XAI as Decision-Boundary Instrumentation for AI Governance

High-risk AI systems are increasingly deployed as proprietary black-box services, creating a structural accountability gap at the core of emerging governance regimes: Article~26 of the EU AI Act and the NIST AI Risk Management Framework's MEASURE function both require per-decision operational monitoring, yet deploying institutions interact with models exclusively through inference APIs and cannot inspect model internals. This paper proposes a telemetry architecture that repositions XAI methods as governance infrastructure rather than explanation tools. Five measurement signals, derived from STAMP systems-theoretic safety analysis and bounded by the black-box admissibility constraint, generate a structured, machine-readable evidence snippet synchronously at each decision event. Eight engineering demonstrators covering all Annex~III risk categories of the EU AI Act validate the architectural claim across structurally distinct deployment domains. The result transforms accountability from an aspirational governance principle into an operational requirement: integrity-checked, reproducible, contestable, and auditable at the per-decision level.