The reassuring second look
A model writes an explanation, rereads it, and says it is sound. Sometimes that second pass catches a contradiction. Sometimes it merely repeats the same mistaken premise with greater fluency. The ability to critique one's own words is useful but not a reliable independent oracle.
The problem is not that a model can never detect an error. It is that generation and self-judgment may rely on the same missing fact, prompt framing, or reward to sound persuasive. A confident self-check can conceal a shared cause of failure.
Separate actor from checker
The PHS W15 rationale argues for separate responsibilities: active reasoning layer, persona substrate, Memory evidence, Governor evaluation, execution authority, and receipts. It explicitly says this defense in depth does not guarantee safety and can fail through correlated errors. The enforcement standard prohibits self-approval for consequential actions and asks for external, attributable evidence.
A second model can help, but two models may share training, prompt, or retrieval errors. Independent tests, deterministic validators, source inspection, and real-world readback are stronger when the claim calls for them.
Pause here: Which error could the same model catch by rereading, and which error needs a genuinely different source or test?
Use an external question
Choose the check that matches the risk. Arithmetic can be recalculated by code. A published page can be fetched and compared. A philosophical interpretation can be challenged by a reader with a different perspective. A model's self-review remains one input, not the final judge of its own success.
The red guide may reread his directions. It is wiser still to ask the traveler where the path actually led.
My answer: Self-review can improve an answer but cannot reliably certify it. Match important claims to independent evidence, validators, and authorized review.