Provenance-Attributed Inference Token (PAIT): A Protocol for Token-Level Inference Provenance and Identity-Conditioned Inference in Generative AI Systems

This document specifies the Provenance-Attributed Inference Token (PAIT) protocol, a wire-level protocol for token-level provenance attribution and identity-conditioned inference in Large Language Model (LLM) systems and other generative AI systems. PAIT defines: (1) an Agent Identity Token format that binds a globally unique agent identifier to a hierarchical authorization level by means of an asymmetric digital signature; (2) a Provenance Manifest Record format that associates each output token of an inference session with verifiable attribution data referencing training-corpus segments; (3) a Trust Telemetry Signal format for aggregated session metrics emission to external monitoring endpoints; and (4) a wire protocol state machine governing the interaction between a requesting agent and a generative AI endpoint. PAIT is intended to address the transparency obligations imposed by emerging regulatory frameworks for high-risk and general-purpose AI systems, in particular those concerning the provenance of AI-generated content at sub-document granularity.