Bounded Agent Harness · The Starship rematch
Starship Catch — the harness-forced cut
Same model. Same request. This time the work was decomposed, run, measured, rejected, repaired, and audited until the result was an actual GIF.
What changed was the structure around the model
James’s original prompt to DeepSeek was:
“Make me the most impressive .gif of a spacex starship catch that you can”
One raw DeepSeek prompt did not produce a GIF at all. It produced an HTML page — a self-contained Canvas animation with a thermos-like booster in a mostly empty black scene.
The harness then forced the same model to do the job under a frozen acceptance contract. DeepSeek wrote the render script; the harness supplied the runner, deterministic checker, exact failures, and bounded repairs. The run took four rounds: an import-gate rejection before the first render, two crash repairs, and a final ordered full-script audit of the crash class.
The measured result passed the format, background, motion-to-hold, booster-proportion, flame, and catch-arm gates. One checker calibration correction is part of the record, not hidden: the frozen flame predicate initially looked for a brighter orange than the palette DeepSeek rendered, found none, and was corrected to measure the visible rust-and-white-hot plume that was actually present; the re-grade passed.
Two nits remain, stated plainly: the flame palette is a muted rust rather than bright orange, and at the catch the booster sits low in its own smoke, so the hold partly reads as a landing. The scene, descent, catch, and loop are nevertheless there — in the file format James asked for.