Boundary-Aware Assertion Control Under Fixed Semantic Verification Constraints

Autonomous systems deployed in open-world environments often operate under fixed semantic verification boundaries: only a restricted class of external evidence is admissible for certifying asserted outputs at inference time. In such settings, a nonzero fraction of deployment inputs may be uncertifiable, meaning that no admissible justification exists regardless of model accuracy or confidence. This work studies how assertion itself should be governed once such grounding limits are acknowledged. Rather than treating assertion as a default outcome of inference or as a confidence-based abstention mechanism, assertion is modeled as an architectural exposure decision subject to externally imposed verification and throughput constraints. The paper introduces a boundary-relative notion of audited divergence, measuring the rate at which asserted outputs fall outside the admissible verification boundary. Assertion allocation is formulated as a constrained optimization problem: minimize audited divergence subject to a required minimum assertion rate. The formulation depends only on deploy-time observable quantities and does not invoke correctness probabilities, ground-truth labels, or model-internal confidence. Under a precisely defined non-recognizability regime, where inference-time signals carry no information about certifiability, we prove a structural optimality result: among all admissible assertion policies satisfying the throughput constraint, audited divergence is minimized by a threshold policy with respect to estimated uncertifiability, with randomized tie-mixing when required. Deterministic and stochastic controllers achieving the optimal tradeoff are derived, along with robustness bounds under estimation error. A concrete deployment scenario and a deterministic worked example demonstrate how boundary-aware assertion control functions as an implementable engineering mechanism in audit-constrained systems. The results are model-agnostic and apply independently of learning paradigm or calibration scheme, provided the verification boundary is fixed. This work establishes assertion control as a distinct architectural layer for managing exposure under bounded semantic verification, complementing, but not replacing learning, calibration, or boundary expansion methods.