The gap between 75.1% step completion and 34.0% full-plan completion is the number I would keep in front of any team planning a pilot. A model can look capable step by step and still fail most complete plans, because one wrong step breaks the whole sequence. That is the difference between a good demo and a system you can rely on.

Your point about testing the Arabic interface independently is also important for MENA teams. Equivalent prompts in Arabic and English should give equivalent behaviour, but that cannot be assumed — it has to be tested with the same scenes and the same scoring, or a team can end up with a product that works well in one language and quietly underperforms in the other.

I would add one more test to your falsifiable pilot: deliberately impossible instructions, as you suggest, plus a check on what happens after a failure. Does the model recover, repeat the same mistake, or drift? How a system fails matters as much as how often it succeeds, especially if anyone is tempted to move simulation into a control loop later. Keeping that boundary explicit is the right call.