DUO Framework: Discover, Understand, Output

Generative AI can produce a confident-looking deliverable before the user has enough context to judge it. DUO addresses that failure mode by separating context building from decision-making and production.

It is not a fixed three-step prompt. It is an iterative conversation loop for work where evidence, situational knowledge, trade-offs, and human judgement materially affect the result.

In Discover, the agent expands context and identifies sources, unknowns, risks, options, and questions. In Understand, the human adds situational context, probes the evidence, challenges assumptions, and makes defensible decisions. Output begins only when the human authorizes production from the agreed context.

A gap found during Understand returns the work to Discover. A gap found during Output returns it to Discover or Understand. The human decides when the artifact is ready for review and when the work is complete.

• Discover: expand the context before optimizing for an answer.
• Understand: apply human knowledge, judgement, and direction.
• Output: produce a fit-for-purpose artifact under human governance.

For some exploratory work, the human may temporarily withhold a preferred conclusion or artifact so the agent maps the domain before converging. DUO calls this controlled goal withholding.

Material safety, privacy, legal, regulatory, professional, jurisdictional, confidentiality, evidence, time, and budget constraints must still be disclosed. Goal withholding is optional and experimental; it should not be presented as a proven method for preventing bias.

A DUO output should identify its audience, purpose, evidence standard, assumptions, exclusions, human decisions, unresolved items, checks, and review state. Consequential work still requires qualified review and primary-source verification.

DUO version 0.2 is a theoretically grounded public draft, not an empirically validated named model. The living Gist includes the full specification, quick reference, worked examples, evidence notes, and a validation protocol designed to test the framework without confusing fluency or user confidence with accuracy and decision quality.