Recent generative AI systems can produce highly plausible interpretations of linguistic and operational inputs, yet plausibility alone does not guarantee execution admissibility. In real-world settings, a fluent or high-confidence output may still be semantically ill-typed, structurally inconsistent, policy-violating, or insufficiently justified for action. This paper presents a layered neuro-symbolic interpretation runtime that explicitly separates candidate generation from execution eligibility. The proposed framework combines candidate proposal, typed semantic role filtering, well-formedness constraints, policy arbitration, action gating, and audit tracing. Candidate interpretations are generated by a proposal layer, filtered by semantic and structural admissibility, and assigned one of three operational outcomes: ACCEPT, REVIEW, or REJECT. Hard constraints are applied prior to score-based ranking, preventing high-confidence but invalid candidates from being selected. On an internal calibration set (20 cases), the runtime achieved 20/20 decision-match accuracy with explicit trace outputs preserved for every decision.
Myeong Jun Jo (2026) studied this question.