The Constitutional AI Protocol (CAP) for Agentic AI Systems

This document specifies the Constitutional AI Protocol (CAP), a kernel-enforced prohibition architecture for agentic AI systems. CAP defines a three-tier prohibition model: Tier 0 absolute prohibitions derived from near-universal treaty consensus, Tier 1 jurisdiction-specific prohibitions declared by operators and verified by auditors, and Tier 2 voluntary operator ethical standards. CAP evaluates every action request twice -- once on the AI agent's action before Cedar policy evaluation, and once on the human principal's decision before kernel execution -- ensuring that neither agents nor human principals can authorize absolutely prohibited actions. This document also specifies Jurisdictional Conflict Resolution (JCR), the mechanism by which the kernel surfaces irreconcilable jurisdictional conflicts to human principals via the Human Escalation Mechanism [I-D.sato-soos-hem].