NFL 2026 forward test: the commitment as published
Published Aug 6, 2026, before Week 1, and reproduced below word for word. The results, every graded row and the grading rules are on the record.
House punctuation only: each em dash in the original is shown as a comma. The record now calls Arm A the Broad test and Arm B the Focused test. Same rows, same grading; only the names changed.
This is a pre-registered test, not a betting board.
Before the 2026 season we wrote one rule down and froze it: on six NFL player-prop markets, flag the under when our projection lands 10–20% below the line, in weeks 1–9 only. Then we stopped touching it. Every row below was logged before kickoff and graded afterwards, win or lose.
Every row carries play = 0 . Nothing on this page is a recommendation, and we make no claim about returns. We are finding out in public, with the losses attached, whether a rule that looked real in a backtest survives a season it has never seen. Most of the ones we have tested did not.
- The rule was fixed in advance. The band, the side, the six markets and the week window were written down before Week 1 and cannot change mid-season. Changing any of them would void the test, so we won’t.
- The model is frozen. Every projection comes from one committed artifact trained on data through 2025, asked and never retrained. Its hash is published below.
- The record is append-only. A graded row never disappears. Losses, pushes and DNPs sit in the same tables, at the same weight, as wins.
Frozen model artifact sha256 5f6c085ac3e9b1d9… · trained on data through 2025 · board last published 2026-11-05 23:45 UTC
That hash is the point. It is stamped on every board we publish, so anyone can check that the model which produced Week 1’s rows is byte-for-byte the model that produced Week 9’s. A mid-season retrain would change it, and would void this test.
The two arms, from the same page
Two pre-registered arms over the same season. Arm A is the broad control, all six markets. Arm B is the narrower mechanism it was supposed to be about. Reporting both is the point: if only the narrow one works, that is much weaker evidence than it looks.
- Arm A, all six markets
- Arm B, receiving family (+ rushing only in a flat/down rushing year)
Units are a measurement at the price we logged, on a flat 1-unit basis. No bet was placed on any of it. A pick can appear in both arms; the two records overlap by design and must not be added together.
One later sentence on how the focused test’s rushing inclusion is decided is not reproduced (owner-only since Oct 14, 2026).
The week window, from the same page
The pre-registered window is weeks 1–9. No new pick can qualify from Week 10 on, so this week is projection-quality rows only. The window was set in advance, the original signal was an early-season regime effect that died after Week 9 in backtest, and extending it now to chase results would be exactly the mistake this test exists to avoid.