Public evidence for AI Act deployer obligations before enforcement: A baseline from Estonia, EU procurement, and AI policy documents

Abstract

Version 1.1 (23 May 2026) is an author hand-pass over v1-ejrr-submission. The empirical findings, source tables, and reference list are unchanged. v1.1 tightens the abstract, introduction, discussion, and conclusion in the author’s voice; preserves the central claim that public evidence-readiness can be measured before AI Act enforcement pressure changes the documentation environment; and the closing positions remain plain: the result is a baseline, not a verdict; snapshots age, researchers should say when they took them. Manuscript EJRR-2026-0121 remains under review at the European Journal of Risk Regulation. Repository pointer change: the GitHub working-repository URL (github.com/sapsan14/bseade) is removed from this deposit’s related identifiers because the repository is private; the reproducibility anchor is the OSF mirror at https://osf.io/dh9gy/. Audit-stage working paper. This is version 1 (2026-05-21) of an empirical legal-policy article submitted on the same date to the European Journal of Risk Regulation (Cambridge University Press, manuscript ID EJRR-2026-0121, status Under Review). It is posted here as an open-access preprint so that the standards-body community (ETSI, CEN-CENELEC, EU AI Office) and the AI-governance research community can cite the pre-enforcement baseline during the EJRR peer-review cycle. Article scope. The article reports a reproducible audit-stage baseline of how the EU Artificial Intelligence Act (Regulation (EU) 2024/1689) deployer-obligation themes (Articles 10, 12, 13, 14) are surfaced in publicly available documentation in the weeks before the 2 August 2026 enforcement milestone. Three workstreams: WS1. 96 Estonia public-sector AI deployment records from kratid.ee, the public register operated by RIA (Riigi Infosüsteemi Amet) as part of the Estonian Kratt programme. Estonia is treated as a critical case (strongest available test bed), not as an EU-representative sample. WS2. 750 Tenders Electronic Daily (TED) procurement notices, narrowed by a conservative two-stage filter (Common Procurement Vocabulary code gate plus mandatory multilingual AI vocabulary) to 2 retained AI-relevant tenders. WS4. 248 passages from 41 AI policy documents (AI Watch, EU AI Office, member-state policy sources). Headline results. Conservative throughout. Estonia deployment records show a low public evidence-readiness distribution across Article 10/12/13/14 signal categories (score 0–4, mean 0.844 out of a possible 4). In the retained TED set, strict AI Act references are 0/2. In the policy-document passage set, strict cryptographic-evidence language is 0/248. These results do not establish legal conformity or internal operational practice; they provide a public evidence baseline against which post-enforcement documentation can be compared. Reproducibility. The full data, processing scripts, generated outputs, deterministic 90-record IRR sample manifest, and machine-vs-machine kappa noise-floor baseline are deposited at the BSEADE OSF project (https://osf.io/dh9gy/, CC BY 4.0). Source code and reproducibility scripts are at github.com/sapsan14/bseade. Audit-stage caveat. The current draft is labelled audit-stage because independent human paired-coding has not yet been performed. The validation plan is committed as a condition of final journal acceptance (see Section 3.6 of the manuscript) and will report per-field Cohen's kappa in the methods supplement of the revised version. A machine-vs-machine kappa noise-floor baseline is reported as a lower comparison band only, not as a substitute for the human paired-coding result. Companion artefacts. The same content is mirrored as a manuscript snapshot in the Tyche Research Vault (papers/bseade-paper-a-v1/) and is queued for arXiv (cs.CY) and SSRN posting. Author affiliation. Anton Sokolov, Tyche Institute, Tallinn, Estonia (https://tyche.institute), anton.sokolov@tyche.institute, ORCID 0000-0003-2452-7096. The author works as a Public Key Infrastructure engineer in his day job; the present research is conducted in his independent research capacity at Tyche Institute and does not represent or reflect the views of his employer. AI use declaration. Claude (Anthropic) and Codex (OpenAI) coding-agent sessions assisted with prose drafting. All empirical claims and final wording are the author's responsibility. Citation. Sokolov, Anton (2026). Public evidence for AI Act deployer obligations before enforcement: A baseline from Estonia, EU procurement, and AI policy documents. Zenodo working paper, version 1. CC BY 4.0.

Ähnliche Arbeiten