Agreement
Do independent models repeat the same core claim?
Repeated claims are not automatically true, but agreement is a useful signal when the prompt includes enough context and the claim can be checked.
Benchmarks
What AI Debate measures before trusting a consensus
AI Debate is not a claim that four models make an answer correct. It is a workflow for surfacing agreement, preserving disagreement, and showing what still needs verification.
Disagreement
Which assumptions produce different recommendations?
Disagreement is preserved in the result so users can see whether a recommendation depends on risk tolerance, missing data, or model-specific framing.
Latency
How long did the panel take to respond?
Provider response time is part of product quality. A slow but more complete answer may be useful for planning; a quick answer may be better for lightweight drafting.
Human review
What changed after comparing the answers?
The most useful case studies show what a single model missed, what the panel agreed on, and which claims still require outside sources.
Current benchmark posture
The public product reports provider answers, cross-review notes, contradictions, and the final synthesis. The next benchmark layer should publish fixed prompt suites with model versions, response time, consensus changes, and human-reviewed error notes.