calibrated SPORTS
NFLReplayFilm Room

Replay · Film Room

Every rule the engine has replayed, and the 4 it cannot run.

4
Rules replayed · 1999–2024
4
Rules it cannot run, with reasons
18,486
Bets across the 4 replayed

Replayed rules

4 rules replayed (verdicts: Loses 3, Unclear 1); 0 with an ROI interval wholly over zero. 4 more are listed that the engine cannot run.

LosesN = 13,680

Fade every over

prop under on receptions/receiving_yards/rush_attempts/tackles_assists/sacks, 2023-2025 REG weeks 1-18, 0 condition(s), consensus_close at close

ROI −3.3% [−5.0, −1.8]Hit 52.5%

UnclearN = 2,256

Rush attempts unders

prop under on rush_attempts, 2023-2025 REG weeks 1-18, 0 condition(s), consensus_close at close

ROI −1.8% [−5.0, +1.0]Hit 52.4%

LosesN = 1,150

Chase the streak

prop over on receptions/receiving_yards/rush_attempts/sacks, 2023-2025 REG weeks 1-18, 1 condition(s), consensus_close at close

ROI −13.2% [−18.2, −8.5]Hit 47.0%

LosesN = 1,400

Home underdogs of 3 to 7

spread underdog on spread, 1999-2025 REG weeks 1-18, 2 condition(s), consensus_close at close

ROI −5.6% [−10.7, −0.4]Hit 48.8%

AUDIT-2026-09-24, the Lab library launch set · lab index written Sun, Sep 27, 2:14 AM ET · universe built Sun, Sep 27, 2:14 AM ET

How this is computed
Each rule is written in the Lab’s strategy language and run by the same engine a typed rule will run, over the seasons it names, at the closing price, one unit flat. The verdict is set by a fixed rule from the ROI interval, the luck check and the holdout, not by judgement: “Loses” means the ROI interval lies wholly under zero; “Unclear” means the ROI interval spans zero and the luck check is outside the 5th to 95th percentile, or could not be run.Every card’s track is on one scale, with zero marked. ROI is published only where it averages at least 10 cleared bets’ prices.

Side by side

Any two rules, the same five figures, on the same terms.

CompareRush attempts undersChase the streakHome underdogs of 3 to 7withFade every overRush attempts undersHome underdogs of 3 to 7
FigureChase the streakFade every over
VerdictLosesLoses
Bets1,15013,680
Hit rate47.0% [44.1, 49.8]52.5% [51.7, 53.4]
ROI · 95%−13.2% [−18.2, −8.5]−3.3% [−5.0, −1.8]
Ruleprop over on receptions/receiving_yards/rush_attempts/sacks, 2023-2025 REG weeks 1-18, 1 condition(s), consensus_close at closeprop under on receptions/receiving_yards/rush_attempts/tackles_assists/sacks, 2023-2025 REG weeks 1-18, 0 condition(s), consensus_close at close

Rules it cannot run

4 rules are in the library that the engine cannot run, each with its reason, shown rather than hidden.

Cannot run

Follow the model's leans

  • refused by lab.strategy.check: side 'model_lean' needs walk-forward model predictions, and none are persisted for the historical seasons (the stored predictions are the live model, in-sample on history)
Cannot run

Buy the middle

  • not expressible in lab.strategy/1: a middle is TWO legs at two books at two thresholds; lab.strategy/1 is a single-leg language (one side, one price source), and the register figure is taken over every fresh pre-kickoff snapshot of the raw archive, where the universe holds one close per claim. No strategy in this schema expresses it, so no run of this engine can reproduce it - the library card must cite the R16 script, or the schema must grow a legs field (a lab.strategy/2).

Measured elsewhere: R16 / brief 023 Part 2, research/bookvbook.py: spreads middle mean EV -2.54pp empirical, -2.65pp realized, n=10,860 over 848 games (data/repairs/bookvbook-2026-09-17.txt)

Cannot run

Arbitrage between books

  • not expressible in lab.strategy/1: the same two-leg shape as the middle

Measured elsewhere: R16: receptions arbitrage mean size +1.09pp

Cannot run

Unders in the wind

  • refused by lab.strategy.check: condition on 'weather.wind_forecast': unavailable - no forecast weather is joined to the universe; observed weather would be look-ahead for a pre-kickoff rule

Conditions no rule can use yet

The engine’s catalogue of conditions (lab/nfl/catalogue.json) is not published, so the conditions it marks unavailable, and its reasons, are not listed here. None is written in their place.

Held out

2025 has not been revealed for any of the 4 replayed rules.

  • PendingFade every overheld out: no figure in this result includes the 2025 season
  • PendingRush attempts undersheld out: no figure in this result includes the 2025 season
  • PendingChase the streakheld out: no figure in this result includes the 2025 season
  • PendingHome underdogs of 3 to 7held out: no figure in this result includes the 2025 season
How this is computed
The latest complete season, 2025, is held out of every figure until it is revealed, once, and the verdict above is then replaced by the held-out one. A rule is stored exactly as the engine approved it, with its hash, so a changed version cannot pass for the one listed here. Rules are not yet graded forward week by week; that is not connected, and nothing here says otherwise.

Hypothetical results · nothing here was bet · not advice · Deliberately restricted: no bet carries its book price, its book, its stake or its profit, and the file has no equity curve. The prices are The Odds API's, whose terms (read 2026-09-24, updated 31 Aug 2026) say nothing either way about displaying historical odds, so they stay on the server and only aggregates leave it. A winning flat bet's profit is its price minus one, and consecutive points of a cumulative curve difference back to it, so those are withheld for the same reason. ROI and units appear only where they average at least 10 cleared bets' prices. Do not add these fields back without a decision that the terms allow it.