Operationalizing the EU AI Act through eIDAS Trust Services Primitives: A Reference Mapping for High-Risk AI Systems

Version 1.1-preprint update. v1.1 strengthens §3.6 (crypto agility and post-quantum readiness) with the European Commission PQC coordinated-roadmap Recommendation and Estonia's April 2026 ROAD2PQ national migration roadmap. The revision frames CADI/CBOM and crypto-agile architecture as emerging public-sector procurement requirements for long-lived AI Act evidence stacks, while preserving the paper's article-by-article mapping and worked-example structure. An editorial pass on 2026-05-23 applied targeted hand-authored insertions across the abstract, §1.1, §2, §3.6, and the article-by-article mapping in §4. The insertions sharpen the procurement-side framing ("ask for the CBOM, the key-history plan, and the re-signing story before signing the contract"), name the adversarial-falsification posture as the healthy way to read the mapping table, and add the small but load-bearing observation that signed artefacts do not turn careless judgments into careful ones. The argument structure and the article-by-article rows of the mapping are unchanged. The competing-interest disclosure and the reference list are unchanged from v1.0.1. The PDF rendering has been rebuilt from the editorial-pass markdown source; the page count is 42 (v1.0.1 was 41). Prepared as a Zenodo new-version under concept DOI 10.5281/zenodo.20257971. What is new in v2.0.1 (CSI derivative — DOI typo fix). Errata-only republish over v2.0-csi: the §1 footnote's Zenodo DOI references are now correct (concept DOI 10.5281/zenodo.20257971; v1.1 versioned DOI 10.5281/zenodo.20265919). All other content is identical to v2.0-csi. This version is a leaner venue derivative of the v1.x preprint line, prepared for submission to Computer Standards & Interfaces (Elsevier). The intellectual contribution is the same standards-gap reference mapping; v2.0 reshapes it into a shorter, IMRaD design-and-evaluation paper with new empirical content. Main changes from v1.1-preprint (versioned DOI 10.5281/zenodo.20265919): IMRaD restructure. Nine numbered sections: Introduction, Related work (standalone), Methodology, Architectural view (§4.1–4.6), Article-by-article reference mapping with a layer × article matrix (§5), Contested mapping decisions (§6, collecting the rows around Articles 10, 14, and 50 in one place), Evaluation (§7), Discussion and limitations (§8), Conclusion (§9). New §7 Evaluation, fully written. A worked MCP trace traversed end to end, reaching seven AI Act articles on one primitive set (§7.1); a conformance check across two independent EATF reference verifiers on an 11-vector public corpus, with identical verdicts package by package (§7.2); a single-machine performance measurement on an Intel Core Ultra 5 135U for classical RSA-4096 and hybrid (RSA-4096 + ML-DSA-65) signing and verification (§7.3). Hybrid signer implemented and measured. The EATF reference signer was extended to emit an ML-DSA-65 (NIST FIPS 204) signature alongside the classical RSA signature; v1.1 only projected the hybrid cost, v2.0 reports it: sign 9.0 ms median, verify 4.2 ms median, package 11.3 KB. New references and §2 expansion. Verifiable-inference as a heavier point on the cost curve (Kang et al.); model cards (Mitchell et al.) and datasheets (Gebru et al.) for the §6.1 boundary; an empirical pre-enforcement evidence-readiness baseline; a parallel-domain case study in adaptive educational AI. Length and form. About 4,000 words shorter than v1.1, with the §1 abstract reduced to 199 words to fit the Elsevier 250-word cap; Figure 1 (layer × article matrix) added in vector form; a table of contents and clickable in-text citations were added for the venue PDF. Declarations updated. The Generative AI declaration is narrowed to "structural editing and language polishing" only; the Competing interests, Funding, and Data availability statements are unchanged in substance. Status. Submission-ready manuscript prepared for Computer Standards & Interfaces; not yet peer reviewed, not submitted to the journal at the time of this deposit. The v1.x preprint line remains accessible at the versioned DOI above for readers who want the longer background treatment. What is new in v2.0 (CSI derivative). This version is a leaner venue derivative of the v1.x preprint line, prepared for submission to Computer Standards & Interfaces (Elsevier). The intellectual contribution is the same standards-gap reference mapping; v2.0 reshapes it into a shorter, IMRaD design-and-evaluation paper with new empirical content. Main changes from v1.1-preprint (versioned DOI 10.5281/zenodo.20265919): IMRaD restructure. Nine numbered sections: Introduction, Related work (standalone), Methodology, Architectural view (§4.1–4.6), Article-by-article reference mapping with a layer × article matrix (§5), Contested mapping decisions (§6, collecting the rows around Articles 10, 14, and 50 in one place), Evaluation (§7), Discussion and limitations (§8), Conclusion (§9). New §7 Evaluation, fully written. A worked MCP trace traversed end to end, reaching seven AI Act articles on one primitive set (§7.1); a conformance check across two independent EATF reference verifiers on an 11-vector public corpus, with identical verdicts package by package (§7.2); a single-machine performance measurement on an Intel Core Ultra 5 135U for classical RSA-4096 and hybrid (RSA-4096 + ML-DSA-65) signing and verification (§7.3). Hybrid signer implemented and measured. The EATF reference signer was extended to emit an ML-DSA-65 (NIST FIPS 204) signature alongside the classical RSA signature; v1.1 only projected the hybrid cost, v2.0 reports it: sign 9.0 ms median, verify 4.2 ms median, package 11.3 KB. New references and §2 expansion. Verifiable-inference as a heavier point on the cost curve (Kang et al.); model cards (Mitchell et al.) and datasheets (Gebru et al.) for the §6.1 boundary; an empirical pre-enforcement evidence-readiness baseline; a parallel-domain case study in adaptive educational AI. Length and form. About 4,000 words shorter than v1.1, with the §1 abstract reduced to 199 words to fit the Elsevier 250-word cap; Figure 1 (layer × article matrix) added in vector form; a table of contents and clickable in-text citations were added for the venue PDF. Declarations updated. The Generative AI declaration is narrowed to "structural editing and language polishing" only; the Competing interests, Funding, and Data availability statements are unchanged in substance. Status. Submission-ready manuscript prepared for Computer Standards & Interfaces; not yet peer reviewed, not submitted to the journal at the time of this deposit. The v1.x preprint line remains accessible at the versioned DOI above for readers who want the longer background treatment. Version 1.0.1 (18 May 2026) is a minor revision of v1.0-preprint (17 May 2026, archived under the same concept DOI). v1.0.1 preserves the analytical claims, the article-by-article mapping, all numbered tables, and the architectural view of v1.0. It applies the following cleanup so that the public preprint no longer carries internal editorial scaffolding: Removes §2.6 (Reviewer-risk register) and §2.7 (Venue positioning), which were authoring-stage tables aimed at peer reviewers and venue selection rather than at readers of the paper. Introduces a new §2.6 (Claim boundaries) that consolidates the public-facing claim-scope language — preserving the precision distinctions between “supports evidence for” and “satisfies”, between “uses eIDAS vocabulary” and “is an eIDAS trust service”, and between “reduces future-verification risk” and “guarantees decade-scale validity”. No changes to §3 (architectural view), §4 (article-by-article mapping), §5 (contested decisions), §6 (worked example), §7 (open rows), or the references list. This revision affects positioning only; the substantive contribution of v1.0 stands. The competing-interest disclosure remains as in v1.0. v1.0 preprint (17 May 2026) of Operationalizing the EU AI Act through eIDAS Trust Services Primitives: A Reference Mapping for High-Risk AI Systems. This working paper maps selected high-risk obligations in Regulation (EU) 2024/1689 (the EU AI Act) to cryptographic and trust-service primitives drawn from eIDAS/eIDAS 2.0, ETSI EN 319-series standards, IETF RFC 3161 timestamping, W3C Verifiable Credentials, JSON canonicalization, and post-quantum signature practice. Its central contribution is an article-by-article and layer-by-layer reference mapping for producing independently verifiable evidence about AI system behavior. Version v1.0-preprint is derived from publication-prep rc4 and adds a structured failure-case appendix for broken evidence packages, completes a URL verification pass, and tightens claim-risk wording from compliance guarantees toward evidence-support formulations. The mapping is implementation-agnostic. The Agent Trust Framework (EATF) is used as a worked example because its artifacts are publicly observable. Tyche Institute is a research entity, not a trust service provider or qualified trust service provider, and this paper does not claim that EATF or any implementation certifies legal compliance. Note added 9 August 2026 — repository links in this PDF. This version cites github.com/tyche-institute/eatf (and its lib-python and test-vectors paths) as publicly observable artifacts. That repository was made private on 23 July 2026 during a GitHub organisation consolidation, so those three links now return 404. The public substitute is github.com/tyche-institute/eatf-verifier (Apache-2.0), which carries the reference verifiers, test vectors and CLI surface described here. Every other link in the PDF was re-checked on 9 August 2026 and resolves. The deposited PDF is left unchanged so that the bytes continue to match this record.