Leo's Lab
Learning Dashboard
Slice every graded game by the dimensions that matter: where the model is right and wrong, whether the edge is real, and which reads actually predict. Measurement only; it never changes a pick.
The daily learning summary (data/learning-summary.json): headline, findings, official-pick gates, and the multi-day record, straight from the same file the pipeline writes every night. This did not appear anywhere on the site before 2026-08-08.
Are we improving?
20-game rolling win rate on the model's side. Dashed = 50%.
Win rate, record, Brier and average edge for each value. Click a row to see the games behind it.
From data/calibration/confidence_report.json, written nightly by scripts/grade-confidence.js (the "Grade confidence" workflow step). Does the Lab Rating actually predict the outcome, and which of its components are earning their points?
Which Lab Rating components earn their points
Worst-rated losses
The highest Lab Rating scores that still lost, and the points breakdown that produced each score.
Leo is the official model: calibrated, and it drives real picks. v3.1-bullpen ("shadow") runs beside it on every graded game and never drives a pick; its numbers are public only on the shadow view. This compares them game by game and night by night: who called it, whose probability was closer to the truth (Brier score, lower is better), and where they disagreed.
Night by night
Click a date to see every game that night: both models' probabilities, which side each favored, and the actual result. A red Brier delta means shadow was worse that night; green means shadow was better.
Pick any past graded date and game to see the real inputs the model used (starter quality, bullpen state, offense form, market price) and the shadow model's read alongside Leo's. Then try swapping a pitcher, a bullpen, or an offense in the what-if panel below and see both models' probabilities recompute live.