Routing Without Hallucination : A Structural Grammar for AI in Institutional Systems

Routing Without Hallucination: A Structural Grammar for AI in Institutional Systems proposes a minimal, domain-agnostic grammar for representing how rule-dense institutions actually operate. Rather than treating institutional reasoning as narrative interpretation, the paper formalizes four recurring structural primitives—rails (corridor families), subjects (attribution containers), conversions (authorized reconfigurations), and triggers (state-change conditions)—as a “slot architecture” that governs how admissible inputs are routed into obligations, liabilities, or entitlements. The 4-3-3-2 grammar is not a taxonomy of tax law. It is an axiomatic extraction of the operational logic shared by rule-dense systems such as taxation, immigration, benefits administration, and platform compliance. Tax law serves as a high-pressure laboratory in which these primitives become visible: when narrative interpretation is stripped away, institutional work reduces to corridor selection, subject attribution, constrained transformation, and trigger-based switching. The paper positions this grammar as a protocol layer for AI in institutional settings. Instead of allowing generative systems to produce unconstrained prose, the model requires every output to be expressed as a routing trace: an auditable, replayable, and contestable intermediate artifact that records rail assignment, subject attribution, conversions applied, and triggers invoked. This directly addresses the “hallucination” problem by preventing narrative substitution—coherent text that is unmoored from institutional structure. Methodologically, the grammar converts tacit professional expertise into an explicit, teachable, transferable, and machine-legible form. It enables: Teachability: novices learn to read systems by filling structural slots rather than memorizing doctrine. Transferability: while local counts vary across domains, the slot architecture remains stable, allowing the same reading practice to migrate across regimes. Machine legibility: AI systems are constrained to operate within verifiable institutional boundaries. The result is a minimal executable syntax for institutional reasoning—an evidence-to-structure architecture that makes complex systems readable to humans and governable for machines.