WHY THIS EXISTS

One answer can hide a point of view.

Accord asks models built by American and Chinese labs to answer independently, then shows both the shared conclusion and the views underneath it.

It does not make an answer neutral by vote. It makes agreement, dissent, and framing easier to inspect.

01

The models

Hosted model catalog

Choose a preset panel or configure your own mix.

O
GPT-5.6 LunaOpenAI
openai/gpt-5.6-luna
G
Gemini 3.7 FlashGoogle
google/gemini-3.7-flash
D
DeepSeek V4 FlashDeepSeek
deepseek/deepseek-v4-flash-0731
O
GPT-5.6 SolOpenAI
openai/gpt-5.6-sol
A
Claude Sonnet 5Anthropic
anthropic/claude-sonnet-5
D
DeepSeek V4 ProDeepSeek
deepseek/deepseek-v4-pro-0813
Z
GLM 5.3 FlashZ.ai
z-ai/glm-5.3-flash

Accord assigns a moderator suited to the selected panel. The moderator does not replace the source answers; each remains available in full.

02

The method

1

Independent answers

Every panelist receives the same question and answers without seeing the other models. This reduces anchoring and prevents one model from steering the room before synthesis.

2

Consensus synthesis

A moderator compares the completed answers for agreement, disagreement, assumptions, uncertainty, and counterarguments. It writes a synthesis while preserving material dissent.

3

Narrative comparison

Each answer is compared with the panel’s shared framing. The score looks at emphasis, omitted context, assumptions, implied causality, and value judgments—not wording or tone.

03

The color

The model pills show narrative divergence.

Shared framingDistinct framingOne-sided story

The warmer the pill, the more its underlying story differs from the room. A minority conclusion is not automatically biased, and consensus is not automatically true. The score is an AI judgment relative to this panel—not an objective fact-check.

04

Limits

What this cannot guarantee

  • Several models can repeat the same false claim or inherit similar training-data biases.
  • A moderator can omit important dissent or introduce its own framing.
  • Narrative-bias scores are comparative and probabilistic. They are not measurements of truth.
  • For medical, legal, financial, or safety-critical decisions, verify claims with qualified sources.