LLM◆ AI-generated · Sourced
TRACE: A Typed Schema for Auditable Agent Reasoning Records
TL;DR
The arXiv paper TRACE proposes a typed, versioned schema for reasoning traces plus one operating rule—no durable state change without a record—so an agent's commitments become auditable, reusable, and accountable.
Bottom line
TRACE (Typed Reasoning And Commitment Evidence) tackles a real pain point: agents make decisions but can't show 'on what basis'. It turns the reasoning process into recordable, auditable, reusable structured evidence.
Core claims
- The paper argues bluntly that reasoning is not in the language model—the autoregressive mechanism natively computes association; chain-of-thought and RL inherit its limits; the machinery of classical reasoning theory is absent.
- TRACE fills the gap with 'fields plus tests': a TraceRecord structure, an eight-stage writing procedure, a gate-first measurement regime, and the TRACE-Bench protocol.
One hard rule
No durable state change without a record. This turns the schema from a passive document into an operating interface: what a record guarantees, a consumer must honor.
For teams doing agent orchestration that need compliance auditing and traceable decisions, this record-consumer contract is worth a look.
Sources (compliance trail)
https://arxiv.org/abs/2607.12480