← Paper Trail

Correction · filed 1 October 2026

Week 1 was graded wrong.

Not the picks — the grading. The record published after the Week 1 games counted things it should not have, and failed to count things it should have. Both errors are described below, with the corrected numbers and the reason each one happened.

---

First: the sheets were never touched

Both published sheets still hash to the values that were posted before kickoff. Anyone can download them and check.

The commitment held. The error was downstream of it, in how the committed picks were later counted.

---

Error one — picks were graded that were never published

The board writes a pick to its log every time prop markets are loaded. Markets also get reloaded after a sheet is built, to capture closing prices. Each of those reloads could write new picks — picks that were never on the sheet and never had a hash attached to them.

The Week 1 grade counted everything in the log. That swept in 5 best bets and 3 worth-a-look picks logged between 10 and 15 September, after the Week 1 hash was already public.

The five uncommitted best bets went 4–1

---

Error two — picks were published that were never graded

The sheet builder took every ungraded pick in the log regardless of which week it belonged to. The night before the Week 1 build, the Week 2 schedule had been loaded, so Week 2 game lines were written onto the Week 1 sheet.

The giveaway is on the sheet itself: it lists 25 games. Week 1 had 16. The extra nine are Week 2 games.

That put 11 Week 2 picks and 21 Week 2 refusals on the Week 1 sheet. On 17 September — before any of those games were played — a clean-up tool removed ungraded picks made by an older model version, and deleted all 32 rows. They were published, and then never graded.

Graded now, at the prices the sheet printed, against the final scores

---

The corrected numbers

The sheet is the commitment, so the honest record is every pick the Week 1 sheet published, graded — no more and no less.

The true figure is 2.47 units worse than the one originally published. The worth-a-look record, the skip record and the closing-line rate for Week 1 used the same contaminated set and are being recomputed on the corrected basis.

Week 2 does not change

Week 2 was graded by hand, keeping only picks logged before the sheet was built. It reproduces the sheet's own header exactly — 47 best, 72 look, 747 refused — and stands at 21–26, −8.61 units. No Week 2 figure moves.

---

What has changed, and what has not

One fix is live. Four are not yet built, and this page will not pretend otherwise.

Shipped and running

• A sheet covers one week's games only. The builder now filters every row by the week it belongs to, and reports anything excluded rather than dropping it silently. The Week 4 build this morning reported zero rows from other weeks.

Not yet built

• Storing the exact set of rows a sheet committed, so the record counts those rows and nothing else.
• Marking a pick logged after its sheet was built, keeping it, and excluding it from the record rather than relying on timing.
• Giving the closing-line refresh its own path that cannot add a pick.
• Preventing the clean-up tools from removing any row that appeared on a published sheet.

Until those four exist, the commit boundary is applied by hand every Tuesday, and that is a weaker guarantee than code. It is written here so the gap is on the record rather than only in a to-do list.

---

Why this is published at all

Nobody asked for this correction. No one had noticed, and the error ran in the favorable direction — the published record looked better than the real one.

A record that is only corrected when someone catches it is not a record. The whole claim here is that the number can be checked by anyone, which is worth nothing if a mistake in our favor quietly stays on the board.