Verification as a Control Primitive for Frontier AI
A systems perspective on enforceable authority in high-capability environments.
A verification-gated boundary.
Conceptual reference model. The report is deliberately mechanism-neutral.
actionMachine capability
evidenceAuthority claim
decisionAllow / deny
executionConsequential state
CONTROL THESIS Capability and authority are separate properties. Consequential execution should remain conditional on verifiable evidence.
Abstract
Frontier AI safety is often discussed as a problem of model behavior, policy compliance or organizational governance. Those layers matter, but they share a structural limitation: they can fail at the moment a capable system crosses from reasoning into consequential action.
This report examines verification as a complementary control primitive at that boundary. The argument is conditional and architectural. If high-consequence actions require evidence that is independent of the digital claim being made, and if policy can deny execution when that evidence is absent or invalid, then some failure pathways can be converted from open-ended trust problems into bounded authorization problems.
The report develops a conceptual hazard model, a general reference architecture and an evaluation agenda. It does not claim to solve alignment, eliminate catastrophic risk or validate any specific commercial mechanism. The central conclusion is deliberately narrow: stronger intelligence increases the value of externally verifiable authority.
A systems thesis, not a product specification.
The publication separates the public research argument from protected implementation. It addresses control objectives, system boundaries, failure modes and testable questions without disclosing proprietary mechanisms, confidential architecture or deployment specifications.
- IncludedConceptual hazard model
- IncludedMechanism-neutral reference architecture
- IncludedEvaluation and research agenda
- ExcludedProprietary implementation details
Cite this report
Braun, N. (2026). Verification as a Control Primitive for Frontier AI (Version 1.0). Zenodo. https://doi.org/10.5281/zenodo.21454914