The MIT Research Behind Delibera's Architecture

The Research Question

Can you build an AI system whose output is more accurate than any of its constituent models — not by averaging, not by ensemble voting, but by adversarial deliberation?

The honest answer, before we started, was: we didn't know.

The Setup

Three independent frontier models, each given the same matter but a distinct analytical mandate. Structured rounds of cross-critique. Mandatory dissent. Gap-driven search for external verification.

We evaluated against the Phare Hallucination Benchmark and a proprietary adversarial misinformation suite. The results, after iteration, were consistent and surprising.

The Result

• 82.6% hallucination resistance on Phare (350 samples)
• 13.1% hallucination rate on adversarial misinformation (791 samples) vs.

22.3% for the next-best model — a 41% reduction

• The only system tested that expressed uncertainty (26% of challenging cases)

The interesting finding wasn't just lower hallucination — it was that the improvement came from the *disagreement* between models, not from any single model getting smarter. The dissent produced the accuracy.

What it means for you

Two things:

• When three independent models agree, the answer is very likely correct.
• When they disagree, that disagreement is itself valuable — and it's usually

what you need to make a better decision.

Delibera preserves both: the consensus where it exists, and the disagreement where it doesn't.