Hierarchy as Constraint on Revision Validity: A Conceptual Account of Why Layered Adaptive Systems Learn When Flat Ones Merely React

This paper proposes that hierarchy in adaptive systems functions to direct error-driven revision by constraining the revision distribution toward low entropy and high validity simultaneously — away from both randomness and rigidity. This claim is formalized as the Hierarchical Constraint Principle (HCP). The paper makes five operationally distinct contributions. First, it provides a three-criterion definition of functional hierarchy — temporal separation of layers, top-down constraint on revision priors, and class-level content — with explicit satisfaction thresholds that enable principled adjudication of ambiguous cases, including a formal analysis of why standard end-to-end trained deep networks satisfy Criterion 1 and 3 but are borderline non-satisfying on Criterion 2 at threshold, with that weakness predicting a specific and independently testable generalization limitation. Second, it introduces a two-dimensional operationalization of revision quality: revision entropy H(E) measuring the distribution's concentration or dispersion, and constraint validity V(E) measuring whether concentrated choices are structurally directed toward viable options. This two-dimensional framework is required to distinguish healthy directed revision (low entropy, high validity) from the two failure modes the HCP predicts: collapse (high entropy, low validity) and freeze (low entropy, low validity). Third, the paper formalizes two distinct failure modes — hierarchical collapse, in which higher layers become unavailable under load producing random revision, and hierarchical freeze, in which higher layers remain present but lock under threat producing rigid revision. Both are failures of hierarchical constraint but in opposite directions. Both reduce adaptive coherence. Both predict different behavioral signatures and require different interventions. The distinction resolves a latent tension in prior accounts that described both phenomena under the single term 'compression.' Fourth, it provides a Monte Carlo power analysis demonstrating that the predicted entropy difference between prefrontal-lesion and control groups (Cohen's d = 0.65–0.80) is detectable at β = .80 with n = 28–34 per group, and extends the experimental protocol to include a freeze condition alongside the collapse condition, enabling complete testing of the HCP's distinguishing predictions. Fifth, it positions the HCP's two failure modes within Bennett's (2026) weakness framework, which provides formal mathematical grounding: survival under novelty is monotone in unobserved buffer size, and hierarchical constraint accumulation is the mechanism by which adaptive systems maintain large valid buffers. Collapse reduces buffer validity. Freeze reduces buffer size. Both destroy the genuine weakness that hierarchical constraint provides. The HCP's failure prediction — that higher-layer disruption increases revision entropy specifically — is not made by hierarchical reinforcement learning, perceptual control theory, Simon's near-decomposability, or Dennett's Tower. The freeze prediction — that identity threat produces low-entropy concentrated revision on structurally invalid options — adds a further distinguishing prediction that capacity accounts cannot explain. Keywords: hierarchy, directed learning, structural constraint, revision validity, revision entropy, hierarchical collapse, hierarchical freeze, functional hierarchy, hierarchical reinforcement learning, perceptual control theory, predictive processing, metacognition, weakness maximization