aip

The Agent Integrity Protocol (AIP) is a real-time analysis framework designed to evaluate the reasoning (thinking blocks) of AI agents before they execute actions. It extracts thinking content from LLM responses (supporting Anthropic, OpenAI, Google, and a regex fallback), evaluates it against an Alignment Card using an analysis LLM, and returns integrity verdicts (clear, review_needed, boundary_violation) to enable intervention between turns. The protocol includes provider adapters, session windowing for drift detection, and a conscience prompt system that injects boundary and fear values as evaluation criteria. It is positioned as a complement to the retroactive Agent Alignment Protocol (AAP), providing real-time detection rather than post-hoc verification. The repository includes Python and TypeScript SDKs, a specification, examples (basic checks, gateway integration, adversarial scenarios), and claims alignment with standards like ISO/IEC 42001, IEEE 7001, and the EU AI Act. It explicitly states limitations: it does not prevent injection, does not interrupt streams, and relies on models exposing thinking blocks.