deep-design-3x2

Topics in the source

• Deep-Design-3x2
• Why it is shaped this way
• The procedure
• Pass 1: deep design (Opus)
• Pass 2: external review (ChatGPT), blind first, then critique
• Re-measure the starred claims (you, or a fresh Sonnet agent)
• Optional: a second Opus pass
• Merge (you)

At least three solutions, two passes, a re-measure, one merge, decisions back to the user. Pass 1 (Opus) keeps three real options alive and writes down every claim its pick rests on. Pass 2 (ChatGPT) answers the problem blind before it sees pass 1, so it is an independent sample, not a critique of a framing it inherited; then it critiques. The starred claims are re-measured by a different method. The merge is yours, not another agent's: you hold the conversation with the user and know their rulings. This was Deep-Design-3x3 until 2026-10-18 (a second, fresh Opus pass sat between these two). It was cut to two passes on measured evidence, below. "3x3" still means this skill. Evidence: providers/chatgpt/research/multi-model-eval/ in the Dispatch repository (literature, case mining, two code-review replays, a design judgment pilot; summary in CONCLUSIONS.md there). A third pass buys little. Replaying 42 real escaped bugs: one Opus review caught 0.42; any second pass raised it to 0.51–0.55; three Opus reviews unioned reached 0.54, about +0.03 over the best two-pass shapes. Where one review was already near ceiling (a 97% pilot), no extra pass helped. The literature agrees: gains saturat

This is an excerpt. Refer to the complete source and its current requirements.