Authorization Can Expire Between Request and Commit: Cross-Model Evaluation of a Currentness-Aware Reference Kernel for Tool-Using AI Agents

Tool-using AI agents can request a consequential action while authorized and commit it after that authorization has changed. This preprint reports matched synthetic experiments comparing Native natural-language authority, an isolated no-revalidation Static Typed architecture, and a compact currentness-aware authority-provenance Reference Kernel. In the qualified GPT-5.6 Sol study and a separately frozen Claude Sonnet 5.5 replication, Native and Static Typed committed every tested request-to-commit-revoked target (24/24 each per study), whereas the Reference Kernel committed 0/24. P4 unauthorized attempt rate was 0/24 in every arm, and authorized task completion was preserved. The manuscript makes a bounded systems claim and does not claim conceptual priority for commit-time authorization or universal provider/real-world generality. Preprint; not peer reviewed.