EXP-10: Finite state-conditioned curriculum

Initialization and objective

Warm start from CM-6, with fresh optimizer and scheduler. The 86,235,648-parameter decoder uses a 512-ID vocabulary and FEN-derived current state plus optional 16-ply history. Only three move-token targets are supervised. The input structure is unchanged from CM-6; there are no reasoning traces.

Full exposure accounting

29,447,421 supervised target exposures; 1,197,528,454 input slots and 1,207,344,261 serialized slots. Slots are not distinct tokens or supervised targets. The schedule retained its final partial batch, consumed each planned entry, and stopped naturally after 9,851.83 seconds (2h 44m 12s), rather than repeating to fill its eight-hour ceiling.

Phases use opening: more than ten pieces and ply below 16; middlegame: more than ten pieces and later ply; endgame: at most ten pieces. Actual source availability determined this mix; no duplication filled a phase quota. Ordinary source origins were balanced after exclusions. No opening-book targets were included in this first stage.

Selection and limitations

Pooled fixed-slice NLL fell from 0.699053 to 0.618410 at selected update 19,000 (9,728,000 exposures). The completed terminal weights had NLL 0.620069. The follow-up starts from the selected weights, not the terminal weights. Raw exact preserving accuracy changed from 86.8% to 87.0%; the paired 95% bootstrap interval spans zero. Selection used validation NLL, not Sunfish results.

Validation has 8,192 ordinary and 2,000 exact positions. Exact roots and protected source games are excluded from this continuation, but are not guaranteed unseen by inherited CM-6. Unique counts apply within this stage, not lifetime training. Canonical exact labels retain complete WDL-preserving alternatives; they do not promise fastest mate or conversion.

Selected stage-one checkpoint: eb22fc7888c98e0dd0675bc8ffe5a4ef7554039ca3ab1c59c5f0978fdcc35670. Its pre-pass serialization 92cc373b8326f0086b3869d48fc8e5878730ea4c9b6a0ea2cdb0164fca762b14 has identical model weights. These intermediate artifacts do not receive another CM identifier. Continue to EXP-11.