Week 3, Settled

The early‑season window

Before the numbers, the calendar. Many of the model’s most important inputs are rolling windows of the current season: how an offense and a defense have performed over their last four games, and their last eight, adjusted for who they played. Through Week 4 those windows are not yet full, and the model leans on last season as its prior for the parts of a team that have not yet shown themselves this year. Weeks 1 through 3 are, by design, the stretch where it knows the least, and we said as much before the season.

Looking back over the last three seasons, the model’s best stretch has come once those windows fill, from Week 5 through Week 8, and its confident calls have landed most often in exactly that span. That is an expectation, not a promise, and we will grade it in public like everything else. But it is the reason to read the first three weeks as a wildcard we anticipated rather than as the shape of the season.

Two tracks, graded side by side

This was the first full week of the card carrying both Model picks and Analyst picks, each graded on its own. Neither track cleared its break‑even. The model track went 2–7–3 on the card; the analyst track went 5–12. Across everything published, the split is below.

What the analyst track got right was volume: Jordan Love’s attempts, Drake London’s catches, Jaxon Smith‑Njigba’s catches, Quinshon Judkins’ rushing yards, and Monday’s under. What it got wrong was game lines, which went 1–5, and the plays the whole market had already seen coming: Dalton Schultz and C.J. Stroud with Nico Collins out, Seattle laying eight and a half against a backup quarterback. When the line has already moved on the news, the news is no longer an edge.

The most instructive play of the week is the one we argued ourselves out of in writing and kept on the card anyway. The model had Chicago to win outright on Monday night from Wednesday onward, at prices as long as +160, and our written read leaned the other way: a third‑string quarterback was starting, and that seemed to settle it. The pick stayed live at +154, and Chicago won 27–7 with Case Keenum throwing for 247 yards and two touchdowns. The lesson is not that the model is always right. It is that when the market has already moved five points on a story, the story is no longer an edge — and it is not a reason to fade a number either. That goes into how the two tracks are weighed from here.

Where the week actually went

Ten markets became published plays. The two that finished ahead were the small ones: pass attempts went 1–0 and passing touchdowns 1–1. Receptions carried the volume again and, again, the split inside it is the story.

Receptions unders went 7–6–1 for +2.50u. Receptions overs went 4–7–3 for −4.50u, and the pattern is the one Week 2 showed: plus‑money overs on secondary receivers, where the model’s number sat furthest from the market’s. Anytime touchdown went 0–7 with one void, at the half‑unit rule, which is what the rule is for. Three weeks of this is enough to act on, and the section below does.

Tiers and days

Thursday was the week’s bright spot: all three card picks in Atlanta at Green Bay landed, and the night went 4–1 for +2.80u. Sunday carried the volume and the losses. Monday’s four plays finished ahead, on the under and Chicago’s outright win.

For the third week the higher tier did not separate from the lower one. Across three weeks the picks where the model’s number sits furthest from the market’s have landed least often, not most. That is the single most important thing the record has told us, and it changes how the card is built.

What is different in Week 4

Three weeks is still a short record, so these are the changes the record already supports, in the order they matter.

Every play, all 56

Every play, ordered by game day, at the stake printed on the card at the time. Analyst marks the research track; Strong edge and Edge are the model’s tiers. Nothing here is selected after the fact: this is the same list that was live before kickoff. A void is a play that never truly settled and returns the stake; Week 3 had five.

Swipe the table sideways for price, result and units.