Blind test: deliberative multi-perspective AI system loses to a single frontier model call

Wonderful_Bite1139 · reddit · 2026-10-03

The author built TRAMA 2.0, a Multi-Perspective Deliberative System (MPDS) on top of a frontier model: a single entity that decides how much to think, runs three internal "faculties" over each question, then makes a final decision, with long-term memory, web search, and full logging.

A blind test pitted four systems on the same questions (e.g., how a town should spend €200k once: LED street lighting vs two years of free after-school programs), judged blind by Claude Opus 5.5 against six pre-set criteria:

Key lessons: search mattered more than deliberation, and three perspectives on the same model failed to catch an arithmetic mistake. Limitations are openly stated (one question, judge from yet another lab), with a 3–5 question follow-up planned.

Original post →

More from Research

Research channel →