Today's slate

One distribution per probable starter. The three lines below are fixed thresholds we set — 4.5, 5.5 and 6.5 for every start — not sportsbook lines; this engine consumes no odds feed. P(over) derives from the full strip by summation, never estimated separately. Each starter publishes when the opposing lineup posts, so the opposing-starter and matchup features are built against the nine he will actually face; if a lineup is not posted 60 minutes before first pitch, the prediction publishes from a roster-filtered projection instead. The archive records which (opp_nine_source). (An earlier lineup-aware feature was withdrawn 2026-07-19 — see the engine-era note on the record — but the publish trigger it describes is unchanged.)

Published 2026-07-21 20:02:57 UTC. Graded tonight.

A full hits distribution per starting batter against the announced starter. From 2026-07-19, batter predictions publish only from posted starting lineups, as each team's card posts (typically 1 to 4 hours pre-game). The 2026-07-18 slate was the transition and was still 94% projected, so it is NOT part of the posted-only era despite following 7/17. A side whose lineup never posts pre-game gets no predictions that day. The 7/17 rows were projection-based, as declared at the time; the provenance of every row is labeled in the archive (lineup_source), so this claim is checkable rather than asserted.

Published 2026-07-21 19:31:36 UTC. Graded tonight.

P(1+ home run) per starting batter, the classic low-base-rate market. About 13% of graded starters have homered, so probabilities live in the 2 to 30% range, exactly where calibration is hardest to fake. Fully posted-lineups-only from 2026-07-19; the 2026-07-18 slate was a transition (94% still projected), and 7/17 was projection-era. Park HR context is a model feature, built from our own data and strictly prior. Probability bars on this board are drawn on a 0 to 40% scale for legibility; the printed number is the truth.

Published 2026-07-21 19:32:19 UTC. Graded tonight.

The full run-total distribution per game, published before first pitch and graded by CRPS — the proper score for a whole distribution. Every P(over) below is a summation over the published strip, never a separate model, which is what stops this board and the run distributions from contradicting each other.

Published 2026-07-21 20:00:41 UTC. 2 of 15 scheduled games — a game appears once both starters are announced and its lineups post. Graded tonight.

These lines are fixed at the same values for every start, not set to each pitcher's coin-flip point the way a market would. That places them well for a typical starter and badly for the extremes — an ace is a near-certain over at all three, a soft-tosser a near-certain under — so line Briers are flattered at both ends relative to what market-set lines would produce. Log-loss and the calibration curves score the whole distribution and do not have this problem. Benchmarking against de-vigged closing lines is goal CL-1 in the benchmarks, and is not live yet.

Probabilities, not picks. Not betting advice.