04 · Research
The daily card only shows picks we put real money behind. Everything else lives here — markets the model produces, scored against outcomes, transparently held to the same standard before any of it earns a spot on the card.
Some markets the model is good at; others it isn't ready for. Rather than just shipping picks we're not confident in, we track them publicly in shadow mode— generate the pick, record it, resolve it against the outcome the next day, but don't bet a dollar. A market only graduates onto the daily card after it clears a fixed promotion bar (typically ≥150 settled shadow picks plus a clean error rate).
P&L below is paper for prop markets and shadow for derived markets — no money at stake. CLV is measured for runline only; alt totals and F5 share infrastructure with the main closing-line capture but aren't fetched at T-10 by design until they promote, and prop closing odds aren't consistently captured by our data feed.
Derived from the moneyline ensemble + totals model via Pythagorean inversion. A side-conditional isotonic calibrator (v2) ships in the inference path but stays inert in the picked-prob range until ≥150 closed picks accumulate. Promotion gate: that sample plus a clean Brier read against full-game equivalents.
Pitcher K over/unders. The model produces picks; they're paper-traded so we can score the projection without committing units. Roughly break-even so far — close to the bar, but not enough edge to displace ML/totals.
Hitter hits over/unders. Hardest market we track — line liquidity is thin and prop juice eats most of the edge. Currently net negative in paper, which is exactly why we don't bet it.