How the V9 model produces a pick

V9 isn't a single model. It's a pipeline that runs every evening, and a gate that decides whether the run's output is good enough to publish. Either everything passes and the slate ships, or something fails and the slate is held back.

The pipeline, end to end

• Ingest: raw box scores, lineups, injury reports, weather, line history, and book odds are pulled and reconciled against the prior day's snapshot.
• Features: ~394 numbers per player are derived — recent form, matchup-adjusted baselines, pace, usage, defensive context, lineup effect.
• Ensemble: a stack of gradient-boosted trees and a transformer head co-predict per-market projections and the variance around each one.
• Calibration: the raw probabilities are isotonically remapped so that a 60% prediction actually wins about 60% of the time on hold-out data.
• Pricing: each projection is compared to every book's current line for that market; only differences large enough to clear the noise band become picks.

The readiness gate

Before any pick is shown to a user, the run has to pass a list of checks. Examples: every required upstream source delivered fresh data, no feature distribution shifted more than its allowed band overnight, the calibration set's reliability is inside tolerance, and the prior week's settled picks landed inside their predicted variance.

The dashboard's badges ("available / limited / unavailable") reflect the gate's verdict on each individual data source, so when a pick is shown, you can see exactly which inputs were healthy when it was made.