What the numbers mean.

Projection and probability are different.

The projection estimates a player’s stat total. Probability estimates the chance of a particular side at one exact line. A projected average above the line does not automatically mean More is the model’s most likely outcome.

Sport-specific models use available player history, recent performance, and relevant opportunity or role inputs. Where supported, matchup, availability and market evidence inform the model or its confidence checks. Missing data limits the read. Each hub shows its model version and available evidence; not every displayed split directly changes the probability.

AI explanations do not set or rewrite the statistical model’s probabilities. An estimate is not a promise.

Read hit rate with its sample size.

Hit rate = wins ÷ (wins + losses). Pushes are excluded from that denominator. Tracked includes pending records; settled can include pushes. A small sample or a selected group of signals can look stronger than the full model.

The public Model Ledger contains recorded model signals, which can include multiple props on the same game. Those results are not independent, and individual-prop hit rate is not a multi-leg payout or profitability measure. The Bozzfather’s curated targets are a separate record.

Historical/backtested results must be labeled separately from predictions recorded before games. We do not manufacture results to fill an empty dashboard.

Data reliability is not win probability.

The decision summary keeps the statistical estimate separate from data reliability. Supported, Limited and Needs refresh describe source freshness, sample coverage and availability checks. They are not measured accuracy percentages.

Pricing edge appears only when comparable, fresh prices exist on both sides of the same exact line from at least two books. It is the model probability minus the median no-vig market probability, in percentage points. A difference between your line and a market line is in stat units, not probability edge. The model can use market inputs, so agreement is not independent confirmation.

The evidence drawer labels model inputs, decision gates, descriptive comparisons and experimental research. Streaks and recent hit rates are historical descriptions. They are not added to the production probability as a second bonus.

What qualifies as pregame evidence?

The public performance view uses first server-recorded exact-market forecasts for each exact model version, with the server receipt arriving before the recorded game start. Local records, research experiments, late uploads and legacy rows without pregame evidence remain separate. This establishes receipt timing against the stored game start, not an independent audit of every input.

The ledger compares probability bands, sport, prop, side, signal and model version. It reports sample sizes, unique games, hit rate and calibration. A small or correlated sample does not establish a stable advantage. Pushes, voids and pending outcomes are excluded from win/loss hit rate; calibration uses the eligible cohort stated on the dashboard.

Scheduled settlement reads supported final player/game stats when enabled. Unavailable stats remain pending or require review. Corrections preserve outcome revisions; past outcomes are not silently rewritten. No historical results have been manufactured for this release.

How a Paper Pick gets graded.

A frozen entry records the exact player, game, prop, line, direction, and available model snapshot. More wins above the line; Less wins below it. An equal result is a push. Missing or unsupported final data is marked for verification instead of invented.

Paper Picks uses custom simulation rules: all legs winning pays the total-return multiplier you entered; any loss loses the simulated stake; otherwise, a completed ticket containing a push refunds the stake. Pending legs stay pending. These rules are not any third-party platform’s payout rules.

Simulated ROI uses net BB divided by BB risked on decided entries. Bozzle Bucks have no cash value and no real wager is placed.

Start with one game.

1. Choose a game and player.
2. Select the prop, check the current context, and enter your exact platform line.
3. Analyze, then add your selected side to Paper Picks.
4. Review and freeze 2–6 legs. Return after the games to check results.