self-verification

Self-verification is the model checking its own answer before handing it over, the way you re-read an email or plug a solution back into the equation. After producing a draft, the model is prompted to look for mistakes: does this satisfy every constraint, does the arithmetic hold, did I answer the actual question. When it spots a flaw it can revise. On many tasks this second pass catches careless slips and lifts reliability for free, using nothing but the model's own reading of its work.

It is not a cure-all, and it is worth knowing the failure modes. A model blind to its own error in the first place is often blind to it on review, and the same overconfidence that produced a wrong answer can produce a wrong all clear. Verification works best when checking is genuinely easier than solving, or when an outside oracle does the judging: running the code, evaluating the formula, comparing against a known constraint. Hardest of all is honest verification of fuzzy claims, where there is no crisp test to fail.