EXP-14 / completed / CM-8 / checkpoint trajectory
EXP-14: CM-8 checkpoint trajectory and final Sunfish evaluation
Final selected CM-8: 6 wins, 2 draws, 92 losses; all 100 completed.
| Checkpoint | W / D / L | PGNs |
|---|---|---|
| 14,000 | 1 / 0 / 24 | 2h games · identity and audit |
| 28,000 | 2 / 1 / 22 | 4h games · identity and audit |
| 42,000 | 4 / 0 / 21 | 6h games · identity and audit |
| 46,875 (selected) | 6 / 2 / 92 | final games · identity and audit |
The final selected checkpoint’s matching 25-game subset is 2W / 0D / 23L. It is derived from the 100 games, not 25 additional games. Intermediate evaluations are actual games at different checkpoints on overlapping openings; do not pool these as independent strength trials. Exact checkpoint hashes, suite/opening IDs and separate run IDs are in each game’s provenance and run manifest.
Protocol and diagnostics
Frozen 50 openings paired by color; Sunfish classic e396948acbd80cdcbad3dccfb07e878fba9df8f4 requests 1,000 nodes (soft budget), BF16 state-v2 joint legal decoding, 256-ply cap including book autoplay, symmetric immediate draw claims. Adaptive Rust 1/3/5/7-ply mate helper: 50,000 nodes, 250 ms, four workers; verified all-defenses proofs. Book autoplay is excluded from model-query diagnostics. The helper budget does not bound the whole turn.
Final: 2,846 model queries, 96.84% raw greedy legality, 12 proven overrides; zero unfinished games. Complete audited metrics.
These runs compare complete systems on selected fixed openings. They do not isolate input representation, supervision density, curriculum, initialization, inherited training, precision or helper/search configuration. They establish neither causal proof nor universal strength ranking. Unknown bounded searches are not proof of no mate.
Inspect the evidence
All run identities, checkpoint/suite hashes, protocols and results · Allowlisted source ZIP · Original/rendered source hashes · Training source · Evaluation source · Consolidated game viewer.
Original code is Apache-2.0; the external Rust helper retains its separate GPL terms. Host paths are redacted; datasets, optimizer state and private launch configuration are not bundled. Source identity is inspectable, but this is not a turnkey training replay. Missing game-start timestamps remain unknown.