Logged Is Not Done: Verifiable AI Operations for the EU AI Act Era (Part 01 of 15)

Abstract As AI agents start to act in European critical infrastructure (telecom networks, payment systems, cloud platforms and industrial robots), a recurring gap appears. A system can log that an operation was authorized and acknowledged, yet still be unable to show that the evidence behind it was current, or that the reported result matches what actually happened. This gap matters to obligations under the EU AI Act, including: risk management (Article 9); record-keeping and logging (Article 12); human oversight (Article 14); accuracy, robustness and cybersecurity (Article 15); post-market monitoring and serious-incident reporting (Articles 72–73). This publication presents the foundational part of a 15-part technical series, also submitted as an IETF Internet-Draft. It defines a vendor-neutral evidence-consistency framework built on these elements: every consequential operation receives a stable identity and a declared verification objective; each observation carries its provenance, resource generation, time interval and clock uncertainty; domain-specific checks produce a scoped three-valued result: SUPPORTED, CONTRADICTED or UNRESOLVED. UNRESOLVED is a first-class outcome. Missing or stale evidence is recorded honestly rather than reported as success, which gives auditors, oversight personnel and incident investigators a reconstructable decision record. The framework keeps verification separate from authority. A positive result never grants permission to act; any consequential effect stays under a separate, accountable control boundary. Acknowledgement, completion, physical outcome and financial settlement are reconciled as independent evidence. A three-layer architecture keeps shared evidence semantics apart from twelve domain profiles: radio access and network incident recovery; telecom sessions and charging; hardware and DMA command paths; multi-agent negotiation and asynchronous tools; robotics and physical safety; multilingual emergency instructions; browser agents and agent migration; autonomous commerce; accelerator reset. The material is also relevant to engineering work under NIS2, the Cyber Resilience Act, DORA and the Machinery Regulation. The PDF includes: a reader's summary with diagrams; an analytical latency and capacity model with a measurement plan; 12 proposed verification tests; threat and edge-case profiles; a searchable catalogue of 639 engineering mechanisms. The document describes technical mechanisms and makes no legal-compliance, certification or conformity claim.