Skip to content
Danphe Council

Research

Research notes

How we test the council and what the numbers say, including where they don't flatter us. Written by Danphe Intelligence Lab.

Unseen questions tested
483
five domains
Smart mode accuracy
85.1%
locked test
Latest7 min readEvaluationPreregistered

The locked accuracy test: what a council of models does and doesn't add

We froze our hypotheses before running 483 unseen questions across five domains, plus 100 long-form questions. Three of four hypotheses passed. Here is everything, including the one that failed.

Read the note
Smart mode
85.1%
Claude Opus 5.5 alone
83.4%
GPT-6.1 Sol alone
81.0%
hypotheses passed
3 of 4

How we test

  • Preregistered

    Hypotheses, rules and data are frozen before a locked test runs. Nothing is re-run afterwards.

  • Honest comparisons

    We compare against the best single model, and say where one model alone is the better tool.

  • Numbers you can trace

    Sample sizes, intervals and dates are published with every result.