Abstract Autoregressive large language models (LLMs) are fundamentally susceptible to semantic drift, hallucination, and cumulative error propagation due to the unconstrained nature of token-by-token generation. This paper introduces the Invariant Agency Protocol (IAP), a novel computational architecture that mitigates hallucination by separating internal exploratory variance from final token crystallization. Drawing from the principles of Bidirectional Constraint Closure (BCC) and Constraint Topology Medicine (CTM), the protocol treats intermediate transformer layers as a bounded space of possible histories (W). By establishing an internal "invariant agency"—a transient, high-variance latent state—the model is permitted to explore divergent probabilistic paths. However, a strict mathematical projection operator (P_C) applies the system's core invariants prior to the decoding head. Trajectories that violate these constraints are allowed to naturally dissipate within the hidden state layers, ensuring that only ground-truth, constraint-aligned data collapses into the actualized output string (H_{actual})