Sandbox

## Motivation As noted in `docs/ml_self_improvement.md` §2.1, the 267-key `smrt_player_season_v2` stats are the raw material for every derived attribute (pace, technique, decisions, etc.). Currently these derivations use hand-tuned formulas. This card replaces them with a calibrated gradient-boosted model trained on the residuals between formula-derived attrs and ground-truth anchor ratings stored in `attr_calibration_anchors`. ### (a) Input signal Features: per-player season aggregates from `smrt_player_season_v2` (progressive carries, pressing intensity, pass completion by zone, duel win-rates, xG-chain involvement) + position one-hot + league tier. Labels: `attr_calibration_anchors.attr_value` for the ~N anchor players where human or consensus ratings exist. ### (b) Model + loss **LightGBM regressor** per attribute axis (8 visible attrs), trained with **Huber loss** (δ=1.5) to be robust to outlier anchors. One model per attr; hyperparams tuned via 5-fold group CV grouped by `player_id` to prevent season leakage. ### (c) Eval metric Held-out MAE and Spearman ρ on 20% of anchor players (stratified by position). Target: MAE < 1.0 on the 0–20 scale and ρ > 0.72 vs. current formula baseline. ### (d) Shipping gate Offline eval passes threshold on held-out anchors → shadow-score all 893 k players → A/B compare IAMS index distribution shift < 5% p95 → enable for new players only → full rollout after one nightly cycle with no anomaly alerts.