Architecture ยท audit record
A spine for accountable AI.
The output is not the record. The record is what lets the output be judged.
Skill Core captures the claim, binds the evidence, preserves the authority boundary, observes the outcome stream, and updates trust when reality answers back.
Pattern language
Eight fields, one audit-readable event.
The architecture is AI-native but not AI-limited. Once a claim has evidence, context, authority, and an outcome trail, it can be scored wherever reality later supplies a verdict.
- 01assertedClaim
The model answer, agent action, decision, prediction, or recommendation.
- 02boundEvidence
Sources, observations, retrieved context, data, tool calls, and reasoning traces.
- 03scopeAuthority boundary
Role, permission, policy, escalation, and the point where review becomes mandatory.
- 04contextContext
Model, version, source set, date, prompt state, environment, and operating conditions.
- 05observedOutcome stream
The later measurements and events that make the original claim checkable.
- 06resolvedReality validation
The comparison between what the AI claimed and what actually happened.
- 07learnedTrust update
The change in model, source, or workflow weight after evidence meets reality.
- 08typedFailure semantics
The explicit reason a claim failed, exceeded scope, degraded, or lost support.