Can the Question and Judgment Show Why They Still Hold?

AI systems are increasingly capable of producing strong answers, yet answer quality alone does not establish that the question was adequate, that the evidence remains applicable, that the judgment is still qualified, or that authority to act remains valid after conditions change. This public preprint presents an externalized qualified-state architecture for preserving the operational structure around a judgment rather than only the generated answer. The architecture separates output, judgment, authority, and execution; represents question, evidence, premises, unknowns, dependencies, applicability, closure, authority, and version as explicit state; and supports selective requalification when qualified change occurs. The paper also reports fresh PoC evidence. In the tested suites, the underlying strong LLM was already highly accurate on final-answer endpoints. The judgment-layer configuration preserved task performance while adding question qualification, typed-state integrity, evidence-chain reconstruction, lifecycle control, and selective requalification. In a separate 48-case system-level resource study, direct LLM use and the judgment-layer configuration used the same overall number of model calls (48 → 48), while total tokens decreased from 19,842 to 18,334. First-pass mean latency increased modestly from 2,418.8 ms to 2,732.3 ms (~1.13×), whereas subsequent-change mean latency decreased from 4,291.0 ms to 2,642.5 ms (~0.62×, about 38% lower). The central engineering distinction is that strong AI may answer a question correctly while the question itself remains inadequate for the decision purpose. The proposed architecture therefore treats “questioning the question” and externalizing judgment into an inspectable, reconstructable, reusable, and requalifiable state as separate engineering problems. A governance crosswalk maps the architecture to selected requirements in the EU AI Act, NIST AI RMF, ISO/IEC 42001, Japan’s AI Guidelines for Business, and GDPR. These mappings describe technical support only; they do not constitute legal compliance, certification, conformity assessment, formal safety assurance, or legal advice.