Definition of done

An Intelligence change is complete when a reviewer can reproduce the decision, recover the alternatives and exclusions, challenge its stability, and see why the output recommends, downgrades, holds, escalates, or refuses. A plausible winner is not enough.

Completion by decision surface

Changed surface	Required evidence	Blocking omission
candidate record or filter	complete candidate universe, validation, exclusions, and fingerprints	only the selected candidate survives review
metric or component score	orientation, scale, missing-data behavior, boundaries, and explanation	a number changes without a semantic account
ranking policy	normalized policy, deterministic order, ties, constraints, and alternatives	hidden defaults or unstable tie-breaking
recommendation posture	support, contradiction, downgrade, hold, escalation, and refusal cases	every input produces a positive recommendation
confidence or calibration	named corpus, predicted-versus-observed behavior, and uncertainty	score language implies probability without calibration
sensitivity or regret	plausible policy and evidence perturbations, reversals, and alternative cost	recommendation hides material instability or downside
learning or adaptation	immutable prior policy, new policy identity, outcome lineage, and before/after evidence	historical decisions are rewritten in place
review artifact	evidence revision, candidate set, policy, challenge findings, authority, and round trip	report cannot reconstruct the decision path

Decision evidence loop

flowchart LR
    I["immutable input evidence"] --> C["candidate universe"]
    C --> P["named policy"]
    P --> D["decision and alternatives"]
    D --> H["challenge, sensitivity, calibration, regret"]
    H --> A{"authority and posture honest?"}
    A -->|yes| R["complete review artifact"]
    A -->|no| B["downgrade, refuse, or add evidence"]

Run the focused candidate, judgment, posture, or learning tests, then the package boundary tests for Foundation, Core, Knowledge, Runtime, and Lab contracts touched by the output. A recommendation that crosses into assay or execution planning also requires consumer proof; Intelligence tests cannot grant downstream authority.

Completion record

Preserve the input evidence revision, candidate set and exclusions, normalized policy, component results, alternatives, challenge findings, sensitivity, calibration source, regret assumptions, posture, and human or Lab handoff. State checks not run and keep the public claim inside that envelope.

Not complete

The work remains incomplete when expected rank movement is accepted without an explanation, missing evidence silently becomes a neutral score, an adverse scenario disappears from the packet, or a learned policy overwrites the decision context that produced earlier recommendations.