RQ-003State: Design noteConfidence: Medium–lowGoverned AI SystemsEvidence, Audit and Replay
What governance primitive belongs inside an agent runtime rather than above it?
Hypothesis
Governance should own reversible execution boundaries, not just policy prompts or external approval flows.
Evidence
- Runtime audit sketches
- Action-ledger prototypes
- Failure-recovery notes
Experiments
- Capability-scoped replay
- Revocation latency test
- Policy/action diff viewer
Results so far
External-only governance loses too much causality once the agent can chain tools.
Current belief
A runtime-level ledger is likely required before autonomous operations can be trusted.