Trust, audit, governance
Current authority, exact action approvals, active containment, durable effects, shared limits, and scoped evidence.
Human Oversight for Agent Fleets: Confidence Is Not an Audit Policy
A new audit-allocation paper shows that self-reported confidence can make limited human review worse than random and that tiny audit budgets can become rubber-stamping. Production oversight needs risk gates, stratified random coverage, correlation-aware learning, and a measured non-vacuity test.
Replay Is the Real Audit Log
Why "we have logs" is not an audit story, and what a hash-chained Decision Record plus canonical replay actually buys you when an incident hits.
Build an AI Agent Replay Test with the Context Pack Compiler
Run four TypeScript tests for repeatable context compilation, tool permissions, approval thresholds, and missing evidence. Includes code and replay limitations.
Approval-Mode Tiers: A Compatibility Taxonomy You Can Actually Ship
Why ad-hoc approval gates rot in production, how five v1 compatibility tiers improve operations, and why native ActionRisk remains multidimensional.
Reversibility Is the Missing Safety Primitive for AI Agents
Prevention decides whether agents may act. Reversibility lets them survive being wrong through reversal contracts, compensation, and blast-radius caps.