We don’t guess.
We measure.

The rest of the season is already sealed

Nine calls, one per remaining callable round, committed in a single act under the frozen v4 configuration. The lap times stay sealed until each session runs; the hashes are public now, bound together under one manifest.

Locking race by race would have left one hole nothing in the scheme could close: nobody can tell how many circuits were considered and skipped. A model that quietly declines the ones it feels shaky about produces the same evidence as one that calls everything, and scores better. Removing a circuit from a sealed set changes the manifest, so the option stops existing.

What it costs is said out loud: nine calls made in July cannot learn anything from the season. That is given up on purpose. And the expected result is on the record in advance: three to eight hits out of nine, centred on 5.5, with a 37% chance of at least one strong miss. Nine out of nine would mean the bands were too wide to be wrong, which is a failure of calibration rather than a triumph.

Las Vegas is not among them, and the reason is worth reading before the season starts. Its 2025 reference session ran wet, pole set on intermediates, so a dry-line prediction built on it is not a prediction. From v3 the model refuses that input rather than producing a number from it. But the defect list also records the uncomfortable half: this circuit had already been named in writing as the likeliest miss, and the rule that removed it was written afterwards, from facts that were public the whole time. So the season is nine calls, the baseline keeps its tenth, and the head-to-head is declared as nine rounds rather than ten.

And the model has a rival it does not beat yet

Take a circuit's 2025 pole time and multiply by one constant, 1.0323. That rule, scored under the same seven folds as the physics, lands at σ 0.99% against the model's 1.414%. Published because it is true, not because it flatters, and two re-freezes have widened the gap rather than closed it.

Seven circuits cannot separate those two numbers, so this does not say the trivial rule is better. It says the physics has not earned its complexity yet, and that is the honest state of the evidence. The baseline has its own ten sealed calls and the model has nine, so they overlap on nine rounds and the head-to-head is declared over those nine. Las Vegas, where the model abstains and the baseline does not, is reported on its own rather than folded into the comparison. The paired test was declared before any of it runs, so the season settles this instead of an argument — though nine rounds is not enough power to settle it, which the null-model document says outright.

What this is not trying to be

A Formula 1 team simulates its own car with its own CAD, its own wind-tunnel maps and tyre data nobody outside the paddock sees, on a rig with a driver in it. This is not that and will not become that. Getting closer to a real lap time than a works team is not the goal, because it is not a contest that can be won from public GPS traces.

There is a floor under everyone, and it is measurable. Two team-mates in identical cars qualify a median 0.37% apart, and that gap is form on the day, traffic, and how the track came to them. No model of a car can see any of it. So 0.37% is the best any car model could ever do, this one currently sits at 1.414%, and the honest ambition is the distance between those, not perfection.

And the limitation is not vague, it is the dominant error source and it can be measured. Everything here is built from a position trace at 7.7 Hz, one sample every eight metres at racing speed. That is the whole geometric input.

Follow it into a corner. A slow hairpin has about twenty samples in the entire corner, and the raw trace carries GPS noise, so curvature has to be smoothed before it can be used. That smoothing spans 75 metres of arc; ninety degrees of the same hairpin is 39. The filter averages over more corner than the corner has, and what comes out is an average of a curvature the data never resolved.

Nothing about careful coding fixes that. It is what a racing line reconstructed from public GPS costs, and it is why moving the filter by one step shifts the predicted lap by up to 2.57% — wider than σ itself. The biggest open problem here is a data problem dressed as a numerical one. The tyre says the same from the other side: a team gets a full model under NDA, this one fits two numbers from a single lap, and downforce is not identified at all.

What is unusual here is not the accuracy. It is that the number goes out first and gets graded in public. Teams do not publish a pole prediction before qualifying and let it be scored, and the reason is not that they cannot: a call that misses by a second is a headline on Monday. Being able to afford to be wrong in the open is the whole advantage this project has, and it is what makes the record an experiment rather than a demonstration.

The clearest evidence of that is on this page. Multiplying last year's pole by one constant scores better than the physics does, and that is published because it is true. A simulator that never runs the comparison never has to find out.