Clinical Information Fidelity (CIF): A Task-Defined Framework for Evaluating Information Preservation Across Clinical Transformations

Clinical information increasingly passes through automated transformations such as speech recognition, summarization, translation, ambient documentation, and generative reformulation. Existing evaluation methods address lexical accuracy, semantic similarity, clinical entity overlap, factual discrepancies, and document quality, but do not necessarily make explicit which clinical information a particular transformation was required to preserve. This work proposes Clinical Information Fidelity (CIF), a task-defined framework in which transformation-specific preservation requirements are specified independently of the candidate output and evaluated separately from the mechanism used to assess preservation. For source information (X) and transformation (T), a preservation specification (C_T(X)) defines the clinical information units required to survive. Baseline CIF represents the proportion of those requirements satisfied by output (Y). CIF is not proposed as a semantic similarity algorithm, hallucination metric, correctness score, completeness score, or general measure of clinical document quality. It formalizes one specific evaluation dimension: preservation of required clinical information across transformation. Status: Conceptual framework. Not empirically or clinically validated.