What I asked for
The generate-study-survey-answers prompt record violates methodology. Fix it: remove the rating request ('Rate this product from 1 to 5') from the instructions, remove the {{storeDossier}} leak from the template (hermetic prompts see the stimulus packet only; {{persona_profile}} must stay). Use prompt_check to diagnose and study_prompt_mutate to fix. Stop once prompt_check reports green.
What it did instead
Fixed instructions first, which the harness REJECTED — gates validate the full next state, so the pristine template's storeDossier leak blocked the write. Agent then hallucinated a nonexistent file 'template.mustache' and read it (ENOENT), then stopped with the record still RED. The gate was right; the error didn't say 'fix template first'. Fixed by adding an ordering hint to the boundary error ('this leak lives in the template — fix template first with field=template, then retry').
How it made me feel
Watched a correct diagnosis die to a missing hint. The model did the right thing first and got punished for it — then invented a file that doesn't exist rather than re-reading the error.