Documentation standards

Public recommendation language must expose the reasoning path and its authority boundary. A polished conclusion without its candidate universe, evidence revision, policy, alternatives, instability, and refusal conditions is not a reviewable Intelligence result.

Recommendation vocabulary

Term	Required context	Must not imply
score	component definition, orientation, scale, missingness, and policy version	calibrated probability or biological truth
rank	candidate universe, exclusions, tie-breaking, constraints, and alternatives	absolute merit outside the compared set
confidence	calibration method, corpus, interval or category meaning, and known drift	certainty because a value is numerically high
recommend	evidence revision, policy, challenge results, regret, and authority	autonomous approval or guaranteed outcome
downgrade	adverse evidence or sensitivity condition that weakened posture	implementation failure
hold	missing decision-critical evidence and the evidence needed to resume	a weak positive recommendation
refuse	explicit violated precondition or unacceptable uncertainty	missing output or exception

Reviewable narrative

flowchart LR
    Q["decision question"] --> E["evidence and revision"]
    E --> C["candidates and exclusions"]
    C --> P["policy and component results"]
    P --> A["alternatives and challenges"]
    A --> R["bounded posture and authority"]

Examples finish at a decision brief or review packet, not an isolated score. They show at least one alternative, one adverse or contradictory condition, and the reason the final posture is no stronger. When a small plausible policy change reverses the action, the instability belongs beside the recommendation.

Source and authority

Reference Knowledge evidence rather than copying it into a new truth record. Name the policy and configuration that transformed evidence into a decision. Keep Lab feasibility and Runtime execution as downstream evidence. Intelligence may propose or refuse an action; it does not authorize laboratory work or declare execution successful.

Updating a public result

A new evidence revision or learned policy produces a new decision record. Compare old and new outcomes and explain the changed inputs or policy. Do not rewrite earlier rationale as though the later information had always been available.

Known limitations defines the inference ceiling, and definition of done lists the evidence required after a decision-path change.