Don’t Grade Your Own Homework
If a system runs while you sleep, the obvious question is: how do you know the work is any good?
The tempting answer is to let the system check itself — find its own mistakes and quietly fix them. It sounds efficient. It’s actually how you lose the thread, because a system that grades and corrects its own homework leaves you no record of what went wrong, and worse, its checking shares the blind spots of its doing. Self-certified work is exactly as trustworthy as the certifier — which is the thing you were trying to verify.
The pattern that works is separation: the checker checks, reports, and does not fix. One part of the system does the work. A separate part compares the result against what was supposed to happen and files what it finds — honestly, including the failures — for review. Fixing is a decision, and decisions route to whoever holds judgment: usually you.
This mirrors how trustworthy human organizations already work. The auditor doesn’t rewrite the books; the inspector doesn’t patch the building. Not because fixing is bad, but because the same hands finding and fixing destroys the record that lets anyone else trust the result.
The payoff is the whole autonomy bargain: you can let a system run unattended precisely because you’ll get a truthful account of what happened — not a tidy story from something that already cleaned up after itself.
Related
- From Operator to Reviewer — the role that consumes those honest reports.
- AI That Doesn’t Wait to Be Asked — why unattended work needs this rule.