First Cognitive Layer: The Judgment Layer Before Execution in Deep Learning Research Workflows

We propose a deep learning research method whose unit of work is judgment: making a determination where something remains unsettled and allowing it to endure. Purpose becomes conditions that must be preserved: situations with different semantic conse quences cannot be conflated without loss, such as a variable being absent versus present but not delivered. Proposition 1 formalizes this impossibility. Each distinction is assigned a position, and the component contract o = (I, O, pre, inv, eff) specifies what combinations must preserve and what they may change. We implement the method in a workbench; ten groups of runtime records report matching inputs, weights, and predictions over 40 matched task pairs, with 3% runtime overhead, 13% storage overhead, and adaptation costs in one-off tasks. Because judgment is fixed at positions, obligations can be separated from their bearers. A public entry point accepts specifications written by humans and AI under one semantics, so humans and AI bear the same responsibility at the same position: parity of public rules and obligations, rather than parity of capability. Beyond deep learning, judgment is a method in its own right. We deliver reusable rules and distinctions that must remain separate, not a unique architecture, a mini mal operation set, or a performance advantage.This is a preprint submitted to ACM Transactions on Software Engineering and Methodology. It has not yet been accepted for publication.