Test 02 · ran 2026-07-11
One prompt. Four models.
A referee that bites back.
The verdict
The model didn't matter — until the task got open-ended.
Test 01 claimed that for well-specified, checkable tasks, model choice stops mattering. This test was built to find the edge of that claim — same checkable core (physics with an exact analytic answer), plus something no spec can fully pin down: an autopilot that has to fly and land the model's own game. The checkable half converged again: all four implementations sit at ≈0% physics drift. The open-ended half split the field 20/20, 18/20, 5/20, 0/20. And the two models that tested their own work before shipping are exactly the two that landed.
The race, on video
four terminals, one screen, one DNFWall-clock times and the crash are documented in Speed & temperament. The video plays on X — this site makes no external requests.
The contestants
one of them changed drivers mid-raceThe scoreboard
every number from each app's own self-test panel| Model | Physics drift | Autopilot | Time | Lines | Cost |
|---|
Physics drift = deviation of 10 s simulated free fall from the analytic ½gt². Autopilot = landings out of 20 seeded episodes, graded by the model's own game. Time = wall-clock from the side-by-side race recording.