EXP-7 / completed / Completed experiment · 29 October 2026

EXP-7: Six-hour curated-data continuation

Elite continuation: 27 wins, 9 draws, 63 losses and one unfinished game under the frozen 100-game Sunfish protocol.

This remains a non-reasoning next-move imitation model. Evaluation alone does not establish general strength, internal reasoning, or a live deployment change. The benchmark used the frozen three-ply helper, not production’s seven-ply search.

Checkpoint lineage

Foundation 857bc93bfffb287dc67e2845a14c214a4691bc1a008dae1fa4d044e5fccc87d2 → validation-selected Elite c0a3f67ac4b364276028950b2010c5644e03efb3ca0f1228b32c36b8b5340238. Weights continued with fresh AdamW and scheduler; this was not optimizer resume.

Data and compute

Curated Nikonoel Elite, newest first from November 2025 back through part of October 2024; cutoff 2024-10-16T01:56:24 UTC. No additional rating, BOT, speed or completed-result quality filters. Standard-start legality, context compatibility, exact deduplication and historical holdout protection still apply.

3,721,125 training games / 1,000,000,401 targets. Two RTX 4090 GPUs; 5h 58m 05.915s; 109,393 updates; 3,500,576 unique games and 940,645,697 targets consumed, no repeated epoch. The selected checkpoint is earlier: update 103,873, 893,361,947 targets. Validation NLL 0.5300231569 on 37,959 games / 10,227,153 targets; initial 0.6061916479. The last checkpoint was not separately validated and was not selected.

Evaluation

CM-5 27W / 9D / 63L / 1 unfinished; CM-4 21W / 9D / 69L / 1 unfinished. Fifty fixed openings, both colors, identical three-ply helper and engine settings. Elite’s opening-46-white and foundation’s opening-03-white reached the 256-ply cap. All 100 new PGNs passed final hash and legal replay audits.

Observed Elite score 31.5–32.5 points versus foundation 25.5–26.5, depending on unfinished outcomes. The +5–7 point difference is not statistically conclusive: all paired approximate 95% intervals under endpoint assumptions include zero. Extra training and data selection are confounded; no equal-compute foundation-only control was run. No model-evaluated test set or Elo claim.

Raw greedy legality: 3,228/3,321 model queries (97.20%); helper overrides: 46. Foundation: 3,043/3,183 (95.60%); helper overrides: 29. Different visited positions limit causal interpretation; helper-assisted play is separate from the learned policy.

Replay Elite games · Foundation report · Frozen evaluation harness · Rust preprocessing source

Configuration and reproducibility

Recorded training configuration

{
  "architecture": "unchanged causal decoder-only Transformer",
  "parameters": 85940736,
  "context_tokens": 1024,
  "vocabulary": 128,
  "tokens_per_ply": 3,
  "initial_checkpoint_sha256": "857bc93bfffb287dc67e2845a14c214a4691bc1a008dae1fa4d044e5fccc87d2",
  "optimizer": "fresh AdamW, not optimizer resume",
  "precision": "BF16 autocast; FP32 model and optimizer",
  "microbatch_per_rank": 4,
  "ranks": 2,
  "accumulation": 4,
  "games_per_update": 32,
  "learning_rate_floor": 0.00001,
  "selection": "lowest completed validation NLL",
  "benchmark_games": 100,
  "benchmark_helper_plies": 3,
  "production_helper_plies": 7
}

Shared model/training code (foundation snapshot) · Frozen evaluation harness · Exact Rust codec snapshot and build instructions.

Inspect and reproduce

Download the structured experiment record and selected evidence · Download all 100 Elite PGNs · Open the unified replay viewer.

Combined PGN SHA-256 ff8c6c993c66826f935112c075df4f0f3892eb3ecda9ff0a1c851d114de9e72f.

Foundation initialization checkpoint, dataset and reproduction artifacts · Published model/data packages. These are explicitly versioned earlier packages: Elite weights, Elite dataset exports and its exact launch receipts are not yet published as a reproduction bundle. The Elite checkpoint identity and selected public training/evaluation evidence are provided here; do not substitute the earlier packages and claim an exact replay.