Self-review underperforms structurally, not just anecdotally. A model reviewing its own just-generated output still has its own generation reasoning present in context, which makes it measurably less likely to question decisions it already made moments ago — it's inclined to confirm its own reasoning rather than challenge it.
Remedy hierarchy, most to least effective:
- An independent review instance without the generator's reasoning context — most effective, because it has nothing to defend.
- Extended thinking on self-review — weaker; more deliberation from the same biased starting point still starts from that bias.
- "Review critically, assume bugs exist" prompting — weakest; wording alone doesn't remove the underlying context bias.
Multi-pass review splits a large review into per-file local passes plus a separate cross-file integration pass for data-flow analysis that spans files. One single pass over many files at once produces shallow findings and sometimes outright self-contradictory conclusions — flagging a pattern as wrong in one file while implicitly approving the identical pattern elsewhere in the same review, simply because attention thinned out differently across the files.
Confidence-reporting verification passes — having the model self-report a confidence score per finding — are genuinely useful for routing attention (review the low-confidence findings first), but unreliable as an escalation trigger on their own. Worth holding this in mind alongside Domain 5's escalation-pattern guidance: self-reported confidence shows up as an unreliable proxy in both places, for the same underlying reason — the model can be confidently wrong on exactly the cases that are hardest.
The discriminator: independent context beats better instructions, and attention dilution is a capacity-to-attend problem, not a capacity-to-hold problem — so a bigger context window doesn't fix it. Fourteen files fit fine in a large window; the review is still shallow and contradictory, because the issue was never whether the files fit, but how much genuine attention each one got.