Why architecture matters here
Reflection fails without loop bounds (cost + latency), without verifier (self-affirming loop), and without eval (unmeasured value). Architecture matters because these guardrails convert reflection from research demo to product.
The architecture: every piece explained
The top strip is the core loop. Initial answer generated. Self-critique identifies flaws. Verifier tool — external check like code runner. Revise produces new answer.
The middle row is safety. Loop bound caps iterations. Convergence check stops when stable. Diverse critics from multiple perspectives. Ranking picks best.
The lower rows are ops. Cost + latency — reflection uses more tokens. Eval — did it help? Ops — task fit + budget + governance.
End-to-end flow
End-to-end: math problem answered. Verifier (calculator) checks arithmetic; error found. Self-critique identifies where. Revise. Verifier confirms. Return. Eval shows reflection lifts task F1 +8% at 3x cost.