Research

From Obligation to Verifier Profile: A Computational-Legal Method for Making EU AI Act Duties Testable

Anton Sokolov · v0.2 — under review · May 2026

Computational legal studies repeatedly meets a step it rarely documents: the move from a legal obligation to a test a machine can run. The paper proposes a small grammar that takes one obligation, reads out the evidence a third party would need to confirm it, names a cryptographic primitive that can produce a matching verifiable artifact, and records the result as a verifier profile — with an explicit scope test that sends obligations cryptography cannot reach out of scope rather than forcing them. A verifier profile supports evidence for an obligation; it does not certify legal compliance or replace Article 43 conformity assessment.

A research article proposing a reproducible method for the translation step computational legal studies usually performs by hand and leaves unrecorded: turning a legal obligation into something a machine can test. The method is a small grammar (four steps and one scope gate) that maps an obligation to the evidence a third party needs, to a cryptographic primitive that can produce a matching verifiable artifact, and to a compact verifier profile recording assumptions and what a verifier must check. To show the grammar does real work, the paper instantiates it on the high-risk obligations of the EU AI Act using the mature trust-service primitives of eIDAS and eIDAS 2.0; the instantiation doubles as validation, with two independent verifiers agreeing on a public conformance corpus and single-digit-millisecond costs. The contribution is the method, not any single mapping it produces. The evidence boundary is explicit: a verifier profile supports evidence, it does not certify legal compliance, prove oversight effective, or replace Article 43 conformity assessment. Submitted to the Cambridge Forum on AI: Law and Governance on 28 May 2026; manuscript ID CFL-2026-0048.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026verifierprofile,
  author       = {Sokolov, Anton},
  title        = {From Obligation to Verifier Profile: A Computational-Legal Method for Making {EU} {AI} Act Duties Testable},
  year         = {2026},
  month        = may,
  note         = {Research article; submitted to the Cambridge Forum on AI: Law and Governance (CFL-2026-0048, under review)},
  institution  = {Tyche Institute}
}

World Cards for Evidence-Ready AI Systems: Descriptor Governance under European Technology Law

Anton Sokolov · v0.1 — under review · May 2026

Public and semi-public descriptors — model cards, dataset cards, agent cards, API descriptions, software bills of materials, provenance records, credentials, catalogues, governance factsheets — support disclosure, but they do not automatically become inspectable evidence. A descriptor that can be read is not the same as a descriptor that can be checked. The world-card method treats such descriptors as evidence infrastructure and compares them on a common footing under European technology law, without claiming legal compliance, trust-service status, or production validation.

A refereed article examining a practical legal-technical problem for European AI governance: a wide range of public and semi-public descriptors are meant to support disclosure, yet they rarely become evidence that a later reviewer can independently inspect. The article proposes a world-card method for comparing such descriptors as evidence infrastructure, locating the analysis in European technology law — the EU Artificial Intelligence Act, eIDAS, and the broader descriptor-governance landscape. The evidence boundary is explicit: the article does not claim legal compliance, trust-service status, notified-body status, or production validation. Submitted to the European Journal of Law and Technology (Refereed Articles) on 29 May 2026 as OJS submission 1247.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026worldcards,
  author       = {Sokolov, Anton},
  title        = {World Cards for Evidence-Ready {AI} Systems: Descriptor Governance under European Technology Law},
  year         = {2026},
  month        = may,
  note         = {Refereed article; submitted to the European Journal of Law and Technology (OJS submission 1247, under review)},
  institution  = {Tyche Institute}
}

Making AI Agents Legible to the State: X-Road, Carrier-Bound Evidence, and the Legal Architecture of Delegated Autonomy

Anton Sokolov · v0.4 — under review · May 2026

A successful carrier transaction is not evidence of agent authority. Public-sector interoperability infrastructures such as Estonia's X-Road / X-tee can identify organisations, sign, timestamp, and log a call, yet still leave opaque whether the actor behind it was a human officer or a software agent acting under delegated authority. Closing that gap needs a separate carrier-bound evidence profile — records that bind route identifiers, payload hashes, delegation claims, policy scope, time evidence, key status, and independent verifier appraisal — rather than turning the carrier into a native AI-agent verifier.

A legal-computational case study using Estonia's X-Road / X-tee data-exchange infrastructure to ask how AI agents can become legally and administratively legible at the boundary of public-service exchange. The article keeps X-Road in the role it is good at — governed carrier, organisational trust fabric, and audit witness — and proposes that agent-specific evidence live in a separate carrier-bound evidence profile that can be signed, hashed, appraised, and reconciled against the carrier record. It connects the design pattern to the EU AI Act's regulatory-sandbox logic, public evidence-readiness gaps in AI deployment records, and the institutional role of public-key infrastructure as governance infrastructure. A synthetic lab case and a public X-Road development-stack exchange are used as illustration, not as a production assurance claim. Submitted to the Cambridge Forum on AI: Law and Governance on 28 May 2026; manuscript ID CFL-2026-0047.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026legible,
  author       = {Sokolov, Anton},
  title        = {Making {AI} Agents Legible to the State: {X-Road}, Carrier-Bound Evidence, and the Legal Architecture of Delegated Autonomy},
  year         = {2026},
  month        = may,
  note         = {Case study; submitted to the Cambridge Forum on AI: Law and Governance (CFL-2026-0047, under review)},
  institution  = {Tyche Institute}
}

A Verification Commons for AI-Agent Evidence Packages: Negative Controls, Offline Reproducibility, and Claim Boundaries in EATF

Anton Sokolov · v0.9 — under review · May 2026

↳ companion-of: sokolov2026breakablereceipts

Weak or overclaimed AI-agent evidence is a practical digital threat: a dashboard entry or mutable log row cannot be checked once the producing service is gone. Treating EATF as a non-commercial verification commons — open evidence-package formats, offline verifier libraries, conformance vectors, and negative controls — makes failure visible, while explicit non-claims hold the line that a valid package is not truth, a signature is not compliance, and a hosted testbed is not a regulated trust service.

A non-commercial case study that frames weak audit trails for AI-agent actions as a digital-threat problem. EATF (the Agent Trust Framework) is treated not as a trust service but as an artifact: open evidence-package formats, verifier libraries, conformance vectors, Model Context Protocol (MCP) attestation profiles, and negative controls. Its core object, the Agent Evidence Package (AEP), bundles canonical bytes, digests, metadata, and signatures so that a later reviewer can rehash bytes, check signatures, inspect timestamps and boundaries, and reproduce failures offline. The paper locates EATF's gap among identity, authorization, runtime-evidence, and governance-event projects, reports bounded bench evidence, and states explicit non-claims. The Breakable Receipts synthetic lab serves as a companion negative control; X-Road is a supporting carrier-boundary example only. Submitted to ACM Digital Threats: Research and Practice on 28 May 2026; manuscript ID DTRAP-2026-0077.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026verificationcommons,
  author       = {Sokolov, Anton},
  title        = {A Verification Commons for {AI}-Agent Evidence Packages: Negative Controls, Offline Reproducibility, and Claim Boundaries in {EATF}},
  year         = {2026},
  month        = may,
  note         = {Case study; submitted to ACM Digital Threats: Research and Practice (DTRAP-2026-0077, under review)},
  institution  = {Tyche Institute}
}

Breakable Receipts: A Synthetic Case Study in Layered AI Attestation Evidence

Anton Sokolov · v0 — under review · May 2026

Layered AI attestation evidence separates into three kinds — what hashes and signatures bind byte-for-byte, what domain replay can recompute, and what still needs human judgment — and controlled break cases show which layer catches each failure, including a post-quantum ML-DSA-65 verification path.

An artifact-first synthetic evidence lab. A runtime tutoring event is packaged as an EATF Agent Evidence Package (AEP) carrying an OVERT-style runtime receipt, then deliberately broken in controlled ways and re-verified through two independent paths: cryptographic envelope checks and domain-level semantic replay. The executed corpus includes transitional classical-signature cases, explicit ML-DSA-65 post-quantum positive and negative cases, deterministic mutation sweeps, and ordered chain checks. The paper makes no conformance claim with OVERT or any external attestation scheme; its contribution is a reproducible method for exposing where AI governance evidence becomes independently testable, where it merely remains well-documented, and where policy judgment is unavoidable. Submitted to the ESORICS 2026 RAISE workshop on 28 May 2026 as Submission 957.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026breakablereceipts,
  author       = {Sokolov, Anton},
  title        = {Breakable Receipts: A Synthetic Case Study in Layered {AI} Attestation Evidence},
  year         = {2026},
  month        = may,
  note         = {Synthetic case study; submitted to the ESORICS 2026 Workshop on Real-world AI Security and Engineering for Cybersecurity Systems (RAISE), submission 957},
  institution  = {Tyche Institute}
}

Adopting AI Act Transparency in Three Stages: A Planning Typology for SMEs

Anton Sokolov · v0.1 — under review · May 2026

↳ companion-of: sokolov2026registereddisclosure

EU AI Act transparency obligations can be calibrated to a three-stage adoption typology — registered disclosure → live attestation → continuous compliance audit — that lets SMEs ship a meaningful compliance posture at each stage and lets regulators enforce against the stage actually reached, rather than against uniform Article 17 compliance from the high-risk applicability date.

A Translational policy essay proposing a three-stage adoption typology for AI Act transparency obligations and arguing that enforcement toward small and medium-sized enterprises (SMEs) can be calibrated to staged adoption rather than to uniform compliance from day one. Each stage is a meaningful compliance posture in itself, not a stub on the way to 'real' compliance. The typology is a planning artifact for SMEs and a calibration vocabulary for regulators. The argument's policy implication, developed in §6, names the institutional preconditions — particularly the stage-3 auditor profession — that still need to be built before the AI Act's full compliance posture is reachable in practice. Companion to an empirical case study on the same typology under double-anonymous review at ACM Digital Government: Research and Practice (DGOV-2026-0116); the two artifacts are deliberately distinct. Submitted to Data & Policy on 27 May 2026 as DAP-2026-0199.

doi:10.5281/zenodo.20402526 → concept DOI: 10.5281/zenodo.20402526

@misc{sokolov2026typology,
  author       = {Sokolov, Anton},
  title        = {Adopting {AI} {Act} Transparency in Three Stages: A Planning Typology for {SMEs}},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.1},
  doi          = {10.5281/zenodo.20402526},
  url          = {https://doi.org/10.5281/zenodo.20402526}
}

Registered Disclosure as a First-Stage AI Act Deployment: A Three-Stage Adoption Typology for SMEs, with a Case Study in a Small Audit SaaS

Anton Sokolov · v1.0 — under review · May 2026

↳ companion-of:sokolov2026mapping

EU AI Act transparency obligations can be adopted incrementally: stage 1 — registered disclosure — is reachable for roughly three percent of a representative SME codebase, with zero new direct dependencies, using twenty-five percent of an open framework's audit-event API surface.

A three-stage adoption typology for EU AI Act transparency obligations — registered disclosure, live attestation, continuous compliance audit — with a worked example of stage 1 in the eAudit building-audit SaaS at eaudit.ee. The integration of the open EATF framework added approximately 2 434 lines of code (3.08% of the project) and zero new direct dependencies. The paper is framed around the typology rather than the case so that the typology is useful to other SMEs independent of this deployment. The paper carries a first-stage / family-affiliation disclosure: the author is a co-developer of both the framework and the SaaS, and the operator (TADF Ehitus OÜ) is family-affiliated. Submitted to ACM DGOV on 26 May 2026; manuscript ID DGOV-2026-0116; double-anonymous review.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026registereddisclosure,
  author       = {Sokolov, Anton},
  title        = {Registered Disclosure as a First-Stage {AI} {Act} Deployment: A Three-Stage Adoption Typology for {SMEs}, with a Case Study in a Small Audit {SaaS}},
  year         = {2026},
  month        = may,
  note         = {Case study; submitted to ACM Digital Government: Research and Practice (DGOV-2026-0116, under double-anonymous review)},
  institution  = {Tyche Institute}
}

PKI as Governance Infrastructure: A Conceptual Framework

Anton Sokolov · v0.3 — under review · May 2026

The Estonian 2017 ID-card cryptographic crisis behaved like a constitutional event, not an IT incident — as a five-dimensional governance framework would predict.

A conceptual framework that treats public-key infrastructure as governance infrastructure rather than a security technology. Five dimensions — trust production, infrastructural power, mediated institutional trust, two-way accountability, and path dependence — each pinned to one load-bearing theorist (Mann, Bodó / Shapiro / Giddens, Scott, Star and Ruhleder, Pierson). The Estonian e-state is the worked case; the framework is portable to any national PKI. Submitted to Halduskultuur — Administrative Culture on 23 May 2026 (submission #419).

doi:10.5281/zenodo.20381436 → concept DOI: 10.5281/zenodo.20381436

@unpublished{sokolov2026pki,
  author       = {Sokolov, Anton},
  title        = {{PKI} as Governance Infrastructure: A Conceptual Framework},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.3; submitted to Halduskultuur (submission \#419)},
  institution  = {Tyche Institute}
}

Trust Fields and Public Administration: A Second Reading of Halduskultuur in the eIDAS 2.0 Era

Anton Sokolov · v0.3 — under review · May 2026

↳ second-reading-of: sokolov2026pki

The five discrete dimensions of the PKI-as-governance frame (#419) lift to six locally-measurable observables of a continuous trust field — the right ontology for eIDAS 2.0, the AI Act, and the post-quantum migration window.

The explicit second reading of #419 (PKI as Governance Infrastructure). The five-dimensional frame is re-read as a six-observable diagnostic of a continuous trust field — amplitude, phase, polarization, coherence, propagation kernel, gauge — that public administration can apply without further mathematics. The Estonian e-state is the worked case across the 2017 ROCA crisis, the 2020s QTSP procurement transition, and the active EUDIW and PQC migrations. The paper proposes eAIDAS (electronic AI Declaration and Attestation Services) as a Commission Implementing Act target for AI-attestation regulated at the field-equation level. Submitted to Halduskultuur — Administrative Culture on 25 May 2026 (submission #421); editor deference offered on co-review timing with #419.

doi:10.5281/zenodo.20381438 → concept DOI: 10.5281/zenodo.20381438

@unpublished{sokolov2026trustfields,
  author       = {Sokolov, Anton},
  title        = {Trust Fields and Public Administration: A Second Reading of Halduskultuur in the {eIDAS} 2.0 Era},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.3; submitted to Halduskultuur (submission \#421)},
  institution  = {Tyche Institute}
}

Closing the Evidence Gap: An Administrative-Culture Reading of AI Act Operationalization in the Estonian e-State

Anton Sokolov · v0.3.4 — under review · May 2026

↳ extends: sokolov2026pki

The gap between what the AI Act asks public administration to verify and what eIDAS-grade trust services can actually produce is not a technical gap — it is an administrative-culture gap, and halduskultuur is the substrate in which it will either be closed or be replaced by compliance theatre.

An administrative-culture reading of the operational mapping between the EU AI Act and eIDAS trust services. Four administrative competencies are named (reading evidence, compelling its production, cross-checking against the system, maintaining the evidence chain across time); three compliance-theatre risks are anticipated (vendor capture, cargo-cult deployment, performance theatre); five halduskultuur substrate-readiness markers are offered as a diagnostic. The Estonian e-state is read as the worked case across three scenarios — the 2002–2017 substrate-building period, the 2017 ROCA event as substrate-under-stress, and the 2020s QTSP tender transition as substrate-adapted — and three concrete moves for Estonian public administration are proposed. Routed to JeDEM rather than Halduskultuur to keep the Halduskultuur queue at #419 + #421 (editorial-density management); JeDEM's e-government and AI-governance scope is the natural non-Halduskultuur home. Submitted to JeDEM — eJournal of eDemocracy and Open Government on 25 May 2026 (submission #1282, double-blind).

doi:10.5281/zenodo.20381440 → concept DOI: 10.5281/zenodo.20381440

@unpublished{sokolov2026evidencegap,
  author       = {Sokolov, Anton},
  title        = {Closing the Evidence Gap: An Administrative-Culture Reading of {AI} {A}ct Operationalization in the {E}stonian e-{S}tate},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.3.4; submitted to JeDEM (submission \#1282)},
  institution  = {Tyche Institute}
}

Verifiable but Not Yet Credible: Reading Cryptographically-Attested Administrative Artifacts as Objects of Halduskultuur Inquiry

Anton Sokolov · v0.1 — under review · May 2026

↳ extends: sokolov2026pki; programmatic-sibling: sokolov2026evidencegap

Cryptographic verifiability and institutional credibility are two distinct properties of an administrative artifact, and the gap between them is the new methodological object for halduskultuur scholarship in the eIDAS 2.0 / AI Act / EUDIW era.

A methodological essay for halduskultuur researchers. It proposes the crypto-attested administrative artifact (CAAA) as a unit of analysis, distinguishes verifiability from credibility as travelling-separately properties, and uses the open Agent Evidence Package specification at eatf.eu as a methodological lens through which a researcher can see the translation gap concretely. Three research-design moves follow: attribute tracing, translation-gap mapping, and cross-jurisdictional comparison. Cites Yeung 2018 and Dunleavy & Margetts 2013 as the PA-of-technology canon the article extends. Programmatic sibling of JeDEM #1282 (Closing the Evidence Gap) with explicit duplication firewall: 0 long-line overlap, distinct unit of analysis, distinct method.

doi:10.5281/zenodo.20381442 → concept DOI: 10.5281/zenodo.20381442

@unpublished{sokolov2026trusttranslation,
  author       = {Sokolov, Anton},
  title        = {Verifiable but Not Yet Credible: Reading Cryptographically-Attested Administrative Artifacts as Objects of {Halduskultuur} Inquiry},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.1; under review at Halduskultuur (submission #422)},
  institution  = {Tyche Institute}
}

Unified Postquantum Trust Field Theory: A Field-Theoretic Reading of Attestation Infrastructure for the eAIDAS Era

Anton Sokolov · v0.2 — under review · May 2026

↳ companion-of: sokolov2026fieldweforgot; extends: sokolov2026trustfields

Public-sector attestations behave as quanta of a continuous trust field with six locally-measurable observables; three falsifiable predictions follow for the PQC migration, EUDIW uptake, and the polarisation structure of AI-decision contestation in 2027–2030.

The engineering-and-regulatory companion of the Halduskultuur Trust Fields paper. The five discrete dimensions of governance trust lift to six locally-measurable observables of a continuous trust field — amplitude, phase, polarisation, coherence, propagation kernel, gauge — each carrying an engineering specification and a regulatory specification. Three structural isomorphisms with physical field theory (locality, gauge invariance, conservation under continuous symmetries) are demonstrated, phenomenological field equations for the pre-, hybrid-, and post-PQC regimes are formulated, and the eAIDAS (electronic AI Declaration and Attestation Services) regime is proposed as the regulatory target. Initially drafted for First Monday; on 2026-05-26 First Monday's submissions page was found closed ("This journal is not accepting submissions at this time"), and the manuscript was re-routed to Big Data & Society under double-blind review on the same day as Manuscript ID BDS-26-0556.

doi:10.5281/zenodo.20381444 → concept DOI: 10.5281/zenodo.20381444

@unpublished{sokolov2026pqtft,
  author       = {Sokolov, Anton},
  title        = {Unified Postquantum Trust Field Theory: A Field-Theoretic Reading of Attestation Infrastructure for the {eAIDAS} Era},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.2; submitted to Big Data \& Society (BDS-26-0556, under review)},
  institution  = {Tyche Institute}
}

The Field We Forgot to Name: A Practitioner's Note on Trust as Physics, Not Plumbing

Anton Sokolov · v0.1 — under review · May 2026

↳ companion-of: sokolov2026pqtft; short-essay-of: sokolov2026trustasdata

After two decades of operational work in national identity infrastructure, public-key infrastructure stops looking like plumbing and starts looking like a field — and an eAIDAS that arrives by registry rather than by field equation is the wrong regulatory architecture for the next cycle.

The practitioner-register companion to Unified Postquantum Trust Field Theory. An essay-form account of why public-key infrastructure has, over the last two decades of operational work in national identity systems, stopped looking like plumbing and started looking like a field. The essay defends three structural properties that field-theoretic thinking gives the regulator — locality, gauge invariance, conservation — and warns against an eAIDAS that arrives by registry rather than by field equation. Submitted as a companion to Unified Postquantum Trust Field Theory to First Monday on 25 May 2026. An expanded research-article version was deposited 2026-05-26 as the canonical preprint (see *Trust as Data Infrastructure*, DOI 10.5281/zenodo.20399235).

doi:10.5281/zenodo.20381446 → concept DOI: 10.5281/zenodo.20381446

@unpublished{sokolov2026fieldweforgot,
  author       = {Sokolov, Anton},
  title        = {The Field We Forgot to Name: A Practitioner's Note on Trust as Physics, Not Plumbing},
  year         = {2026},
  month        = may,
  note         = {Practitioner essay v0.1; submitted to First Monday (under review)},
  institution  = {Tyche Institute}
}

Trust as Data Infrastructure: Local Verification, Portability, and Conservation in Digital Identity and AI Governance

Anton Sokolov · v0.2-bds — Tyche Working Paper (cascade preprint) · May 2026

↳ expanded-of: sokolov2026fieldweforgot; companion-of: sokolov2026pqtft; companion-of: sokolov2026trustfields

Cryptographic trust infrastructure is a data infrastructure in its own right: PKI, trusted lists, wallets, revocation services, audit logs, and attestation formats determine whether public data claims become democratically contestable evidence — not background plumbing.

A conceptual research article bringing cryptographic trust infrastructure into critical data studies and infrastructure studies. Develops a field vocabulary — locality, gauge invariance, conservation — as diagnostics for public evidence infrastructures, and reads X.509 path-construction failures, OCSP outages, trusted-list update gaps, root rotations, and the Estonian ROCA crisis as breakdown moments that reveal democratic vulnerabilities in data governance. The convergence of eIDAS 2.0, the EUDI Wallet, the AI Act, and post-quantum cryptography migration motivates the analysis. The 5,217-word research-article expansion of the shorter First-Monday practitioner essay (companion-of: sokolov2026fieldweforgot); engages boyd & Crawford, Kitchin, Iliadis & Russo, Star–Bowker, Edwards. Deposited 2026-05-26 under CC BY 4.0 as the canonical preprint anchoring the cascade-fallback journal submission packets prepared for JeDEM Reflections (~3,985 words), Big Data & Society (the deposited 5,217-word research article form), and Information Polity (~5,000 words digital-governance adaptation).

doi:10.5281/zenodo.20399235 → concept DOI: 10.5281/zenodo.20399235

@misc{sokolov2026trustasdata,
  author       = {Sokolov, Anton},
  title        = {Trust as Data Infrastructure: Local Verification, Portability, and Conservation in Digital Identity and {AI} Governance},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.2-bds},
  doi          = {10.5281/zenodo.20399235},
  url          = {https://doi.org/10.5281/zenodo.20399235}
}

Trust Fields for Digital Government: Attestable Evidence, Verifier Portability, and Administrative Accountability

Anton Sokolov · v0.1 — under review at Information Polity (IPO-26-0133) · May 2026

↳ companion-of: sokolov2026trustasdata; expanded-of: sokolov2026fieldweforgot

Digital-government trust infrastructure — identity wallets, trusted lists, revocation services, AI-system logs, electronic signatures, preservation services, post-quantum migration plans — should be specified, procured, and supervised as public evidence infrastructure, evaluated against three locally-checkable administrative properties: locality, gauge invariance, and conservation.

The e-government / digital-governance adaptation of the Big Data & Society research-article variant (sokolov2026trustasdata), reframed for the Information Polity audience. Public administration, not critical data studies, is the load-bearing reader: how public bodies should evaluate whether wallet, trusted-list, audit-log, and post-quantum migration infrastructure remains locally verifiable, portable across legitimate substitutions, and durable enough for administrative appeal. The conceptual triad — locality, gauge invariance, conservation — is read against public-key infrastructure episodes (X.509 path-construction failures, OCSP outages, trusted-list update gaps, root rotation, the Estonian ROCA crisis). Five practitioner takeaways are addressed to public managers, policy analysts, and IT architects. Submitted 2026-05-27 to Information Polity (Sage / IOS Press) as ScholarOne manuscript IPO-26-0133; same conceptual argument as the Zenodo-deposited Big Data & Society variant but with distinct framing, audience, and word count (5,119 vs. 5,217). Zenodo record will not be updated while this submission is under Information Polity review (Sage prior-publication policy).

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026trustfieldsdgov,
  author       = {Sokolov, Anton},
  title        = {Trust Fields for Digital Government: Attestable Evidence, Verifier Portability, and Administrative Accountability},
  year         = {2026},
  month        = may,
  note         = {Working paper v0.1; submitted to Information Polity (IPO-26-0133)},
  institution  = {Tyche Institute}
}

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

Anton Sokolov · v1.1 · May 2026

Weeks before the 2 August 2026 enforcement date, deployer obligations surface unevenly across 96 Estonian kratid.ee records, 750 EU procurement notices, and 248 passages from 41 AI policy documents.

An empirical legal-policy working paper measuring how EU AI Act deployer obligations (Articles 10, 12, 13, 14) surface in public documentation: a pre-enforcement evidence-readiness baseline. Submitted to the European Journal of Risk Regulation (manuscript EJRR-2026-0121). v1.1 preserves the empirical findings and tightens the framing; the reproducibility anchor is the OSF mirror.

doi:10.5281/zenodo.20357762 → concept DOI: 10.5281/zenodo.20329188

@misc{sokolov2026deployer,
  author       = {Sokolov, Anton},
  title        = {Public evidence for {AI Act} deployer obligations before enforcement: A baseline from {Estonia}, {EU} procurement, and {AI} policy documents},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v1.1},
  doi          = {10.5281/zenodo.20357762},
  url          = {https://doi.org/10.5281/zenodo.20357762}
}

Cryptographic Attestation for AI Agent Governance under the EU AI Act: A Survey of Approaches and Standards

Anton Sokolov · v2.5 · May 2026

A structured taxonomy of cryptographic attestation approaches across MCP, AAIF, OVERT, and verifiable-credential lines, mapped to the AAL ladder and the four standards bodies that decide the trajectory.

A structured survey of the cryptographic attestation landscape for AI agent governance: regulatory baseline, adversary classes, defensive primitives, design principles, the AAL ladder, the OVERT 1.0 × EATF AEP v1 co-emergence pattern, and the standards-body trajectory at CEN-CENELEC JTC 21, ETSI TC ESI, ISO/IEC JTC 1/SC 42, and IETF SCITT.

doi:10.5281/zenodo.20357731 → concept DOI: 10.5281/zenodo.20185410

@misc{sokolov2026survey,
  author       = {Sokolov, Anton},
  title        = {Cryptographic Attestation for {AI} Agent Governance under the {EU AI Act}: A Survey of Approaches and Standards},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v2.5},
  doi          = {10.5281/zenodo.20357731},
  url          = {https://doi.org/10.5281/zenodo.20357731}
}

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

Anton Sokolov · v1.1 · May 2026

An article-by-article mapping from high-risk AI Act obligations to eIDAS, ETSI EN 319, RFC 3161, RFC 8785, and FIPS 204 primitives — evaluated against an open verifier on an 11-vector corpus, with classical and hybrid (RSA-4096 + ML-DSA-65) measurements.

An article-by-article and layer-by-layer reference mapping from selected high-risk AI Act obligations to cryptographic and trust-service primitives drawn from eIDAS / eIDAS 2.0, the ETSI EN 319 series, IETF RFC 3161 and RFC 8785, and NIST FIPS 204 post-quantum signatures, evaluated against an open-source reference implementation: a worked MCP trace end to end, conformance across two independent verifiers, and a performance measurement of classical and hybrid signing. The journal-form derivative (~9,700 words, OSCOLA-style footnotes) was submitted to JIPITEC — Journal of Intellectual Property, Information Technology and E-Commerce Law on 25 May 2026 (double-blind review).

doi:10.5281/zenodo.20357732 → concept DOI: 10.5281/zenodo.20257971

@misc{sokolov2026eidas,
  author       = {Sokolov, Anton},
  title        = {Operationalizing the {EU AI Act} through {eIDAS} Trust Services Primitives: A Reference Mapping for High-Risk {AI} Systems},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v1.1},
  doi          = {10.5281/zenodo.20357732},
  url          = {https://doi.org/10.5281/zenodo.20357732}
}

eAIDAS as Turning-Point Infrastructure: The Cryptographic Attestation Layer in the AI Revolution's Deployment Phase

Anton Sokolov · v0.1 — Tyche Working Paper · May 2026

↳ companion-of: sokolov2026survey; companion-of: sokolov2026eidas

The EU AI Act, eIDAS 2.0, the ETSI EN 319 series, and NIST's post-quantum standards form the embryo of the attestation infrastructure that Perez's turning-point framework requires — and Estonia holds a first-mover window.

A TEP-theoretic reading of the AI revolution's institutional deficit. Using Carlota Perez's techno-economic paradigm framework and Lema and Perez's (2024) direction-driven extension, the paper argues that cryptographic AI compliance attestation — eAIDAS (electronic AI Declaration and Attestation Services) — is the missing turning-point infrastructure for the AI revolution's deployment phase. The EU AI Act's Art. 47 Declaration of Conformity, eIDAS 2.0 qualified trust services, ETSI EN 319, and NIST FIPS 203/204/205 are the building blocks; the profile work and governance mandate remain undone. Estonia's 25-year e-ID investment, ROCA crisis governance experience, and active EUDIW pilots position it as a plausible first mover in the deployment-phase attestation economy. Deposited as Tyche Working Paper v0.1 on 25 May 2026.

doi:10.5281/zenodo.20379597 → concept DOI: 10.5281/zenodo.20379597

@unpublished{sokolov2026turningpoint,
  author       = {Sokolov, Anton},
  title        = {{eAIDAS} as Turning-Point Infrastructure: The Cryptographic Attestation Layer in the {AI} Revolution's Deployment Phase},
  year         = {2026},
  month        = may,
  institution  = {Tyche Institute},
  note         = {Tyche Working Paper v0.1; DOI: 10.5281/zenodo.20379597},
  doi          = {10.5281/zenodo.20379597},
  url          = {https://doi.org/10.5281/zenodo.20379597}
}

On the Crossroads: Marketplace vs. Distributed Trust in Agent Attestation Frameworks

Anton Sokolov · v0.5 — under venue query (IAB + ACM Queue) · May 2026

Audit the nouns before the APIs: flat conformance contracts, public-key history mirrors, and operator-managed trust anchors substitute for a hosted registry without re-creating its structural-power problems.

A design-rationale essay on why an agent-attestation framework should resist the temptation to grow into a hosted registry. The piece argues that "registry" is not neutral UX vocabulary in trust frameworks: in systems designed for offline verification, a convenient framework-level registry can quietly import authority, lookup-privacy risk, and platform governance into a protocol that was intended to remain locally verifiable. Six plausible user stories motivate a hosted registry; five interlocking traps explain why building it would dissolve EATF's design properties; four distributed-trust substitutions (key-history mirrors, operator-managed anchors, conformance vectors, no framework discovery layer) are offered as the road taken. v0.5 is the author-voiced revision; sent as informal pitch to IAB ([email protected]) and ACM Queue ([email protected]) on 2026-05-26 ahead of CACM Viewpoints and IEEE Computer Perspectives portal submissions.

doi:10.5281/zenodo.20357727 → concept DOI: 10.5281/zenodo.20257933

@misc{sokolov2026crossroads,
  author       = {Sokolov, Anton},
  title        = {On the Crossroads: Marketplace vs. Distributed Trust in Agent Attestation Frameworks},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.5},
  doi          = {10.5281/zenodo.20357727},
  url          = {https://doi.org/10.5281/zenodo.20357727}
}

From Bayesian Knowledge Tracing to Verifiable Educational AI

Anton Sokolov · v1.4 (JLA submission) · May 2026

↳ under peer review at JLA; demo at AIED 2026 (non-archival) is sokolov2026aied

An Agent Trust Framework worked example for an educational AI tutor — event model, four-swimlane architecture, EU AI Act evidence mapping, threat model, and replay / tamper benchmarks against a bundled evaluation harness.

A practical-report case study applying the Agent Trust Framework to a Bayesian Knowledge Tracing tutor for primary-mathematics. The v1.4 packet was submitted to the Journal of Learning Analytics on 2026-05-26 (submission ID 9924, Practical Reports track, double-blind). The submission packet covers the full BKT pedigree (Kolmogorov 1933 → Markov 1913 → Baum & Petrie 1966 → Rabiner 1989 → Corbett & Anderson 1995 → Yudelson et al. 2013) plus Vygotsky 1978 on ZPD, the AI-governance lineage (Mitchell, Gebru, Raji, HLEG, NIST AI RMF), the EU AI Act 2024/1689 with eIDAS 910/2014 + 2024/1183 and ETSI EN 319 421, and the full NIST FIPS 203/204/205 post-quantum suite. Open preprint deposit: Zenodo DOI 10.5281/zenodo.20357766. Open artifact bundle (synthetic BKT trace, replay script, tamper-detection script, schema, threat model): Zenodo DOI 10.5281/zenodo.20273730. (An IACR Cryptology ePrint Archive deposit submission xxxx/109612 was rejected by the archive editors on 2026-05-28 as out-of-scope; the BKT manuscript is AI-governance + attestation infrastructure, not cryptography proper. JLA peer review is unaffected.)

doi:10.5281/zenodo.20357766 → concept DOI: 10.5281/zenodo.20257996

@unpublished{sokolov2026bkt,
  author       = {Sokolov, Anton},
  title        = {From {Bayesian} Knowledge Tracing to Verifiable Educational {AI}: A practical report on evidence packages, teacher oversight, and {AI Act} audit readiness},
  year         = {2026},
  month        = may,
  note         = {Practical Report submitted to the Journal of Learning Analytics on 2026-05-26 (submission ID 9924, double-blind); under peer review. Open preprint deposit: Zenodo DOI 10.5281/zenodo.20357766. Open artifact bundle: Zenodo DOI 10.5281/zenodo.20273730. (IACR ePrint Archive deposit xxxx/109612 rejected 2026-05-28 as out-of-scope for the cryptography archive; JLA peer review unaffected.)},
  institution  = {Tyche Institute}
}

Evidence-Carrying Skill-State Updates: A Verifiable Trace Protocol for Stateful Adaptive ML Systems

Anton Sokolov · v0.2 — submitted to TMLR (submission 9280; under double-blind review) · May 2026

↳ venue-pivot sibling of sokolov2026bkt (JLA Practical Report); cryptographic-protocol angle decoupled from the JLA deployment study; AI-governance / disclosure-to-evidence angle is sokolov2026disclosureToEvidence at AI and Ethics

A verifiable trace protocol for stateful adaptive ML systems that exports canonical, hash-linked, optionally signed and timestamped evidence records — letting a third-party verifier check trace integrity, BKT-style replay correctness, model and policy version binding, oversight-event presence, and offline reproducibility without trusting the live platform.

An ML-methodology contribution that formalises domain-specific verifier predicates for stateful adaptive ML systems, using Bayesian Knowledge Tracing as a worked-example domain rather than as the paper's topic. The protocol defines event syntax, algorithms, and proof sketches for trace tamper evidence, replay correctness, version binding, oversight-event presence, and offline verification under standard cryptographic assumptions (deterministic canonicalisation, collision-resistant hashing, authentic chain heads, existentially unforgeable signatures). A synthetic artefact bundle (one trace, parameters, schema, replay verifier, tamper test) demonstrates the executable claims. Submitted to TMLR after the IACR Cryptology ePrint Archive declined the earlier framing on 2026-05-28 as out-of-scope for cryptography proper; the TMLR re-framing positions the work as ML methodology with correctness-criterion review rather than as a cryptographic-novelty paper. SaTML 2027 remains the cascade-fallback venue when its CFP opens (~September 2026 deadline expected).

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026evidenceCarryingSkillState,
  author       = {Sokolov, Anton},
  title        = {Evidence-Carrying Skill-State Updates: A Verifiable Trace Protocol for Stateful Adaptive {ML} Systems},
  year         = {2026},
  month        = may,
  note         = {Submitted to {Transactions on Machine Learning Research} on 2026-05-28 via OpenReview; under double-blind review. Cascade-fallback: {SaTML 2027} when CFP opens.},
  institution  = {Tyche Institute}
}

From AI Disclosure to Evidence: Audit-Ready Provenance for AI-Mediated Educational Research

Anton Sokolov · v0.2 — submitted to AI and Ethics; Technical Check stage · May 2026

↳ venue-pivot sibling of sokolov2026bkt and sokolov2026evidenceCarryingSkillState; AI-governance / disclosure-to-evidence framing decoupled from the protocol contribution which is at TMLR; pivoted from A&R desk-reject 2026-05-28

A four-layer integrity framework — narrative disclosure / artefact documentation / event provenance / verifiable evidence packages — that operationalises the move from 'AI was used' to 'here is what AI changed and here is the evidence', with explicit boundaries on what evidence-bearing provenance cannot settle (consent, privacy, fairness, educational validity, legal compliance).

A research-integrity framework paper distinguishing four complementary layers of AI accountability in research workflows: narrative disclosure, artefact documentation (model cards / datasheets), event provenance, and verifiable evidence packages (canonical hash-linked records). Adaptive learning with Bayesian Knowledge Tracing is the worked-vignette domain because BKT update is interpretable enough to make the four layers concrete. The contribution is a bounded, claim-by-claim vocabulary for the research-integrity community: evidence packages can support tamper evidence, version binding, replay checks, and oversight-event verification under stated assumptions; they cannot establish consent, privacy compliance, fairness, educational validity, or legal accountability. The article addresses the recent AI and Ethics conversation on accountability infrastructure (Hartmann et al. EU AI audit ecosystem; Donia et al. event-lifecycle framing; Marko et al. causal transparency; Dominguez Castillo workplace accountability; Visave operational transparency) and the parallel Accountability in Research discourse on AI-use disclosure (Resnik & Hosseini; Charbonneau & Zhang; Moffatt & Hall; Suchikova et al.; Ayala & Hervé-Fernández) without dual-submitting the protocol contribution which is described in the sibling sokolov2026evidenceCarryingSkillState paper at TMLR.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026disclosureToEvidence,
  author       = {Sokolov, Anton},
  title        = {From {AI} Disclosure to Evidence: Audit-Ready Provenance for {AI}-Mediated Educational Research},
  year         = {2026},
  month        = may,
  note         = {Submitted to {AI and Ethics} (Springer Nature) on 2026-05-28 via SNAPP, submission UUID ea8c416e-9bd6-46d1-be9b-d5c44abbe662; under double-anonymous review. Pivoted from {Accountability in Research} desk-reject (GACR-2026-0305, Hosseini) on the same day. Cascade: Research Ethics (SAGE) → Big Data \& Society (SAGE) → Internet Policy Review.},
  institution  = {Tyche Institute}
}

Agent Evidence Packages for Offline-Verifiable AI Audit Trails: A Practitioner Reference Architecture with MATx as One Vertical

Anton Sokolov · v0.2 — draft in preparation · May 2026

↳ parallel security-architecture derivative of sokolov2026bkt; non-overlapping with JLA Practical Report (different framing and audience)

A practitioner-facing reference architecture for offline-verifiable evidence packages (RFC 8785 JCS canonicalisation, hash chaining, RFC 3161 timestamping, hybrid Ed25519 + FIPS 204 ML-DSA signatures, two independent verifiers) for any high-risk AI system that produces per-decision evidence, with MATx used as a single illustrative vertical and explicit non-goals around fairness, pedagogical validity, and legal compliance.

A practitioner feature-article derivative of sokolov2026bkt, written for security engineers, CISOs, and compliance leads at organisations deploying high-risk AI. The architecture allows downstream auditors — internal compliance, market-surveillance authorities, end users — to reproduce model decisions, validate timestamps, and detect tampering at row-level granularity without contacting vendor APIs or central registries. MATx (deployed BKT tutor for Estonian school mathematics) appears only as one vertical illustration; the construction is domain-agnostic and operationalises EU AI Act Articles 12, 13, and 14 for any high-risk AI system.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026ieeespMag,
  author       = {Sokolov, Anton},
  title        = {Agent Evidence Packages for Offline-Verifiable {AI} Audit Trails: A Practitioner Reference Architecture with {MATx} as One Vertical},
  year         = {2026},
  month        = may,
  note         = {Practitioner feature article; v0.2 draft in preparation for IEEE Security \& Privacy Magazine (rolling ScholarOne submission). Derivative angle of sokolov2026bkt, security-architecture register, MATx treated as one vertical example only},
  institution  = {Tyche Institute}
}

The Tutor That Can Be Checked: MATx as a Critical Case of Attestation Infrastructure for Educational AI

Anton Sokolov · v0.3 — submitted to AI & Society 2026-05-27 · May 2026

↳ vertical-application of sokolov2026pqtft (in review at Big Data & Society as BDS-26-0556); parallel STS/Bourdieu critical-case angle of sokolov2026bkt; FAccT/AIES 2027 cycle keeps a separate normative-ethical angle

An STS + Bourdieu critical case: the choice between registry-mediated and offline-verifiable evidence formats for high-risk educational AI is a governance choice, not a technical one — it redistributes audit capacity from credentialed gatekeepers to anyone capable of running a verifier, while leaving the social and professional conditions under which that capacity becomes substantive as the harder open question.

An Open Forum article for AI & Society, applying the field-theoretic reading of attestation infrastructure (developed at the general level in sokolov2026pqtft, in review at Big Data & Society as BDS-26-0556) to one specific vertical: high-risk educational AI under the EU AI Act, with MATx as material. The article draws on Bowker & Star, Edwards, Star & Ruhleder, Hughes, Bourdieu, plus Williamson, Selwyn, Ostrom, Benkler, Mitchell, Gebru, Raji, Diakopoulos, Pasquale, Lessig. It proposes the verification commons as a third governance mode beside state mandate and vendor self-attestation, and surfaces the missing observable uptake as a back-question to the general theory.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026aiSociety,
  author       = {Sokolov, Anton},
  title        = {The Tutor That Can Be Checked: {MATx} as a Critical Case of Attestation Infrastructure for Educational {AI}},
  year         = {2026},
  month        = may,
  note         = {Open Forum article; v0.3 packet ready for {AI \& Society} (Springer, Editorial Manager). Vertical-application derivative of sokolov2026pqtft (in review at Big Data & Society as BDS-26-0556); parallel STS/Bourdieu critical-case angle to sokolov2026bkt},
  institution  = {Tyche Institute}
}

Brokered Trust for AI Agents: Lessons from the eIDAS QTSP-Broker Pattern for an Agent Trust Framework

Anton Sokolov · v0.3 — under review (preprint at tyche.institute) · May 2026

↳ identity-broker-layer companion to attestation-evidence-layer work (sokolov2026csur, sokolov2026bkt, OVERT 1.0); BSEADE Paper A at EJRR addresses agent-ecosystem accountability from a different angle

The federation problem agent-trust frameworks treat as future work has already been solved for human principals by QTSP-brokers — single-API façades under eIDAS that fan out to dozens of Qualified Trust Service Providers and national eID schemes, emit ETSI-baseline signature artifacts, and maintain long-term validity. The pattern's primitives (identity/action decoupling, format-baseline reuse, broker-level long-term-validation responsibility, QSCD/QTSP separation) map cleanly onto open agent-trust design questions, with four concrete EATF recommendations.

A practitioner-facing feature article proposing the QTSP-broker pattern as architectural prior art for AI-agent trust frameworks. Audience: PKI engineers, AI architects, compliance leads deploying or evaluating agent-trust systems. The article positions Brokered Trust at the identity-broker layer and does NOT overlap with the attestation-evidence layer covered by OVERT 1.0 or by the IACR ePrint 2026/109612 / CSUR survey work. Closes with an Estonian-regulatory-sandbox pilot proposal scoped to the Accelerate Estonia experimentation framework.

Archival deposit pending editorial response. The full manuscript is held under submission.

@unpublished{sokolov2026brokeredtrust,
  author       = {Sokolov, Anton},
  title        = {Brokered Trust for {AI} Agents: Lessons from the {eIDAS} {QTSP}-Broker Pattern for an Agent Trust Framework},
  year         = {2026},
  month        = may,
  note         = {Practitioner feature article; submitted to IEEE Security \& Privacy Magazine (Regular track, IEEE Author Portal) on 27 May 2026; ScholarOne Manuscript ID pending},
  institution  = {Tyche Institute}
}

The Agent-Ecosystem Attestation Gap: An Empirical Study of Public MCP-Server and Agent-Framework Documentation

Anton Sokolov · v0.1.2 — In Screening (Wiley Security and Privacy) · May 2026

↳ BSEADE Paper B — empirical-evidence companion to sokolov2026bseade (EJRR-2026-0121, Under Review) which establishes the public-sector baseline; the v0.1.2 Related Work also engages the same MCP/agent-security literature surfaced in sokolov2026brokeredtrust

A cleaned top-200 corpus of public GitHub artifacts self-describing as MCP servers, agent frameworks, or agent applications shows a gap between connectivity claims and externally verifiable evidence. 139/200 (69.5%) contained no explicit attestation-adjacent claim. 61/200 (30.5%) contained at least one claim — most often ordinary audit/logging language (47, 23.5%) or software/supply-chain mechanisms (code signing, container signing, SBOM/provenance, cryptographic export). Only 1/200 (0.5%) contained tamper-evident or append-only logging language, 3/200 (1.5%) action-signing language, and 5/200 (2.5%) documented a threat model. The EU AI Act Article 12 logging signal is rare but nonzero: 1/200 = 0.50% with Wilson 95% CI [0.09%, 2.78%]. The contribution is descriptive, not adjudicative.

Descriptive empirical study of how the public agent-ecosystem documents (or fails to document) the evidence that downstream deployers, auditors, and EU AI Act enforcers will need. The paper engages the established literature on software-supply-chain attestation (Newman/Torres-Arias/Samuel), verifiable computation and trusted execution (Parno/Hawblitzel/Sabt), append-only and tamper-evident logging (Crosby & Wallach/Birkholz), empirical software-engineering on OSS security and supply-chain risk (Decan/Ohm/Zahan/Wermke/Ladisa/Peisert), and the emerging LLM-agent / MCP security literature (Greshake/Liu/Ruan/Chan/Shavit/Hou/Yao). Generative-AI assistance was used only in Wiley submission-variant recalibration from author-approved briefs; corpus construction, claim extraction, statistical analysis, and the entire data-producing pipeline are deterministic Python (declared in §3.5 Methods per Wiley AIGC policy).

doi:10.5281/zenodo.20402716 → concept DOI: 10.5281/zenodo.20402715

@unpublished{sokolov2026attestationgap,
  author       = {Sokolov, Anton},
  title        = {The Agent-Ecosystem Attestation Gap: An Empirical Study of Public {MCP}-Server and Agent-Framework Documentation},
  year         = {2026},
  month        = may,
  note         = {Original research article; submitted to Wiley Security and Privacy on 27 May 2026; Manuscript ID 8132248 (In Screening). Replication packet at https://osf.io/qb38p/; preprint at https://doi.org/10.5281/zenodo.20402716},
  institution  = {Tyche Institute},
  doi          = {10.5281/zenodo.20402716}
}

A curated corpus of retracted, withdrawn, and discontinued artifacts in AI governance

Anton Sokolov · v0.4 — under review at Scientific Data · May 2026

↳ parent-of: sokolov2026nekropolisOpacity, sokolov2026nekropolisRipr

A single unified entry schema brings five public negative-results record families — journal retractions, preprint withdrawals, trusted-service lifecycle changes, archived governance repositories, and reviewed benchmark withdrawals — into one cross-domain corpus of 14,923 records, while keeping source-family labels visible so the evidentiary unevenness between families is never silently pooled.

A Scientific Data Data Descriptor for Nekropolis v0.3 — a 14,923-record curated corpus of artifacts retracted, withdrawn, revoked, archived, or discontinued across AI research and adjacent digital-governance infrastructure between 2018 and 2026. Each record carries an artifact category, lifecycle dates, source-stated reason where one exists, a conservative project-inferred cause label, and hashed source pointers. The descriptor defines the schema, source-family gates, validation checks, deposit layout, licensing, and responsible-reuse guidance for the corpus. Companion analyses of public-record opacity (sokolov2026nekropolisOpacity) and AI/ML retraction causes (sokolov2026nekropolisRipr) are separate manuscripts that read this corpus through narrower slices. Deposit at https://zenodo.org/records/20405512 (CC BY 4.0); pipeline bundle and v0.2 lineage files included in the same versioned record.

doi:10.5281/zenodo.20405512 → concept DOI: 10.5281/zenodo.20405511

@misc{sokolov2026nekropolisDataDescriptor,
  author       = {Sokolov, Anton},
  title        = {A Curated Corpus of Retracted, Withdrawn, and Discontinued Artifacts in {AI} Governance},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.3},
  doi          = {10.5281/zenodo.20405512},
  url          = {https://doi.org/10.5281/zenodo.20405512},
  note         = {Data Descriptor; under technical check at Scientific Data (Springer Nature) as ac8a9bf1-914d-4e68-b645-c4696df63c3c}
}

A source-stratified opacity profile of six public negative-results record families

Anton Sokolov · v0.2 — under review · May 2026

↳ sibling-of: sokolov2026nekropolisRipr; child-of: sokolov2026nekropolisDataDescriptor

Five per-family measurements separate six public negative-results record families into three opacity regimes — reason-rich, templated-status, and empty — which a naive cause-presence check collapses into two.

Empirical scientometrics on the Nekropolis v0.2 corpus (17,115 records of retracted, withdrawn, revoked, or archived artifacts in AI research and AI-governance-adjacent public infrastructure). Argues that source-stratified calibration is required before stated reasons can serve as evidence in downstream metascience or governance-evidence systems. Reproducibility pipeline at https://doi.org/10.5281/zenodo.20408462 (CC BY 4.0). Reads the underlying corpus described in sokolov2026nekropolisDataDescriptor.

doi:10.5281/zenodo.20409800 → concept DOI: 10.5281/zenodo.20353904

@misc{sokolov2026nekropolisOpacity,
  author       = {Sokolov, Anton},
  title        = {A Source-Stratified Opacity Profile of Six Public Negative-Results Record Families},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.2},
  doi          = {10.5281/zenodo.20353905},
  url          = {https://doi.org/10.5281/zenodo.20353905}
}

Computer-generated content and the AI/ML retraction record, 2018–2026: a characterization study

Anton Sokolov · v0.2 — submitted to JERHRE (SAGE) as JERHRE-26-0083 on 2026-05-28 (same-day cascade after A&R desk reject) · May 2026

↳ sibling-of: sokolov2026nekropolisOpacity; child-of: sokolov2026nekropolisDataDescriptor

In the AI/ML slice of the public retraction record 2018–2026 (13,502 Retraction Watch entries in the Nekropolis v0.2 corpus), computer-generated content is the largest single inferred cause (34.8%), ahead of compromised peer review (22.0%), plagiarism (18.6%), and editorial process (13.7%); 84% of CGC-retracted papers were published in 2021–2023.

A descriptive, non-adjudicative characterization of the AI/ML slice of the Retraction Watch record, using the Nekropolis v0.2 working corpus (13,502 records). Notices are treated as accountability records, not as adjudications against individual researchers or institutions. Counts are reported on both the conservative title-level subset and the inclusive AI/ML-adjacent slice.

doi:10.5281/zenodo.20409916 → concept DOI: 10.5281/zenodo.20409915

@misc{sokolov2026nekropolisRipr,
  author       = {Sokolov, Anton},
  title        = {Computer-Generated Content and the {AI/ML} Retraction Record, 2018--2026: A Characterization Study},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v0.2},
  doi          = {10.5281/zenodo.20409916},
  url          = {https://doi.org/10.5281/zenodo.20409916},
  note         = {Author Original Manuscript; desk-rejected by Accountability in Research (Taylor \& Francis) as GACR-2026-0305 on 2026-05-28 without peer review; submitted to Journal of Empirical Research on Human Research Ethics (JERHRE, SAGE) the same day as JERHRE-26-0083, Under Review}
}

Computer-generated content as a structural cause of journal retractions, 2020–2026

Anton Sokolov · v0.1.21 — MetaArXiv preprint deposited 2026-05-28 (26zag_v1, awaiting moderator approval); PNAS Brief Report Direct Submission pending Anton's portal click · May 2026

↳ sibling-of: sokolov2026nekropolisRipr; child-of: sokolov2026nekropolisDataDescriptor

In the full 2018–2026 Retraction Watch / Crossref window (47,120 retractions), the share carrying the source-stated AI-content tag ("Computer-Aided Content" or "Computer-Generated Content") rose from below 0.5% before 2020 to a 2023 peak of 40.8%, overwhelmingly Hindawi mass retractions (5,426 of 5,514, 98.4%). The durable finding is the post-2023 record: 22.1% in 2024 across mixed publishers, and 23.6% in 2025 with the Hindawi contribution at zero — the system-level signal persisted past the one-publisher event. Beyond Lei et al. (2024), whose Scientometrics characterization closed at 28 May 2024, the 2024–2026 numbers are the empirical extension.

A successor empirical observation to Fang, Steen and Casadevall (2012 PNAS) for the generative-AI era. Uses the public Retraction Watch / Crossref CSV snapshot (SHA-256 b790b8e3…0d69468) preserved in the Nekropolis negative-results corpus, archival deposit 10.5281/zenodo.20405511. The reading is administrative: computer-generated content is documented as a systemic cause of retraction across the publishing system, not the residue of one publisher's cleanup. The paper is bounded — it does not estimate the prevalence of AI-assisted writing in submitted manuscripts, only what editors and publishers chose to retract under the specific tag.

doi:10.31222/osf.io/26zag_v1 → concept DOI: 10.31222/osf.io/26zag

@unpublished{sokolov2026nekropolisAiContentRegime,
  author       = {Sokolov, Anton},
  title        = {Computer-Generated Content as a Structural Cause of Journal Retractions, 2020--2026},
  year         = {2026},
  month        = may,
  note         = {PNAS Brief Report Direct Submission pending; MetaArXiv preprint deposited 2026-05-28 as 26zag\_v1, https://osf.io/preprints/metaarxiv/26zag\_v1; DOIs 10.31222/osf.io/26zag (concept) / 10.31222/osf.io/26zag\_v1 (v1) activate on moderator approval},
  institution  = {Tyche Institute}
}