Debate as oversight: give two or more models the same question, have them argue opposing answers under a judge's eye, and let the adversarial pressure surface flaws that solo generation hides. It is scalable-oversight theory turned into an engineering pattern.

The theoretical motivation (Irving, Christiano, Amodei, 2018) is elegant: humans cannot directly evaluate superhuman reasoning, but we can evaluate a debate about it — because in an adversarial exchange, an unsupported claim is an invitation to be attacked. The empirical follow-ups were encouraging with caveats: Du et al. (2023) found multi-agent debate improved factual accuracy and reasoning across GSM8K and chess tasks; Khan et al. (2024) showed non-expert judges could better identify correct answers after reading model debates. But the pattern degrades in known ways — judges favor confident, eloquent debaters over correct ones, and models presented with weak-but-opposed positions sometimes concede incorrectly, treating disagreement itself as evidence.

Production usage therefore treats debate as a quality amplifier with guards, not an oracle. Debaters should be structurally diverse — different models or different prompts — since identical twins agree too easily and correlated errors defeat the whole design. The judge should be rubric-bound and evidence-bound, scoring argument quality against cited evidence rather than vibes. And the protocol needs a stop condition: rounds beyond two or three show diminishing returns and a growing convergence-toward-consensus bias.

Where Debate Earns Its Tokens

• High-stakes judgment calls: fact-checking, risk assessments, research reviews — anywhere a silent error is expensive and a visible disagreement is a feature.
• Critique generation: a debating pair briefed as author and attacker produces sharper review comments than a single critic prompted to "find issues."
• Calibration signals: the debate transcript is a confidence artifact — a one-sided rout means something different from a knife-edge split, and routing systems should use that.

The pattern's deepest lesson mirrors the governance pieces of this blog: unverified self-assessment is the weakest signal an LLM system offers. Structured disagreement converts a model's opinion of itself into evidence a system can act on — the same move maker-checker makes for actions, applied to beliefs.