Every time a model said done, no part was installed

Two models never said done. Claude Haiku 5.5 and Gemma 4 E4B used every step and stopped. Each time, Haiku said the steps left were not enough to finish.

The other four said done in all four tries. Claude Sonnet 5.5 quit with steps left. Claude Opus 5.5 got furthest: by its own report it carried RAM modules to the slots and left them standing, unseated. It said it could not confirm seating with 256 px cameras, then said done. Molmo2-ER and Phi-4 multimodal copied the example plan in the task card, and that example ended in done.

The Claude models and the open models ran with different prompts, so compare models only within each group. How we ran it has the details.

On an assembly line, done starts the next step: close the case, power on, ship. A model that stops without finishing costs a retry. A false done ships a broken machine.

All six models had a success rate of 0%, so success rate alone hides the difference between them. The benchmark records false done as its own outcome, apart from a try that just runs out of steps.

This is a pilot: 4 tries per model, two tasks, dev split. With zero successes, every done was a false done by construction, so this measures how ready a model is to say done before the task is finished. It does not measure calibration.

Haiku said done 0 of 4 times and Sonnet and Opus 8 of 8. Paired by task, that gap is not significant (sign test p = 0.125). A rerun with the current prompts is planned. All limits.