CM-7 / live
State-conditioned curriculum model
Continued from CM-6 through a finite mixed curriculum and opening/endgame/Elite follow-up; selected by validation NLL at follow-up update 2,232.
EXP-11Checkpoint registry
One checkpoint. One stable name.
CM identifies selected model weights. EXP identifies the work that produced or evaluated them. Serving releases and helper changes are tracked separately.
Live and browser checkpoint: CM-7, adopting state-conditioned input. CM-6 remains a research-only predecessor.
CM-6 / research only
State-conditioned sanity checkpoint
Warm start from CM-5 with expanded vocabulary and new input/objective; not promoted.
EXP-8Curated-data continuation
Weights continued from CM-4 with fresh optimizer; historical full-history production model.
EXP-7Broader-data foundation
Trained from scratch on a broader corpus; not continued from CM-3.
EXP-5Full-history baseline
Fresh random initialization; historical browser checkpoint.
EXP-3Training pilot
Fresh random initialization; brief training pilot.
EXP-2Untrained reference
Untrained reference; two local validation games.
EXP-1Fresh random initializations are separate branches. The confirmed weight-continuation chain is CM-4 → CM-5 → CM-6 → CM-7; CM-6 changes the vocabulary and objective. CM-1, CM-2 and CM-3 are not claimed as ancestors of CM-4.
Append-only identifiers for selected research checkpoints and documented experiments, ordered by first experiment. Unselected terminal weights remain explicitly identified by hashes in provenance. Serving releases, helper changes and optimizer artifacts do not receive new CM numbers. IDs are never renumbered or reused.