THE RECEIPTS
Every claim we make, what it's based on, and how it's graded. Hover any number to see the file and the hash it came from.
ALWAYS FREE · NO ACCOUNTNO MANUAL EDITS · EVER
PROVEN-BLIND
HOW WE DID ON A SEASON WE'D NEVER SEEN
The 2025 season was held out of fitting and tuning. The model never saw it while it was being built, so when we walked it forward through the season one week at a time — projecting each week from what was knowable that Thursday — the results were a test, not a description.
And every bar was committed to git before the number that would be measured against it. You can check the order in the repo: the standard is timestamped earlier than the result, so it could not be moved to fit what came back.
PROVEN-BLIND
WHEN WE SAY 80%, IT HAPPENS 80%
A range is only useful if it's honest. Ours are measured, not asserted — and where the measurement says they aren't good enough, we say that instead of publishing them.
RECEPTIONS
Brass tick = the 80% we promised. Bar = what happened, over 3,627 blind predictions.
RUSHING YARDS
Brass tick = the 80% we promised. Bar = what happened, over 1,779 blind predictions.
RECEIVING YARDS
Brass tick = the 80% we promised. Bar = what happened, over 3,394 blind predictions.
ON TWO PROPS WE PUBLISH NO RANGE AT ALL
On targets our range catches the outcome 87.4% of the time when we say 80%. On carries, 89.7%. That is the cautious direction — the ranges come out too wide, not too narrow — but wide is not calibrated, so we don't publish a range claim on those two props until we've fixed it. You'll see the projection; you won't see a range promise we can't back.
It also means the coverage figure we quote — 80.0% — is measured on the three props we actually publish a range for, not pooled across all five. Pooling in the two we withhold would flatter the number using precisely the ranges we're declining to stand behind.
PROVEN-BLIND
WE'RE NOT PARROTING AVERAGES
A model can look accurate by always predicting a player's average. Ours doesn't — it ranks player-weeks better than the trailing average does on all five props, with every confidence interval clear of zero.
Receptions+0.041 [0.027 – 0.056]
Targets+0.035 [0.024 – 0.045]
Receiving yards+0.050 [0.032 – 0.067]
Rushing yards+0.034 [0.016 – 0.052]
Carries+0.027 [0.017 – 0.036]
In plain terms: when we say this week is different for a player, it usually is.
PROVEN-BLIND
THE BREAKOUT SIGNAL
When our flag fires, the player sustains an expanded role over the next four weeks 53.3% of the time — measured on 287 flags inside a graded population of 1,725 player-weeks.
We disclose the weak year: in 2023 the flag came in below the control — it did worse than simply picking players on a hot streak. three of four seasons beat the control, not four. And 93.3% of the flags we show end up gradable; the other 6.7% never resolve, because the player didn't play enough afterward. Whether a player appears again isn't knowable on a Thursday, so we show the flag anyway and count it honestly.
HOW OFTEN WE'RE CLOSE
COUNT PROPS
WITHIN 1
WITHIN 2
WITHIN 3
Receptions
49.0%
78.4%
90.4%
YARDAGE PROPS
WITHIN 5
WITHIN 10
WITHIN 20
Rushing yards
28.0%
49.4%
71.1%
Receiving yards
21.2%
39.7%
67.9%
Half our receiving-yard calls land within 13.0 yards. The average is dragged up by blow-up games — which is exactly the part of football nobody can predict.
LIVE 2026
THIS SEASON, GRADED LIVE
Every Thursday we publish. Every following week we grade. The scorecard below updates itself — we don't touch it.
WEEK 1 PREDICTIONS POST THURSDAY, SEPTEMBER 10.
Grading begins the following week. This table fills itself.
harness: scheduled · tue 09:00 ET · writes are append-only
NOT CLAIMED
WHAT WE DON'T DO
NO
QB PASSING PROPS
We don't model them. (QBs are projected for rushing.)
NO
TOUCHDOWNS
Highest-variance stat in football and we haven't proven a TD model. Until we do, we won't publish one.
NO
FANTASY POINT TOTALS
They're built from touchdowns, so they wait on the same thing.
NO
SEASON PROJECTIONS
Ours doesn't beat a simple pace estimate yet. We're not selling it until it does.
NO
PARLAYS OR COMBINED PROPS
Coming in a later version, after they are validated the same way as everything else.
NO
CLAIMS ABOUT BEATING MARKETS
We've never measured ourselves against closing lines, so we don't claim it.
WHERE THESE NUMBERS COME FROM
Every figure on this page is read from a committed export in the model repo, each carrying the hash it was computed under. Nothing is typed in by hand, and the build refuses to ship a number with no export behind it.
What that does not mean: these bars were not registered before the 2025 season was played. They could not have been — this work began in 2026, after the season finished. Anyone claiming a pre-season lock on 2025 results, us included, would be claiming something that does not exist. What we have is held-out data and a bar that predates its own measurement.
grade tables docs/results/ · exported 2026-08-03
backup model 5.9% of yardage routes · flagged on the card, never dropped