Game contribution
How recorded events were credited after a game. This is descriptive evidence, not a pregame prediction.
Methodology / Versioned and inspectable
How recorded events were credited after a game. This is descriptive evidence, not a pregame prediction.
A stabilized estimate of ability. It needs prior performance, role context and explicit sample-size treatment.
An estimate archived before the result. It is evaluated with calibration, Brier score and log loss.
The canonical model is box-rapm-ivw-v1: a box-score signal and team-demeaned RAPM combined in standardized units with sample-dependent weights. Play-by-play credit, live box-score estimates and tracking adjustments are separately named signals. They cannot overwrite the canonical output.
Its uncertainty calibration and inverse-variance assumptions are not guarantees of optimality. Off-ball defense, screening, teammate effects and missing tracking data remain limitations. A one-game contribution is not an estimate of career ability.
Scoreboard agreement is an attribution check. Predictive evaluation uses chronological data cutoffs and pregame probability archives. Full-season RAPM reused in both halves is not an independent reliability test; the validation page labels such estimates separately.
Portland's October 10 preseason prediction stays intact. Later snapshots preserve their assumptions, source cutoffs and changes. Only confirmed regular-season results count toward the record. NBA Cup schedule placeholders are replaced by assigned games rather than added twice.
The preseason foundation includes analyst roster and availability assumptions. Its season-total simulation tails are unvalidated. We keep that limitation visible alongside the expected record.
A finished score does not require finished writing. Articles remain pending when essential inputs or factual checks fail. Every meaningful dataset has a source timestamp, calculation timestamp, coverage report and model version. Missing information remains missing.