MechaTerp-TRACE: Post-Processing Results from Component-Ablation Experiments

MechaTerp-TRACE is an experimental framework for characterizing how support for language-model outputs is distributed across model components via controlled component ablation analysis. This archive contains the post-processing results produced from MechaTerp-TRACE component-ablation experiments. The deposited outputs contain analysis-ready representations of the experimental traces and derived measurements used to characterize component importance, prompt-level response behavior, cross-prompt reuse, and the spatial and functional organization of model components associated with named-entity generation. The archive includes post-processed outputs derived from baseline-relative ablation measurements, including component/lane-level importance information, cosine-similarity-based response traces, prompt and entity metadata, aggregation outputs, and derived comparison statistics used for analysis across prompts, entities, models, and experimental conditions.