the story
I do not write Python.
I have been arguing about college football rankings since 1997, I run a firm that helps leadership teams decide where AI goes first, and I have not written a line of the code behind this poll. An AI wrote all of it, inside three rules I wrote down before it started.
why I cared
In 1997 Nebraska finished first with the coaches and Michigan finished first with the writers. Both of them were undefeated and neither of them played the other. I have never really let it go.
1997 two national champions
Both trophies were handed out in January 1998. The season they were handed out for is 1997.
The BCS came the next year to take the bias out with a formula.
In my opinion it never worked, because it was just blending the human polls plus some other data. The bias didn’t come out. The formula did math on it.
Even though CFB abandoned the BCS formula the human bias still exists in the polls.
A formula was never the wrong idea. That one just wasn’t better than what it replaced, and the tools to build better didn’t exist yet.
So I wrote three rules down before I looked at a result.
1) You can’t cheat by looking ahead.
2) You have to be only obsessed with accuracy, nothing else.
3) No human bias can be introduced (polls or models that use the polls).
I wrote the rules and drew the lines the model can’t cross. The machine searched inside those lines and chose every number.
That took a Claude subscription, a computer and an internet connection. Any fan can have the same three tools tonight.
The systems this poll is measured against
These are the rating systems I could score this poll against. The number is how often each one picked the winner of a game. 2 of them are above it, and the rows are ordered by that number.
- SRS / Massey69.1%
How much you win by, adjusted for who you played. Nothing about a team beyond the games in front of it.
- Elo68.8%
Every result in the order it happened, updating as it goes. It cannot go back and redo September once October is in.
- This poll68.7%
Every game and every play of the season it is ranking. No ballots, no recruiting rankings, no other rating system.
- Colley67.9%
Who beat who. One of the computer rankings the BCS used. It never looks at the score.
- Win percentage66.1%
Wins over games played. It never looks at who you played.
- Random walker65.0%
Walk the schedule at random, moving at every step to a team that beat you. It never looks at the score either.
- Home team wins56.3%
Pick the home team, every week, all season. Not a rating system. It is the floor everything else has to clear.
4 other systems were scored on the same games and are earlier versions of this model rather than anybody else’s. They are not in the list above, and their numbers are on the full scorecard, 2 of them above this poll as well. Every rating system I could find to measure this against keeps its code private. A fraction behind and completely open is the trade I am making, and closing that fraction is what the season is for. The whole scorecard →
- Built from
- 1,585 games from 2021 through 2023, the same games for every system. every file →
- This view
- thepoll.ai/cfb-poll/story
- Licence
- Ratings CC BY 4.0, code MIT. No account, no email. terms →
where I lost
I had South Carolina 14th in August. The last poll of the season had them 88th.
That was the projection, which is the other ranking this model makes: last season and the offseason, published in August, never re-run. Of the 30 teams the 2025 projection put highest, nobody fell further. And refusing the polls is not free: over 4 preseasons the AP’s own ballot averaged 13.8 of the final top 25 right and this projection averaged 12.8.
The whole post-mortem, including the fix I refused to make.written by the grading run, printed unedited
It says 81st where the heading above says 88th, and both are this team. The heading is the last weekly poll, published before the bowls. This is one week later, with the bowls played, and they climbed back.
The projection had South Carolina 14th and they finished 81st once the bowls were played. Of the 30 teams the projection put highest, that is the furthest any of them fell, and it is worth being precise about why, because the easy explanation is wrong. The model does not read the press, and here that cuts against the easy story rather than for it: the AP had South Carolina 13th in its own preseason top 25, so the writers made the same mistake from a completely different direction. What the model read was South Carolina's own 2024, where they were the 16th best team in the country by its power rating, and that one number was worth 5.9 points to their projection. The model also read what left the roster. South Carolina returned 63.4% of its offensive usage, the 105th lowest figure among the 134 teams with a row, and 88% of its passing usage, which is what a quarterback room turning over looks like in the data. That cost them 1.4 points, the portal took another 0.0, and between them they moved South Carolina from 16th to 14th. The problem is the ratio. Last season's rating can swing a team 28 points and returning production can swing one 5, so a team that arrives 16th cannot be argued down to 81st by the offseason. The grading loop is what settles what to do about that, and this season it settled it the dull way: across the 136 teams the poll ranked, all four terms come back priced about right, the furthest of them 1.4 standard errors from the published value. No coefficient here was wrong. The ratio is a property of the design, and 2025 is the first season that made it cost something. What I am not going to do is turn the returning-production dial up until South Carolina looks right. I checked: every setting that moves South Carolina down also moves Indiana down, and Indiana returned even less than South Carolina did and went 16-0. Penn State and Baylor returned more production than almost anyone in the country, 4th and 3rd of 134, and finished 48th and 75th. In 2025 returning production told you almost nothing, and the fix for South Carolina is not a bigger version of a term that did not work.
What I want back.
Football people can answer football questions. Data people can answer model questions. I want both of them arguing about the same ranking, and I would rather be corrected in public than right in private.
I still do not know who was better in 1997 (Nebraska). I would like the next argument like it settled with numbers anybody can open.
I love college football, I love helping people get things done, and I believe AI is how regular people get their edge back.