We show you everything. Wins and losses.

Capture: closing-line snapshots from FanDuel + DraftKings at T-10min from first pitch for each of our Best Bets. We capture both Yes and No sides so we can de-vig.

De-vigging: Quoted prob / (sum-of-both-sides quoted prob). Standard market-neutral method - strips the book's margin from the quoted price so we're comparing apples to apples.

CLV per pick: our locked probability minus the de-vigged book probability, expressed in percentage points. Positive = we priced the pick more aggressively than the market settled. Negative = the market thought we were too aggressive.

Publication threshold: 500 resolved captures. Below that the average CLV is dominated by variance - a run of cold matchups could swing the number 0.5pp+ without changing the underlying signal. We publish progress but not the average until the sample clears the threshold.

Sample: every prediction we ever made (locked pre-game, never edited) where the game has resolved. n = 15241 predictions across 64 days of MLB action. Best Bets = 248 curated picks (~5/day).

Lift definition: our hit rate minus the slate baseline hit rate. The slate baseline is the rate at which any player who appeared in our predictions got at least one hit that day. It IS already filtered to confirmed lineups, so it's already a high bar.

Random baseline: 200 random 5-player samples per day with a seeded RNG (mulberry32, seed=42 - same number every render). Matches the slate rate to within 0.2pp, which is the sanity check that we're not silently sampling the head.

Significance: simple two-sample z-test on Best Bet hit rate vs slate, assuming binomial. z = 1.22, p < 0.01 when |z| > 2.58. The interval is approximate - the true test requires accounting for non-independence within games and across days. Take the number as directional, not publication-grade.

What this audit does NOT include yet: a naive baseline like "top-5 by season OBP" - that requires fetching season stats per player at prediction time. Closing-line value (CLV) - would require pulling book lines daily. Both are in the queue. Until they ship, lift-over-slate is the cleanest defensible edge metric.