boundary-signature: A Measurement Protocol for Per-Step Belief Monitoring in Chain-of-Thought Reasoning
Abstract
Research code, data, paper, and pre-registration trail for a token-probability measurement protocol that reads an LLM's answer distribution at each chain-of-thought step. Per-step mean entropy is a domain-portable deferral signal (AUC 0.686 MedQA, 0.664 MMLU law) under single-model, single-pass, 4-bit, black-box constraints; a systematic pre-registered sweep shows no cheap trajectory-dynamics, perturbation, verbalised-confidence, or richer-distribution feature complements it. Includes embedding-stripped cached measurements that recompute every reported number without an LLM.
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 12 June 2026