Mapping the Verification Space
Verification for autonomous systems spans several distinct questions. One question concerns permitted actions. Another concerns information flow. Others concern provenance, execution context, or reproducibility. Serious system design keeps these claims separate and connects them through explicit interfaces.
Verification starts by naming the property
Action admissibility
The runtime decides whether an effect is permitted under the active policy state. Finite-state monitors fit temporal constraints and action sequences with explicit event semantics.
Information flow
The system tracks provenance and sensitivity across data movement. Local acceptance rules need a stated bridge to any broader non-interference claim.
Permission state
Calls and data operations reference a defined authorization state. Epoch semantics fix the policy snapshot used for each decision and give revocation a testable meaning.
Artifact provenance
Certificates bind runtime records to the policy and implementation artifacts used for evaluation. Signatures and supply-chain records support independent identity checks.
Execution context
Trusted hardware or cryptographic proofs address claims about the environment in which computation occurred. Those mechanisms answer a different question from runtime policy enforcement.
Replay consistency
Recorded traces support reconstruction of selected executions. Deterministic profiles improve incident analysis and implementation comparison.
Enforcement, attestation, and audit answer different questions
Runtime enforcement decides whether an effect proceeds. Attestation records facts about software or execution context. Audit reconstructs the basis for a prior decision. Each function keeps its own guarantee even under one product surface.
Claim boundaries matter most at interfaces. Binary attestation addresses software identity or execution context. Certificate validity addresses evidence binding to a named policy. Replay addresses reconstruction under a declared trace model. Each claim has a distinct evidentiary target.
The runtime boundary needs four stable identities
Policy identity
Names the specification governing the current decision.
State identity
Names the monitor and authorization state used at the decision point.
Artifact identity
Names the implementation and proof artifacts associated with the verdict.
Trace identity
Names the recorded execution fragment used for replay and external review.
Evaluation follows the claim boundary
Monitor evaluation tests semantic agreement and execution coverage. Information-flow evaluation tests observer views under controlled trace variation. Permission evaluation tests state transitions and revocation semantics. Provenance evaluation tests identity binding. Replay evaluation tests reconstruction fidelity.
| Claim | Primary test | Failure signal |
|---|---|---|
| Action policy | Reference semantics against runtime verdict | Verdict disagreement |
| Information flow | Low-view comparison across controlled traces | Observer-view divergence |
| Permission state | Epoch and revocation transition tests | State mismatch |
| Evidence record | Independent certificate verification | Broken identity binding |
| Replay | Trace reconstruction under declared profile | Reconstruction divergence |
Keep guarantees compositional
Verification infrastructure gains credibility through narrow claims with explicit dependencies. Runtime monitors establish action-policy properties. Provenance records establish artifact identity. Attestation mechanisms establish facts about execution context. The system record links those claims through named artifacts and stated assumptions.
This separation supports replacement of individual components and preserves the meaning of evidence across implementation changes. It also provides external evaluators with a tractable target for independent checking.