Train NAIL Gap-Returner Priors

Train the NAIL-RAPM v1.2 gap-returner candidate through the final completed season, then replay the three frozen evaluation seasons and run the paired bootstrap non-promotion gate.

uv run nba-train-nail-gap-returners --through-season 2025-26
uv run nba-evaluate-nail-gap-returners
uv run nba-bootstrap-nail-gap-returners

Render the production model's 14-panel context-coefficient audit:

uv run nba-audit-nail-v12-context-weights

The training command writes an immutable recursive run beneath artifacts/models/forward_nail_rapm_v12_gap_returner_priors/2025-26/. It also persists gap_returner_projected_states.parquet, which is an audit of the internal annual bridges. The rows are not player-season RAPM observations.

The frozen evaluator compares the completed v1.1 state to the candidate on 2023-24 through 2025-26 without updating either model. The bootstrap command uses game-block resampling within season and requires the candidate's upper 95% full-game RMSE bound to be no worse than 0.5% of v1.1 in the pooled sample and each season before it can be promoted.

See NAIL-RAPM v1.2 Candidate: Gap-Returner Priors for the forward contract and results.