Levers

A lever is something that might move a player’s points. Live = in production (volume_v1q); staged = in volume_v4 (v4_fix), which is not live until the check over 2026 weeks 7–12 and the owner’s yes (OPT-11). A lever is kept only if it beats the model of the day on a season it never saw: MSE not worse and a significant MSE or quantile-score gain. “Gains of X or more ruled out” means the test had 80% power to see a gain that size. “Untestable” means one validation season is too short to see any plausible effect: not disproven.

Where the error is. Shares are the reducible part of that position’s MSE (ranges run between two noise floors). Realistically a pre-lock signal could remove QB 1.9%, RB 3.4%, WR 2.2%, TE 3.0% and D/ST 1.4% of MSE, most of it in share (reports/r9-error-budget.md, corrected after review). Each position says where its remaining error sits, so new ideas can aim at it.

A lever is an idea that might change a player’s projected points. “Live” ideas are already in the model we use now (volume_v1q). “Staged” ideas are in the next model (volume_v4, also called v4_fix). It is not live until we check it on 2026 weeks 7 to 12 and you say yes (OPT-11).

An idea is kept only if it beats the current model on a season the model never saw. Its average squared miss (MSE) must not get worse, and it must show a clear gain in MSE or in the range score. “Gains of X or more ruled out” means the test had an 80% chance of seeing a gain that big, and did not. “Cannot be tested yet” means one test season is too short to see any realistic effect. That is not the same as proven wrong.

Where the misses come from. Each position shows which part of its error we could realistically shrink. The ranges sit between two noise floors (levels of luck we cannot beat). Realistically a signal known before lock could remove 1.9% of QB MSE, 3.4% of RB, 2.2% of WR, 3.0% of TE and 1.4% of D/ST, mostly by getting each player’s share of the plays right (see reports/r9-error-budget.md, corrected after review). New ideas can aim at those spots.

MSE

The average of each miss squared. Smaller is better. A gain means the miss got smaller.

Likely range

The span the true gain probably sits in (a 95% interval). If it includes zero, the gain may be luck.

Looks

How many times we peeked at test data. More looks means more chances for luck, so the bar goes up.

Held-out season (fold)

A past season kept out of the model’s learning, used only to test it.

Implied total

The points the betting line expects a team to score.

D/ST

A team’s defense and special teams, scored as one player.

EPA

Expected points added on a play.

q (false-alarm score)

How likely a finding is a fluke after counting all our tries. Small is good; near 1 means no evidence.

Pre-registered

We wrote the test down before looking at fresh data.