TRACE Schema Enables Auditable Reasoning for AI Agents

The paper presents TRACE, a new framework designed to bring auditable reasoning to AI agents. It consists of a typed, versioned schema for recording an agent's thought process, a standard procedure for generating these records, and an operational rule that mandates a record for any durable state change. The authors contend that large language models, by their nature, compute associations rather than formal reasoning, and existing methods like chain-of-thought inherit these limitations. TRACE addresses this by providing specific fields and tests within its `TraceRecord` structure, along with an eight-stage reference writer. This enables a "gate-first" measurement approach and supports various consumers for memory admission, plan gating, and verdict reuse, all aimed at making agent decisions more auditable. The framework also defines a clear contract between the record and its consumers, establishing guarantees and responsibilities. Illustrative examples, such as tracing a music lesson argument and a flood search-and-rescue scenario, demonstrate how TRACE can separate associative thinking from intervention and prescription, and how it can defer commitments when a predictive model's confidence is contradicted by its own support scores. The paper emphasizes that its contribution is the schema and its contract, not empirical performance claims, with closed-loop evaluation reserved for future work.