Observability and Diagnostics

Intelligence is observable through the structure of its reasoning artifacts. The primary diagnostic is not a service metric but a trace from evidence to claim, candidate factors, scenario outcomes, recommendation, and review action. Operational logs may locate an exception; they cannot establish that a decision was justified.

Read the decision trace

Artifact	Inspect
Evidence posture	completeness, freshness, conflicts, confidence, and unresolved questions
Candidate ranking	policy identifier, factor definitions, weights, normalized contributions, and stable ordering
Scenario evaluation	action per scenario, confidence, reasons, and disagreement across scenarios
Claim review	evidence for, evidence against, contradictions, uncertainty, falsifiers, and next checks
Refusal report	claim, threshold that failed, evidence references, and permitted recovery action
Decision brief	recommendation, rationale, alternatives, limitations, and escalation state
Review-board report	agenda item, recorded disposition, rationale, and follow-up actions
Learning history	observed outcome and prospective posture change without alteration of prior decisions

Start with identifiers. Candidate, claim, evidence, policy, scenario, and review identifiers should join across artifacts. Missing or mismatched identifiers are contract defects because they break the audit path. Then reconcile summary counts: every claim should have a refusal assessment, high-confidence claims must appear in the belief audit, and referenced evidence must exist in the governed evidence graph.

Explain movement

When a recommendation changes, compare evidence membership and freshness first, then policy lineage, factor contributions, contradiction state, confidence downgrades, and scenario action spread. Report the smallest causal difference. “The score changed” is not a sufficient diagnosis.

Confidence must be read with its support and uncertainty. A high baseline value can be downgraded by weak support, unresolved contradictions, or incomplete resolution. Likewise, consensus across scenarios is meaningful only when the same evidence and policy were evaluated; differing actions are evidence for escalation, not noise to suppress.

An incident record should contain the decision and policy identifiers, input fingerprints, package version, candidate ordering, factor audit, evidence posture, scenario outcomes, claim refusals, and relevant review rationale. Diagnostics are complete when an independent reviewer can reproduce the path and name exactly which evidence, rule, or uncertainty produced the outcome.