normalized. The weighted mean returns to the parent, and the root chooses the highest-weight move.
Depth adapts to the time budget. Simulated food uses fresh randomness rather than the real game’s future random sequence. The Jev search time and cost figures are estimates for comparable work, not a completed Jev search benchmark.
Can I run it myself?
The repository includes the training scripts, the trained Echo model (87 KB), the held-out results, the evaluation commands and the game. The current implementation uses MLX on an Apple Silicon Mac. Recreating teacher labels requires access to Jev; running the trained Echo model locally does not make cloud calls.
Reproduction commands on GitHub ↗