When Explanations Write Back: Mechanistic Provenance for Language Models

Abstract

This paper introduces mechanistic provenance: the requirement to distinguish the mechanism version that generated an explanation from the mechanism version later used to validate it after explanation uptake. It derives the identity s = q + C - L, introduces the Baseline Write Barrier, and reports controlled Qwen experiments together with a preregistered SST-2-derived Phi-3.5-mini-instruct replication. Phi meets the individual-model PR-002 criterion; the stronger preregistered cross-architecture gate is not met. Separate prospective experiments do not support privileged introspection. Version 1.1 is a readability and reproducibility update. The manuscript has been edited for a human ML / mechanistic-interpretability audience without changing the numerical results or confirmatory claim boundary. A later archival audit also recovered the retained PCA-006 through PCA-012 developmental execution layer: 13 Hugging Face Jobs executions and 672 parsed family records. Those records remain developmental/supporting evidence and are pinned to Git commit f92a32e356d261c1f17d6a77d615812de2b101a4. No missing observations were synthesized or back-filled.