Calibrate Forward RAPM To Wins

This command builds a retrospective, usage-conditional team-win calibration. It uses only the frozen prior RAPM available before each season, then weights those player priors by the lineup-stint seconds that the team actually used in that season. It does not use age, box-score, draft, physical, or target-season RAPM inputs.

Run it after training the regular-only forward lagged-prior RAPM exemplar:

uv run nba-calibrate-forward-rapm-wins \
  --forward-lagged-run-dir artifacts/models/prior_rapm/2025-26/forward-lagged-rapm-2025-26-20260803T203054Z-c627d89d

The command writes an immutable run under artifacts/models/forward_calibration/2025-26/ and updates latest.json. It is also indexed in local MLflow when automatic tracking is enabled.

The first four seasons with usable frozen priors warm up the calibration. Each subsequent season is evaluated using a weighted linear mapping fitted only on earlier completed team-seasons. The 2025-26 target fit therefore uses 1997-98 through 2024-25 outcomes, never 2025-26 outcomes.

See Forward RAPM Calibration for the model contract, equations, artifacts, and results.