You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The committed replay export data/deploy/finals_live_predictions.csv contains play-by-play win-probability rows for Finals Games 1-3 only (504 / 489 / 480 events; file last updated June 10, 2026, the night of Game 4). The cloud refresh never captured Games 4 and 5 before it died, so the Finals retrospective page (#3, PR #4) shows "not part of the committed replay export" for those games instead of replay charts, and the exported PNGs cover Games 1 and 3 only.
outputs/reports/nba_finals_case_study_summary.csv already contains the locked live start/final leans and results for all five games, so the retrospective's metrics (live final lean 5/5, opening lean 1/5) are unaffected. This issue is about the missing per-event replay curves only.
Fix
Regenerate the Games 4-5 replays locally from the NBA API and re-export the deploy-safe file:
Pre-game predictions and correctness values in outputs/reports/finals_upcoming_predictions.csv are locked history (fixed in Freeze 2025-26 Finals showcase into a retrospective #3) and must not change. Diff the report CSVs after the pipeline run and restore any prediction columns the rebuild overwrites, mirroring restore_completed_pregame_predictions() in scripts/cloud_run_refresh.py.
The live start/final leans for Games 1-3 already committed in nba_finals_case_study_summary.csv must be byte-identical after regeneration; scoring is deterministic given the same model artifacts, so any diff there is a red flag.
No retraining. Same committed model artifacts (models/*.pkl).
Acceptance criteria
data/deploy/finals_live_predictions.csv contains all five completed games.
The retrospective page renders replay charts for Games 4 and 5 (verify via streamlit.testing.v1 AppTest, raw output pasted in the PR).
Optionally add Game 5 (the clincher) as a PNG in assets/screenshots/.
Locked prediction values unchanged (git diff of the report CSVs shown in the PR).
Problem
The committed replay export
data/deploy/finals_live_predictions.csvcontains play-by-play win-probability rows for Finals Games 1-3 only (504 / 489 / 480 events; file last updated June 10, 2026, the night of Game 4). The cloud refresh never captured Games 4 and 5 before it died, so the Finals retrospective page (#3, PR #4) shows "not part of the committed replay export" for those games instead of replay charts, and the exported PNGs cover Games 1 and 3 only.outputs/reports/nba_finals_case_study_summary.csvalready contains the locked live start/final leans and results for all five games, so the retrospective's metrics (live final lean 5/5, opening lean 1/5) are unaffected. This issue is about the missing per-event replay curves only.Fix
Regenerate the Games 4-5 replays locally from the NBA API and re-export the deploy-safe file:
python run_pipeline.py --mode collect_playoff_games --seasons 2025-26python run_pipeline.py --mode collect_playoff_play_by_play --seasons 2025-26python run_pipeline.py --mode run_playoff_case_study_pipeline --seasons 2025-26python run_pipeline.py --mode export_finals_live_predictions_for_deployGuardrails
outputs/reports/finals_upcoming_predictions.csvare locked history (fixed in Freeze 2025-26 Finals showcase into a retrospective #3) and must not change. Diff the report CSVs after the pipeline run and restore any prediction columns the rebuild overwrites, mirroringrestore_completed_pregame_predictions()inscripts/cloud_run_refresh.py.nba_finals_case_study_summary.csvmust be byte-identical after regeneration; scoring is deterministic given the same model artifacts, so any diff there is a red flag.models/*.pkl).Acceptance criteria
data/deploy/finals_live_predictions.csvcontains all five completed games.streamlit.testing.v1AppTest, raw output pasted in the PR).assets/screenshots/.git diffof the report CSVs shown in the PR).