Evaluated checkpoint: CM-17 update 761,686. Play the currently deployed agent on Lichess.
CM-17 / evaluated checkpoint
CM-17: full-move prediction.
The fixed 24-hour checkpoint, not the subsequent training job.
CM-17 reads complete game history using one token per UCI move. Its 12-layer, 12-head, width-768 transformer uses rotary positions (RoPE), RMSNorm and SwiGLU. Context is 512 tokens including BOS; the vocabulary has 1,975 entries, with 1,968 move tokens. Fixed-White win/draw/loss logits are normalized separately and are diagnostic outputs.
At the 24-hour checkpoint, CM-17 finished its matched 26-game test against CM-14 with 18 wins, 5 draws and 3 losses: 20.5/26 points (78.85%). Thirteen openings were played with colors swapped. The user selected this checkpoint for promotion after the completed assessment.
The retired browser implementation was unassisted: legal-move filtering only, no mate helper or endgame takeover. The head-to-head used the separately assisted live-bot setup. Its results are system-level evidence, not unassisted browser strength, a general Elo estimate or proof across all opponents.
Source 6559295d830d560f768276edcab1532883c9d37515cd17563a863712e0431e91.