The Human Layer: What Stats Miss, and How We Will Test It

Where the numbers stop

Team efficiency, schedule strength and roster quality explain a lot about a college football game. Our models measure those things and show their work. What they handle poorly is the material that never reaches a box score: a starting quarterback who is listed as healthy but is not right, a coaching change that lands in the middle of a season, a locker room that has stopped believing, a momentum swing that the stats only catch a week later.

That gap is why we think of Newton in two parts. The first is the quantitative core, which turns team strength into a forecast. The second is a reasoning layer, still in research, meant to read what a formula cannot: the news, the context, and the parts of a game that are hard to reduce to a column.

What the core has already shown

We replay past seasons by forecasting each one using only the seasons before it. The team had already studied the 2024 and 2025 seasons, so these replays are backtests, not blind tests, and none of this is a live record.

On those replays, the consensus betting line still picks more winners than Newton and lands closer on the final margin. Almost all of that gap sits in close games, the near coin flips where nobody picks well. We say that plainly because it is the honest benchmark.

The result we find most interesting is about line movement. In the same replay, when a point spread moved between its opening and closing number, it moved toward Newton v3's side of the opening line more often than away from it. The pattern was stronger when v3 disagreed sharply with the opening number, and the older version of Newton showed only a faint lean, within what chance produces. The numbers are in the line-movement replay. This describes direction only. It does not show that any line moved because of us, and it is not a betting edge. It is also a pattern found looking backward, which is why we have frozen a forward test of it (more on that below).

Not betting advice: this describes how lines moved in past seasons. It is not a guide to placing a wager.

The smallest human input that worked

Most of what we tried did not help, and How Newton v3 Was Built lists what we cut. One human input did help. Knowing that the starting quarterback is out improved held-out 2024 and 2025 results, and in 415 games with a starter out, average margin error fell from 12.48 points to 12.20. The check counted players benched during games as missing, so it is a replay built after the fact, not a blind test.

In the product the input is optional and entered by hand. It has no effect unless an absence is recorded before the cutoff. That is the human layer in its smallest, most honest form: one fact, measured, with a clear rule for when it counts.

Why the reasoning layer waits

The reasoning layer is the part we are least willing to oversell. It is a language model that reads stored facts and news and proposes small, logged adjustments. It cannot be graded on past games, because a model trained on the internet has already read those results.

So the rule is strict. Its proposals are logged before the games they cover. It does not change any published number until those logged proposals beat the core on games played after they were frozen. Ask Newton can explain a pick and the context behind it, but it never changes the numbers. We have no results for the reasoning layer yet, and this note does not claim any.

Judged forward, not backward

Here is what is in motion. On October 17 we froze a line-movement test, committed before any game it covers. It applies to FBS-vs-FBS games kicking off on or after October 18, 2026 at 00:00 UTC, with no model changes. It has not been scored. A separate set of 99 FBS forecasts was frozen on October 16, before the first kickoff of its slate, and it is also unscored. Newton's picks update until kickoff with every version kept, and since October 17 its live record against the market is graded only from picks saved before kickoff. This note contains no live results.

Why freeze tests at all? Because a record means something only if it was written down before the result. A frozen forecast cannot be quietly adjusted after the fact, including the ones that end up looking bad. The same goes for our own rules: when we chose this-season-only rankings, we measured what that choice costs first, and the rankings head-to-head has the numbers.

What "75%" should mean

The 75% figure gets quoted often, so it is worth being exact. A 2005 academic study measured the closing Las Vegas line at 74.7% on 1,582 games between top-division teams from 1998 to 2001, and the best blend of six or seven rating systems at 72.9%. One published power-rating system reports about 75% over ten seasons, a figure it reports about itself. Seventy-five percent is roughly market level. No model has shown it can reach that every season on games between FBS teams.

The market's own hit rate also moves with the season and with which games you count. It is lower when only FBS teams play each other and higher once mismatches against FCS opponents are included. Any accuracy number means little without its cohort. We will state the cohort every time, and we will not change the cohort to reach a number.

The close

Stats tell most of the story, and the core is built to tell it honestly. The human layer is the part we are still earning. Its proposals will be logged before kickoff and judged only on games played after that. If it does not beat the core, it does not touch a published number.