Methodology

The debate pipeline.

Eight stages, multiple models, human review. Built to make disagreement legible.

01

User question

Submitted from the web, API, or admin. Classified for category and ambiguity.

02

Question classification

We tag domain, identify assumptions, and detect framing bias before sending to agents.

03

Perspective selection

Agents are chosen by question type. A geopolitical question gets a Strategist. An ethics question gets a Philosopher.

04

Agent debate

Each agent argues from its strongest form, independently, without seeing each other's drafts yet.

05

Critique & challenge round

Agents read peers and write rebuttals. Weak claims get pruned; strong ones get sharpened.

06

Synthesis

A synthesis pass produces the executive summary, key takeaway, and trade-off map.

07

Source checking

Factual claims are checked against external references; flagged for review when uncertain.

08

Editorial review & publishing

Human editors review flagged debates, approve, or request regeneration before publish.

Limits & failure modes.

  • — Agents can share priors from training data. We mitigate with diverse model families and adversarial prompting.
  • — Source-checking is imperfect; confidence indicators reflect what we can and can't verify.
  • — Synthesis can soften disagreement. Editors are instructed to preserve dissent.