Embedded evaluators and Stop & Review: comparing two approaches to AI governance

As AI systems gain the ability to use tools and carry out multi-step tasks, governance needs more than high-level principles. One proposed approach is to embed evaluators throughout development and deployment, establish capability thresholds and pause operations when risk indicators require further review. This article compares those ideas with the OAGI framework.

Embedded evaluators can assess behaviour and identify deviations. Capability-linked checkpoints require a system to pass an evaluation before its permissions or autonomy increase. A Stop & Review procedure can halt a workflow and send an unexpected situation to a human reviewer. Persistent records can support audits by showing what actions were taken and which rules applied. Each mechanism needs practical definitions and tests; a label alone is not a safeguard.

Governance also depends on coordination across companies, governments and independent researchers. A workable framework should define who can authorize exceptions, what information is disclosed and how disagreements are handled. International coordination is difficult, but high-impact systems do not stop at national borders.

OAGI presents a conceptual structure for staged development and oversight, while other approaches emphasize principles applied to model behaviour. Neither should be treated as proven simply because its terminology sounds comprehensive. The comparison should focus on measurable reliability, independent auditing, traceability and the ability to stop a system before a mistake causes harm. Effective AI governance requires enforceable controls and clear human accountability, not only commitments in policy documents.