Epistemic Relevance Is Not Invariant Across Data Processing Pipelines: A Structural Account of Representational Transformation, Evidence Formation, and Objectivity in Contemporary Science

Contemporary empirical science increasingly relies on complex data-processing pipelines in which observations are not used directly, but are transformed through multiple stages of measurement, operationalization, modelling, inference, and interpretation. While these procedures are often treated as methodologically neutral, this paper argues that they are epistemically formative: they actively shape what can count as evidence in scientific reasoning. The central claim is that epistemic relevance is not invariant across these representational transformations. What appears as relevant information at one stage of analysis may be altered, diminished, or rendered inaccessible at another. This implies that relevance is not a stable property of data, but is continuously reconstructed within specific representational and modelling frameworks. On this basis, the paper develops a structural account of data-driven science in which evidence is not a fixed input to inference, but the outcome of representational operations. Models and preprocessing steps do not merely process information; they define the space of possible evidential relations by determining which differences in the data are visible and which are ignored. This perspective has direct implications for core epistemological concepts. First, evidence must be understood as transformation-dependent rather than as a neutral intermediary between theory and world. Second, scientific objectivity is better interpreted as the transparency and controllability of representational processes rather than as direct correspondence to raw data. Third, underdetermination is strengthened, since not only theories but also data structures depend on representational choices. The paper therefore proposes a shift from a data-centred to a transformation-centred epistemology of science, in which the key question is not how theories match given data, but how different representational systems generate and stabilize comparable forms of evidence.