The model's record

Every call the model makes is written to an append-only ledger before kickoff and graded against the final score at a flat 1 unit, −110, pushes refunded. Nothing is added after the fact and nothing is deleted. Almost nobody in this space will show you a number they cannot edit later — this page is that number, win or lose.

What the model does do: it knows what it doesn't know

A win probability is calibrated if the things it calls 58% happen about 58% of the time. That is a weaker claim than beating the market, and unlike beating the market it is true, and you can check it on this table. 604 in-season games from 2025 — every one from week four on, once both teams had enough drives to fit — graded through the same engine the site serves, fitted only on games played before each one.

One caveat, in the season's opening weeks: until teams have banked enough drives to fit, the model runs a cold-start prior instead, and that path is measured separately and calibrates less well than the table below. This table is the in-season engine. That cold-start path is graded too — 810 games of 2025, Brier 0.2028 against 0.1889 here, with 75.4% of margins inside its 80% band rather than 78.5%. Worse on both, and it is what the site serves until roughly week four — which is the honest reason to read this table as the in-season one.

• · Brier score 0.1889 against 0.2447 for always guessing the base rate — lower is better, so the model carries real information.
• · 78.5% of final margins landed inside the simulation's own 80% band. The target is 80.
• · Regression of actual margin on projected margin: 1.031. Unbiased is 1.0.
• · The market is better calibrated than we are: its Brier on the same games is 0.181 against our 0.1889 (skill score -0.043, and negative means behind). We publish it because a curve shown without its baseline invites you to assume the opposite.

Why closing-line value, and not a win–loss record

Because a record cannot carry the weight people put on it. To distinguish a real edge from luck at −110 you need this many graded calls — one-sided test, 95% confidence, 80% power:

A full season of published calls is a few hundred. So no honest site can settle this with a record inside one season — which is exactly why anyone showing you a hot 12-game run is showing you noise, and why the number we lead with is closing-line value instead. CLV is continuous rather than binary, so it carries far more information per call and can reach significance in weeks rather than decades.

Forward record — published calls

There isn't one yet, and this page will not pretend otherwise. The append-only ledger opened on 2026-07-01, and no published call has reached a final score. 168 calls are logged and live right now; they will appear here graded, win or lose, because they were timestamped before kickoff and cannot be edited afterwards. Check back — that is the entire proposition.

Forward record — paper calls

Pre-registered rules that are logged and graded but deliberately not shown as picks — they are on trial, not in production. Currently 49-38 from 87 — a win rate anywhere between 45.9% and 66.3% is consistent with that, which is to say it means nothing at all. It is shown only because hiding it would defeat the purpose of an append-only ledger.

Backtest — receipts, not a track record

Calls logged after the games were played, graded for transparency and kept strictly separate from anything above. A backtest can be tuned until it looks good; a forward record cannot. Treat these as a description of the method, never as evidence it wins.

How to check this

• · Pricing: flat 1u at -110, pushes refunded.
• · A call counts as forward only if its logged publish time precedes kickoff. Anything else is reported as backtest.
• · Grading is mechanical, against final scores — no judgement calls, no voided bets.
• · Record last graded Sep 10, 2026.

• Model as of Sep 1 · through week 1 · drive simulation

The engine behind these numbers is described in the methodology.