Boundary-First System Architectures for Reliability Under Semantic Uncertainty

This paper develops a boundary-first architectural theory of reliability for autonomous systems operating under semantic uncertainty, including phenomena commonly labeled as hallucination in large language models and other generative systems. In deployment settings where correctness depends on external semantic grounding rather than internal model confidence alone, reliability is constrained by the availability of admissible evidence and verification semantics. The paper formalizes this constraint by treating semantic verification boundaries as first-class architectural objects, distinct from inference mechanisms, calibration, or uncertainty estimation. The central result establishes boundary dominance: under a fixed semantic verification boundary, the achievable tradeoff between assertion rate and audited divergence (boundary-relative failure under admissible audits) is invariant under all transformations of the inference core that do not modify verification reach. Consequently, no strict improvement in certifiable reliability at fixed throughput is possible without explicit boundary expansion. Model-level improvements, calibration, retrieval-augmented generation, or internal critique mechanisms saturate at a grounding-limited reliability frontier whenever uncertifiable input mass is nonzero. The analysis further proves that boundary expansion is the only architectural operation capable of reducing uncertifiable input mass without increasing deferral, and that boundary-free evaluation regimes render reliability attribution structurally non-identifiable. Apparent reductions in hallucination or failure rates observed without explicit boundary representation cannot be causally attributed to inference improvements rather than deferral, audit reinterpretation, or implicit verification changes. The results are structural and deductive rather than empirical. They do not prescribe specific hallucination detectors, verification tools, governance frameworks, or optimization strategies, nor do they model cost, latency, or dynamic boundary evolution. Instead, the paper identifies architectural invariants and necessary conditions under which reliability claims are meaningful when semantic uncertainty is unavoidable. This work is intended as a foundational contribution for regulated, safety-critical, or audit-constrained deployments, and as a prerequisite to domain-specific instantiations that define admissible evidence classes and verification predicates explicitly.