GitHub - makinux/adversarial-panel

A Claude Code skill for multi-model adversarial mutual review — multiple models answer a question independently, attack each other's answers across rounds, and a facilitator synthesizes agreements, live disagreements, and calibrated confidence into one answer. The adversarial structure of GANs, transplanted to inference time. An LLM's answer is most dangerous not when it is wrong, but when it is c

[nguyen-oi]

複数モデルでデベートさせて結論出すの面白い。同一モデルのセルフレビューよりバイアス避けられて説得力ある