[arhancanli]

Self-play search targets can repeat the engine's tactical blind spots, but directly inserting Stockfish's normalized centipawns would also change the trainer's score calibration. Add an isolated teacher-label tool with explicit raw-cp audit and wdl expected-points modes. WDL encodes win plus half a draw through the inverse of the existing sigmoid400 target scale, while retaining raw scores for audit.

The legacy records omit castling, en-passant and repetition state. Uniform fresh-prefix sampling therefore admits only unambiguous irreversible-move positions. Each independent single-thread Stockfish search uses full strength and fresh root state. Labels keep score, WDL, depth, nodes and legal PV from one completed unbounded UCI iteration; interrupted final bounds remain separately auditable. Paired files preserve every byte except the White-relative score and remain outside live self-play. No model, engine default or champion is changed.

Validation: 53 controller and teacher tests pass, including 11 regressions covering omitted state, corrupt records, score/side conversion, bounds, interrupted iterations and WDL scaling. A supervised 18-core pilot labelled 4,096 frozen positions at 100k nodes each. Independent Rust decoding and real trainer decoding agree; 1,558 interrupted roots were recovered without resampling or changing the previous 2,538 exact labels. WDL reanalysis of all 4,096 boards reproduced every raw score exactly. The expected-points encoding agrees with real trainer decoding, with maximum target rounding error 0.000309. A selected one-million-node follow-up covered 59 positions: all 27 mating evaluations remained, eight best moves changed and three expected scores changed by at least 0.1. The 4,480-game strength checkpoint remained unchanged during the reserved 18-core run, and the normal supervisor resumed. GitHub CI passed c67449c, including release workspace/perft/SPRT calibration, WebAssembly and all 52 Python tests. The sampler follow-up 7795cd2 removes the eager candidate-index allocation, rejects duplicate accepted offsets and passes CI with 53 Python tests; teacher scoring and WDL conversion are unchanged.

This is a conditional data-quality experiment, not proof of playing strength. WDL models Stockfish self-play under its reference conditions. Before model comparison, use grouped held-out board/side splits and identical training rows/resources, then require a fresh prescribed strength gate. The existing larger-network experiment, search queue and 14-opponent milestone remain in place.