Skip to content
Danphe Council

Draft — pending legal review

This document is a working draft. It has not been reviewed by a lawyer and is not yet in force. Bracketed text is still to be completed.

AI system card

Draft version, November 2026

What it does

Danphe Council answers questions about documents by combining several large language models from different companies. Models answer independently; their statements are extracted and checked by models from other families against the sources; an editor writes the final answer from the statements that held up. Every statement gets a label (for example “Models agree”, “Not settled”, “Not supported”, “Opinion”).

Intended use

  • Research, review and analysis of documents where a person checks the result: contracts, papers, reports, filings and student work.
  • Finding disagreement and unsupported statements faster.

Not intended for

  • Fully automated decisions about people (hiring, credit, medical, legal outcomes).
  • Use as professional advice without review by a qualified person.

Models

The Free plan uses a fixed council of four models from different companies. Paid plans use a flagship council (Claude Opus, GPT-6.1 Sol, Grok 4.7, Kimi K3) that customers can configure. Models are accessed through Amazon Bedrock and Microsoft Azure.

Evaluation

On a preregistered locked test of 483 unseen questions (November 2026), Smart mode scored 85.1% (95% CI 81.6–88.1%), statistically indistinguishable from Claude Opus 5.5 alone (83.4%, p = 0.13) and significantly better than GPT-6.1 Sol alone (81.0%, p = 0.002). Details and limits are on the accuracy page.

Known limits

  • When all models share a mistake, cross-checking cannot catch it.
  • It costs more and is slower than a single model.
  • On long-form answers in development testing, a single model (GPT-6.1 Sol) had higher statement precision.
  • “Confirmed” labels require a workspace calibration with human-labelled data; without it, the strongest label is “Models agree”.
  • Documents can contain instructions aimed at the models (prompt injection). Inputs are isolated and outputs are checked, but no defence is complete.

Human oversight

Users can send any statement for human review. Every answer has an audit record that shows which models ran, what each said and how each statement was checked.

Data

Customer Content is not used to train models. See the Privacy policy.