Multi-agent language-model systems pass reasoning between agents through messages that are difficult to audit after the fact. This paper presents Reasonledger, an open specification that records reasoning transfer with explicit provenance, a three-way epistemic status, and per-step verification, together with a reference checker that detects five classes of cross-hop consistency failure. An evaluation with a decision rule documented before data collection tested whether the notation improves failure detection over information-matched JSON. No advantage was found, and the LLM arm as run could not have detected one: the capable model flagged every trace, clean or corrupted, in every format, a degenerate operating point, while the weak model flagged almost nothing. Under the fixed rule, the notation's value is standardization rather than detection. An exploratory follow-up produced the central observation. On the capable model and thirty clean Reasonledger traces, a prompt that describes each failure class flagged a fabricated failure on every trace (30/30). Naming the classes without descriptions cut the rate to 8/30. Defining each class and stating what does not count as a failure removed it entirely, 0/30, with detection at 14/15, while a generic instruction to favor precision also reached 0/30 at a detection cost, 5/15. The strict prompt had been written after those false positives were observed, against the same thirty traces, so this experiment alone could not distinguish a generalizable effect from overfitting. A second experiment addressed that directly. All prompt texts were frozen by hash, the design was anchored by a third-party timestamp before any live auditor call with interpretation rules fixed before data, and the auditor ran on sixty same-generator, high-overlap held-out clean traces and fifteen fresh injected failures. The pattern reproduced: descriptive 60/60; minimal 16/60; strict 0/60 with detection 13/15 at class level and 7/15 under exact localization; precision 0/60 with detection 5/15 and localization 4/15; and the previously missing cell, the minimal prompt with the caution text spliced in verbatim, 0/60 with detection 7/15 and localization 5/15. Every localization miss in the run occurs at the correct hop and differs only in item naming, characterized in the appendix. Which prompt property drives the false-positive rate remains open; each condition carries one wording and the design contains no paraphrase control. The specification, checker, harness, both study protocols, and raw results are released in the reasonledger repository (github.com/mpaiello/reasonledger, release commit 781ccd7) and archived at https://doi.org/10.5281/zenodo.21277654. Limits are stated plainly: one model, one trace generator, single-field injections at one failure per trace, and a strict prompt that encodes knowledge of the error classes.
Michael Patrick Aiello (Thu,) studied this question.