PVM-01: learning to use a chess VM.
Act Two is indefinitely paused. PVM-01 remains an experimental checkpoint, not the live player. The Act One CM-14 agent continues on Lichess.
Read the early results and unrun GRPO plan →Research / evidence before conclusions
Model training and harness design are separate lines of work. Each result belongs to its own checkpoint, configuration and protocol.
Act Two is indefinitely paused. PVM-01 remains an experimental checkpoint, not the live player. The Act One CM-14 agent continues on Lichess.
Read the early results and unrun GRPO plan →The CM-14 development record moves from 72% in a 100-game mate-only cohort to 99% in a later 50-game full-system cohort. Protocols differ; this is not a controlled search-only gain.
Sunfish results and dated Lichess ratings →Auxiliary objectives, adaptation and continuation did not establish a stronger replacement for CM-14. The compute-budget explanation remains a working hypothesis.
Outcome and limitations →CM-15 tested a future-move target. Its completed initial evaluation did not establish a strength gain over the matched reference.
CM-15: methods and games →Async GPU search, opponent-turn work and cross-game batching are the next harness direction. Historical CPU results do not validate this implementation.
The design and its tests →How to read an evaluation explains score, win rate, ablations and rating caveats.
Choose a reproduction target before downloading. Published snapshots retain their dates and known gaps.