Before Week 4 we built two versions of this sheet. One was hand-picked the way we had always done it, reading usage and matchups and making calls. The other took the 36 Must Starts straight from a model: projection plus the Vegas estimated team total at quarterback and running back, projection alone everywhere else. We published the model and graded the hand-picked version off the record.
Both went 10 of 36. They agreed on 28 of the 36 Must Starts. On the eight where they differed, each side hit once: the model with the Seahawks D/ST, the hand-picked sheet with Kyren Williams, who finished RB1.
One week is not a verdict. But it fits what the larger test says. Over 18 weeks of results, 2025 from Week 4 plus 2026 so far, we re-ran 15 ways of choosing the 33 quarterback, running back, receiver and tight end Must Starts using only what was known beforehand. Projections hit 40% of the time, 13.2 a week. Projections plus the team total also hit 40%, slightly better at quarterback and running back. Ranking by each player's average over his last three games hit 35%. Ranking by last week's points hit 30%, 10.1 a week, the worst of anything we tried.
That is the uncomfortable part, because last week's box score is what most lineup decisions are made from. Three extra hits a week is the gap between the best method and the most common one.
The hand-picked sheet did one thing better: its Sits went 5 of 11 to the published sheet's 3 of 11. Both sets were judgment, so that says more about which judgment than about models, and we will keep grading it. Until the hand-picked calls beat the model over a real sample, the Must Starts come from the model, and every deviation from it is graded against the pick it replaced.