Dual-Model Consensus: How AI Disagreement Is Resolved

The full executive answer

Neither wins silently — that is the design principle. On the board-tier modules, after the primary Claude analysis passes all its gates, the same prompt goes to GPT-4o, a model from a completely different company trained on different data. Our code then diffs the two structured outputs deterministically — numbers, categories, list sizes — and deliberately ignores narrative wording, because two models never phrase a paragraph identically and that is not a material disagreement. The result is an agreement score and a list of specific divergences, and every consensus run is persisted to its own table.

When agreement drops to eighty percent or below, a third model call adjudicates: for each divergence it rules which side is more defensible to a board — primary, second, or "neither", where the disagreement itself is the finding — with a written reason a director would accept. Critically, the ruling is attached and displayed; the primary output is never silently rewritten. Material disagreement is flagged for human review, never resolved behind the curtain.

Two honest footnotes. The consensus pass is skipped when the primary already fell back to OpenAI — there is no value in diffing a model against itself. And it is an operational cost lever: it can be disabled by configuration under cost pressure, which is a deliberate, documented trade-off rather than an accident. Today it covers ten module families including the scorecard, risk register, compliance, ESG, security posture, roadmap, M&A diligence and boardroom mode.