A Supervisory-Evidence Ontology for Agentic AI under EU Law: Candidate Minimum Conceptual Set and Temporal Extension

Agentic AI has outpaced the ontologies intended to govern it. Commercial and academic ontologies released between 2024 and 2026 cluster around a shared enterprise core of Agent, Skill, Policy, Memory, and Outcome, but none was designed to produce evidence that a European supervisor can ingest. Current supervisory practice relies on ad-hoc documentation produced per controller and per request. This working paper proposes a shared representational layer for agentic AI accountability evidence under EU law, structured in three components. The first is a candidate Minimum Conceptual Set of 23 conceptual slots, separated into an agent-behaviour core (twelve slots) and a supervisory-evidence layer (eleven slots). Under strict reuse-zero accounting these 23 slots correspond to 16 net-new classes plus 7 reuse slots (5 DPV reuses, 2 PROV-O reuses); the Turtle vocabulary contains 69 owl:Class declarations once subtypes and named categories are counted. Each slot is mapped to evidence needs arising under the GDPR, the AI Act, or NIS2, or is motivated by structured reading of a 25-case sample of EU ADM enforcement. The second component is a temporal extension expressed in OWL-Time and made structurally checkable through SHACL shapes for delegation validity, revocation propagation, policy versioning, and evidence decay. The third is an integration layer that reuses GDPRov, DPV, and PROV-O through owl:imports rather than reinventing their concepts. A v1.2 SHACL release ships Profile A (AP-inspired permissive, quantitative) and Profile B (CNIL/German-guidance-inspired stricter, qualitative) alongside a size-based SME proportionality profile. The paper does not claim reference-architecture status. It claims that the synthesis and design choices are defensible, reproducible, and testably better than ad-hoc practice for the teams that would use it. Validation is pre-registered through three open tracks (inter-rater consistency on the case sample, SHACL throughput, structural fit across topologies); these tracks remain pending. Limitations include single-coder empirical base, documented distributive effects that specification work cannot correct, and dependency on external regulatory coherence that is empirically contingent. v0.5.2 corrects three case-sample ECLI citations (B11, B12, B16) and one attribution label; see ERRATA_v0_5_2.md. No ontology, shape, or validation logic changed.