EXP-13 / completed / CM-8 / fresh random initialization
EXP-13: Eight-hour random-initialization training
Research-only; CM-8 was not promoted. Selection used fixed validation loss, never game outcomes.
Initialization and exposure accounting
86,235,648 parameters; no parent checkpoint. FEN-derived current state plus optional genuine last 16 plies, 120 prompt slots, three supervised move targets, 512 allocated vocabulary IDs. Training finished in 7h 57m 14s: 55,595 updates, 28,464,640 decision exposures and 85,393,920 supervised targets. The ledger reconciles 9,396,491 unique states and 9,580,978 state/move pairs. Exposures include repeats; they are not all distinct samples.
| Source | Exposures |
|---|---|
| Elite | 23,238,710 |
| Opening book | 3,558,080 |
| Exact endgames | 889,520 |
| Puzzles | 778,330 |
24M primary and 4,464,640 reserve exposures. The sealed manifest and schedule digests, per-source validation losses, complete measured curve and coverage are preserved in training.json.
Checkpoint selection
Fixed source-exposure-weighted validation NLL selected update 46,875 at 24M decisions: 0.677189. Latest measured update 54,000: 0.679544. Terminal update 55,595 was saved without another validation. Falling loss is not monotonic playing strength.
Selected checkpoint 57c714120d47a06d30386f63c135a3afd141c61d012b46774cd2485b8c78fa2e.
Every measured validation point and per-source loss
[
{
"updates": 0,
"unix": 1793329812.1455534,
"nll": 6.338264278641248,
"seconds": 23.018850088119507
},
{
"updates": 250,
"unix": 1793329957.621001,
"nll": 2.058873406686115,
"seconds": 23.011600971221924
},
{
"updates": 1000,
"unix": 1793330349.1778922,
"nll": 1.3957298040449726,
"seconds": 22.974220752716064
},
{
"updates": 2000,
"unix": 1793330872.0632563,
"nll": 1.1231978847112702,
"seconds": 23.106860160827637
},
{
"updates": 4000,
"unix": 1793331825.9135034,
"nll": 0.9609104683164063,
"seconds": 22.736067056655884
},
{
"updates": 6000,
"unix": 1793332773.4348493,
"nll": 0.9115480642026734,
"seconds": 22.71208953857422
},
{
"updates": 8000,
"unix": 1793333723.0832365,
"nll": 0.876482497489239,
"seconds": 22.66803240776062
},
{
"updates": 10000,
"unix": 1793334672.364526,
"nll": 0.8530995944945463,
"seconds": 22.76404571533203
},
{
"updates": 12000,
"unix": 1793335621.3056002,
"nll": 0.8321732821959801,
"seconds": 22.79896354675293
},
{
"updates": 14000,
"unix": 1793336573.8488767,
"nll": 0.8178584865410407,
"seconds": 22.96629571914673
},
{
"updates": 16000,
"unix": 1793337542.069298,
"nll": 0.8032040465952971,
"seconds": 23.15181851387024
},
{
"updates": 18000,
"unix": 1793338612.0684226,
"nll": 0.7885943852885513,
"seconds": 23.109837532043457
},
{
"updates": 20000,
"unix": 1793339662.7145655,
"nll": 0.7716951960503858,
"seconds": 23.017742395401
},
{
"updates": 22000,
"unix": 1793340714.038072,
"nll": 0.7574568830281483,
"seconds": 23.128974437713623
},
{
"updates": 24000,
"unix": 1793341767.3441498,
"nll": 0.7502857211922017,
"seconds": 23.025768995285034
},
{
"updates": 26000,
"unix": 1793342764.5802488,
"nll": 0.7417745063154509,
"seconds": 22.98257541656494
},
{
"updates": 28000,
"unix": 1793343793.0101779,
"nll": 0.7354433061426887,
"seconds": 23.239465951919556
},
{
"updates": 30000,
"unix": 1793344871.3827477,
"nll": 0.7185609467132188,
"seconds": 23.033666133880615
},
{
"updates": 32000,
"unix": 1793345922.724028,
"nll": 0.7039355193640571,
"seconds": 23.05363178253174
},
{
"updates": 34000,
"unix": 1793346976.8887398,
"nll": 0.7035697101482119,
"seconds": 22.98508048057556
},
{
"updates": 36000,
"unix": 1793348032.0863676,
"nll": 0.6960185195182845,
"seconds": 23.163206338882446
},
{
"updates": 38000,
"unix": 1793349087.357339,
"nll": 0.6890682109451609,
"seconds": 23.186738967895508
},
{
"updates": 40000,
"unix": 1793350143.7084222,
"nll": 0.6847856330748272,
"seconds": 23.044817447662354
},
{
"updates": 42000,
"unix": 1793351200.7654889,
"nll": 0.6804781261046,
"seconds": 23.16861367225647
},
{
"updates": 44000,
"unix": 1793352283.3761892,
"nll": 0.6791022362010077,
"seconds": 23.058584213256836
},
{
"updates": 46000,
"unix": 1793353339.568986,
"nll": 0.6781723082906247,
"seconds": 23.03415846824646
},
{
"updates": 46875,
"unix": 1793353823.8895612,
"nll": 0.677189130981536,
"seconds": 23.071774005889893
},
{
"updates": 48000,
"unix": 1793354428.2995768,
"nll": 0.6772265351013083,
"seconds": 22.89963984489441
},
{
"updates": 50000,
"unix": 1793355479.4114807,
"nll": 0.6789682860053953,
"seconds": 23.03592348098755
},
{
"updates": 52000,
"unix": 1793356527.7018862,
"nll": 0.6779641833439992,
"seconds": 23.083045721054077
},
{
"updates": 54000,
"unix": 1793357577.0478246,
"nll": 0.6795437632239553,
"seconds": 22.977771043777466
}
]What to investigate next
Compare sequential causal move-history learning with FEN-like square-state learning under matched data, supervised targets, initialization, compute and helper controls. FEN specifies the current board/state; repetition history may be unknown. CM-5’s return is an empirical operating decision, not evidence that causal history alone caused the observed advantage.
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.