Epistemic frontiers and the distinction between causality, information, and predictability in pattern recognition

High predictive accuracy is frequently misinterpreted as evidence of causal understanding or population-level signal. Models can exploit spurious correlations, confounding, or protocol-induced artefacts, while post-hoc explanations may faithfully describe model behaviour yet remain misleading about the underlying phenomenon. We propose a framework that separates three layers of evidence: (i) causal relations in the phenomenon, (ii) population-level statistical dependence, and (iii) finite-sample, protocol-dependent predictive effects. This separation clarifies why predictive success and feature attributions do not license mechanistic interpretations without additional assumptions. Under log-loss and Bayes-risk-consistent protocols, the population predictive value of adding a feature equals the conditional mutual information, providing a principled reference for "true signal". Using controlled simulations, we illustrate that bootstrap resampling can induce persistent false positives by amplifying chance correlations, and that SHAP can assign high importance to confounded variables while remaining faithful to the fitted model. These results suggest that "feature importance" is best treated as protocol-bounded evidence, and that interpretation benefits from reporting the protocol, robustness checks, and the intended inferential scope.