Play of the day

The model prices every game and every player market on the board. A play becomes a candidate when our number disagrees with the book by enough to matter — three percentage points on a spread, two on a moneyline, five on a player prop — and only when the edge runs the same way as the side we'd take. That last part does more work than it sounds like. On a moneyline the model's favourite and the better price can be different sides; when they disagree, there's no play.

Then a human looks at it. Plays we pass on never appear here, and the record below counts only what we actually posted. Once a play is up it stays up at the line and the price it was called at, even if the market moves — a record that quietly re-prices itself isn't a record.

Grading is flat: one unit a play, at the posted price. A push counts in neither column. A player who never dressed voids his prop rather than winning it for the under, which is what the book does too. We grade against the book's line, not our own projection, because the line is the number you could have bet.

What's not here, and why. The NBA game total and both team totals are on the matchups board with our number beside the book's, and they are deliberately not eligible for a play. Over the regular season so far the total has graded at −1.7% flat and −9.9% once our disagreement passes three points — it gets worse as we get more confident — while betting every under blind returned +10.5%. A market that loses to its own blind baseline doesn't get staked, however good the gap looks on a card. Rebounds and assists sit out for a different reason: both are judgement calls at the scorer's table, and our box-score source isn't the league's official one.

The samples here are small and will stay small for a while — a play a night is a play a night. Percentages wait until there are enough to mean something, and a unit total of ±2u over a few weeks is noise, not a result. Read it that way.