Game board

Gap vs market (percentage points): raw model probability minus the no-vig market probability for the named side. Positive favors that side relative to the market; negative favors the other side. It is not profit or expected return.

Evidence conviction (0–100): the share of 1,000 resampled prior-season week clusters in which the raw model beats market Brier in the same edge range. At least 50 graded games and 10 weeks are required. A score of 78 means 78% of those historical resamples favored the model—not a 78% chance this bet wins or is profitable. It does not distinguish between opposite sides; the signed favorability supplies direction.

Market-shrunk edge: a challenger that learns how much to trust the raw forecast using earlier seasons only. Zero means it falls back to market, not a verified bet.

Analyst conviction: a separate −100 to +100 weighted pick-consensus metric. No verified timestamped analyst picks are imported yet, so it remains unavailable and has zero influence on forecasts. Missing is not neutral.

All probabilities condition on no push/tie. No current offered payout is verified here, so break-even, expected value, ROI and executable signals are unavailable. Scores and explanatory terms are not causal evidence.