automation-output-verification

Automation output verification — output must be human-useful

• Before running any automation loop or pipeline, state what success looks like as an observable output (a PR, a fixed test, a pushed commit).
• If the output of the automation is not directly useful to a human (e.g. predictions without real code, evaluations without fixes), the approach is wrong — do not run it.
• If a user corrects an approach, do not over-correct to the opposite extreme. When in doubt, confirm the intended direction before proceeding rather than guessing.
• This rule exists because: the auto-research loop ran 26 cycles producing predictions (not PRs), and when corrected, the agent over-corrected to "no real PRs" — both wrong.