The Computational Architecture of Agentive Control: A Dynamical Implementation of the Irreducible Agency Invariant (IAI)

The emergence of behavior that appears internally initiated has generated deep ambiguity across neuroscience, cognitive science, and artificial intelligence. Recurrent systems often display trajectories that resemble spontaneous action or self-directed reasoning, yet similar patterns can arise from reactive dynamics, learned policies, or inherited structure. The Irreducible Agency Invariant (IAI) resolves this ambiguity by providing a formal criterion for identifying when a system genuinely authors its own state transitions. Here I develop the first full computational architecture implementing the IAI inside concrete recurrent systems. The framework decomposes internal and external generators, introduces a mechanistic gain-modulation variable corresponding to agentive control, and defines a counterfactual internal-generator baseline needed to distinguish endogenous influence from impressive (reactive) dynamics. The architecture instantiates the four structural conditions of the IAI including divergence, persistence, spectral coherence, and internal origination, within a unified dynamical model. Divergence and persistence identify sustained departures from the internal baseline; spectral coherence ensures the transition remains within the system’s own stable manifold; and internal origination is formalized through the causal sensitivity test ∂hₜ₊₁/∂Eₜ ≠ 0, which determines whether the system’s evolution depends on its own modulatory signal. Together these conditions operationalize the expressive–impressive distinction, yielding a falsifiable and coordinate-invariant test for endogenous authorship. To demonstrate feasibility, the paper constructs a minimal recurrent prototype with separable internal and external generators, implements a logistic gain-control pathway, and evaluates divergence and causal sensitivity across delayed-response, context-dependent, and conflict-structured tasks. The results show how endogenous gain modulation can reshape flow geometry, stabilize latent commitments, and produce expressive transitions irreducible to reactive computation. By embedding the IAI in an explicit dynamical substrate, this work provides the first principled, operational method for detecting internally originated control in recurrent architectures. The framework unifies insights from phenomenology, global workspace theory, predictive processing, and control theory, offering a computational foundation for distinguishing genuine agency from mimicry in both biological and artificial systems.