An Didrah intelligent-agent answer is not a paragraph. It is several pieces that mislead if they are read apart. This guide fixes the reading order so the autonomy ladder does not climb past “propose” unless the pieces are complete.
Reading order
- Claim: one sentence with a number. If it has no number, ask the question again.
- Evidence: the rows the claim was built from. If the claim does not stick to the evidence, reject the claim.
- Confidence level and its basis. “Medium” means take a reversible decision, not a large bet.
- Recommended action, then risk and data gaps. Do not sign the action before reading the risk.
When the answer should stop
If an AI-governance policy blocks a figure from being published, the answer should stay empty, not get softened. Human control means rejecting an answer, not editing it until it is “presentable.” The difference looks small and matters a great deal for trust in the system.