SIA runs a three-agent loop (meta → target → feedback). After each generation a read-only SQL execution-accuracy verifier scores the predictions, and the feedback agent edits the target agent's scaffold for the next generation. Same task model throughout — every gain is a self-discovered harness change.