Evidence-Grounded Runtime Assurance for AI-Assisted Power-Electronics Design: Three-State Constraint Semantics and Deterministic Validation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI-assisted engineering workflows can assign stronger validity claims than their evidence supports. This study develops an executable evidence contract for preliminary power-electronics design by instantiating established multi-valued monitoring. A deterministic monitor checks input availability, numerical representation, admissible domains, model coverage, and semiconductor-loss allocation before assigning SATISFIED, VIOLATED, or UNRESOLVED to each constraint. A 2 kW, two-phase converter linking a 48–120 V battery to a 400 V bus provides the application. The monitor reproduces all 4096 local verdicts in 512 synthetic records and all 37 separate diagnostic controls. Under identical evidence rules, unresolved-as-pass produces 71.6% conditional unsupported acceptance, and unresolved-as-violation produces 89.4% conditional false violation. Global abstention preserves final decisions but retains only 35.8% local-state accuracy, showing why constraint-level localization matters. Seeded property checks, source-level mutation tests, and controlled revision replays corroborate the specified behavior. The revision workflow requires documented evidence updates through every supported route and invalidates stale evidence after design changes. Prior inputs and labels remain available through a versioned migration audit. The contribution is reproducible domain-specific decision assurance and traceability, not new three-valued logic. The results establish bounded software conformance, not evidence authenticity, independent physical-model validation, or hardware qualification.