Beyond Prediction -- Structuring Epistemic Integrity in Artificial Reasoning Systems