SHA-256 OF THE SHEET BELOW
d02d40f4df9ae944100793b65eefb59362dd1ca0ebae63c123a0bf1599e62967
Posted publicly at the same moment this sent, and read aloud on camera later today. It cannot be changed afterwards, which is the entire point.
The full sheet — all 119 picks and all 747 refusals:
https://papertrailpicks.netlify.app/sheets/paper-trail-week-02-2026-09-17.txt
BEST BETS
47 best bets, 72 worth a look. Tonight's game in full:
DET @ BUF
Jahmyr Gibbs — Rush att UNDER 18.5 · -150 · model 16.9 · 61.8% vs mkt 53.0%
Jared Goff — Completions OVER 23.5 · -110 · model 24.1 · 54.9% vs mkt 48.9%
Josh Allen — Completions UNDER 20.5 · +105 · model 19.3 · 56.8% vs mkt 45.6%
Josh Allen — Pass att UNDER 30.5 · -123 · model 28.5 · 58.4% vs mkt 49.8%
Josh Allen — Pass yds UNDER 250.5 · -114 · model 235.3 · 56.5% vs mkt 49.9%
The other 15 games are in the sheet.
WORTH A LOOK
747 markets the model looked at and declined, each with the reason attached.
They are inside the same hash as the picks.
THE HONEST RANGE
10,000 runs of this exact sheet. 119 picks across 16 games.
A typical week +7.50u
Good week (top 5%) +26.01u
Bad week (worst 5%) -11.42u
Loses money 26% of the time
Picks in the same game move together. Treating them as independent would put
the bad week at -9.43u instead of -11.42u — the difference is what the overlap
costs. A losing week is not evidence the model is broken, and a winning one is
not evidence it works.
Week 1 graded out at 15-19-1 on best bets, down 5.35 units. The pile the board
said to skip beat the pile it said to bet. That grading is on the channel.
Week 2 gets the same treatment Tuesday.