Three days before the YellaWood 500, this site published a Talladega board and told you to ignore it. Ryan Blaney sat on top at 12.0%, the model had no demonstrated ability to pick a drafting race, and the piece said so in the headline.
Kyle Larson then won from 10th. The model had him fourth at 7.8%.
Every board on this site gets written down before the green flag and scored after the checkered, and by that accounting Sunday came out 31.3% better than pulling a name out of the entry list. Second best of the five races scored this way all season, behind only Larson's Kansas win in September.
Nobody here is taking a bow for it.
That 31.3% rests entirely on one car. The number asks a single question, which is how much better the published board was than random chance at identifying the winner, and a 7.8% pick is supposed to come in roughly one time in 13. Sunday was that time. Look at what the board did with the other 36 cars and the picture changes fast.
Of the nine drivers the model rated highest, five never made it to the end. Blaney, the favorite, started 27th and was done after 131 laps in a crash, credited with 38th. Joey Logano won the second stage and finished 34th. Christopher Bell broke a suspension and finished 35th. Chase Briscoe started on the pole, led nothing, and finished 33rd. Austin Cindric crashed out of 27th. Chase Elliott and Tyler Reddick were still running at the end and still finished 36th and 37th.
Running up front had nothing to do with any of it. Chris Buescher led 53 of the 189 laps, more than anyone in the race, won the opening stage, and finished 32nd. Logano led 23 and finished 34th. Larson led 23 and won. William Byron came home second having never led a lap, from 17th on the grid. Ricky Stenhouse Jr. was third from 24th. Cole Custer finished fifth, and the model gave him four tenths of one percent.
Fourteen of the 40 starters retired, a dozen of them in wrecks. That is 35% of the field, the highest of the ten Talladega races in this site's data since 2022, where the usual figure is closer to one car in five.
Which is the race the preview described, right down to the attrition. The model priced the chaos correctly and then benefited from where the chaos happened to land. Five earlier drafting races in 2026, a season the model was never allowed to see while it was being built, came out 3.7% worse than random when they were scored together. This one came out 31.3% better on its own. A single race moving a measurement that far is an argument for collecting more races, not for believing the model has learned to read a pack.
The place it does say something honest is in the shape of the board. At Talladega the favorite got 12.0% and the eighth choice got 5.6%, barely half a step apart across eight cars. For Martinsville on Sunday, the same engine has Blaney at 18.7%, Denny Hamlin at 16.5%, and Larson at 14.7%, with the field falling off a cliff behind them. When it knows something it commits. A flat board is the model shrugging.
As for the title, Larson arrived in Alabama five points clear of Bell and leaves 73 clear of Hamlin, who quietly finished eighth while the contenders stacked up behind him. Bell is 74 back, Briscoe 79, Logano 100, and Blaney 107, with two races left.
So a championship that had been separated by a handful of points for a month got sorted out by which cars had an escape lane when the pack came apart. That is the complaint everyone has about this place, and it is a fair one, and it is also exactly why the one number the model published with a warning label attached is the one that happened to come in.