Measurement 5 of 5 · timed on video
Speed & temperament
Same stopwatch method as test 01: four terminals raced on one recorded screen. This time speed wasn't free. The sprinter shipped in 8:43 and crashed 15 of 20 landings. The two that took twenty minutes spent the difference on testing — and it shows in the score.
Time to done
Wall-clock from prompt to done, timed on the side-by-side race recording. Sorted fastest first. ✓ = tested its own work. Gemini never finished: an API error forced a retry, its quota ran out, the run continued on gemma-4-31b-it past the end of the recording, and then the CLI's own Node process died with a fatal “JavaScript heap out of memory” — a harness crash, not a model verdict.
Speed vs. score
| Model | Time | Autopilot | What the minutes bought |
|---|
The inversion from test 01
In the Fourier test, temperament made no difference — sprinter and deliberator tied on quality, so speed was pure preference. Here the temperaments finally cash out: Grok's sprint produced the field's hardest touchdowns (2.38 m/s, over the limit), while Fable's ten-minute verification loop is directly visible as +15 landings over the sprinter.
The honest caveat
One run per model, twenty episodes per autopilot. This is a reproducible data point, not a benchmark — the direction of the pattern (verification ⇒ landings) is the claim, not the exact digits. Rerun it yourself: the prompt is published and the referee ships inside every file.