Concepts

Understand the deliberation model, consensus scoring, and how four models produce a single structured verdict.

The deliberation model

Decision Memos doesn't just query four models and concatenate their answers. It runs a structured deliberation:

  • Parallel query — All four advisors receive the same question and context simultaneously.
  • Individual analysis — Each advisor responds independently, shaped by their thinking style.
  • Synthesis — A synthesis layer analyses all four responses, identifies agreements and divergences, and produces the unified verdict.
Note
The synthesis is not a fifth model. It uses one of the advisors (The Strategist) with a meta-prompt that instructs it to objectively analyse all four perspectives as a neutral moderator.

Consensus scoring

The verdict includes a consensus level that tells you how aligned the panel is:

  • Strong — "The panel is united." All four advisors broadly agree on the recommendation.
  • Moderate — "The panel largely agrees, with some differing perspectives." Most agree, but there are meaningful divergences.
  • Weak — "The panel is divided." Significant disagreement — the divergences are as important as the recommendation.

A weak consensus isn't a failure — it means the question is genuinely contentious and deserves careful consideration. The divergence map tells you exactly where and why the advisors disagree.

The Briefing

Before deliberation begins, the optional Briefing step generates dynamic follow-up questions tailored to your specific decision. This extracts the context the panel needs for a high-quality answer: constraints, priorities, timeline, scale, and domain-specific factors.

The Briefing questions are AI-generated based on your question — not hardcoded. They adapt to the domain and complexity of your decision.

Criteria

Criteria are the dimensions you want the panel to evaluate. They shape how the advisors frame their analysis. Examples:

  • Scalability — "Can this handle 10x growth?"
  • Team readiness — "Can our team operate this?"
  • Time to market — "How fast can we ship?"

If no criteria are provided, the advisors will infer the most relevant dimensions from the question itself.

Personas vs transparent mode

By default, Decision Memos uses persona mode — advisors are labelled by archetype ("The Strategist", "The Analyst", etc.) rather than model name. This is designed for consumer-facing products where the underlying model is an implementation detail.

In transparent mode, real model names are shown (GPT-4o, Claude 3.5 Sonnet, Grok-3, Gemini 2.0). Use this for developer-facing tools or when model provenance matters.

The Decision Memo

The Decision Memo is the structured output artifact. It's a typed JSON object with a strict schema, designed to be rendered, stored, exported, and shared. It contains:

  • Metadata — Question, timestamp, model versions, and criteria used.
  • Verdict — The synthesised recommendation, consensus level, and display string.
  • Perspectives — Each advisor's individual response, labelled with persona or model name.
  • Consensus map — Agreements and divergences extracted from the synthesis.
  • Risks — Identified risks with mitigations.
  • Next steps — Prioritised implementation actions.