Validate the diagnosability defense for integrity checks

Determine whether the diagnosability defense that records raw filesystem error codes rather than only a summarized verdict can successfully distinguish failure modes requiring different responses, such as EIO and EDQUOT.

Background

The paper proposes a four-dimensional integrity taxonomy for experimental pipelines. One dimension, diagnosability, concerns whether an integrity check preserves enough evidence to support the correct response; the associated defense is to log the raw error code rather than discard it behind a binary verdict.

The identity defense subsequently encountered a known-positive failure, but the diagnosability defense had not yet been tested against the class of failure it was designed to detect. The paper therefore leaves its effectiveness as an explicit live prediction rather than presenting the defense as empirically validated.

References

The second untested defence remains an open prediction, stated as one rather than retired.

— Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay  (2608.19760 - Zhang, 20 Aug 2026) in Appendix, Section “Integrity: the four dimensions, the incident chain, and evidence decay,” paragraph “The prediction arc, in full”; see also paragraph “Defence status”