Pith. sign in

REVIEW 2 major objections 4 minor 53 references

This paper claims that AI-in-the-loop research can be made auditable by recording every activity, claim, and source in a versioned provenance graph whose two invariants are that no claim is parentless and that only humans may grant the top

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 01:47 UTC pith:OITHM3QF

load-bearing objection A serious, honest engineering contribution with a concrete artifact, but its headline claim is only as strong as the self-report it explicitly disclaims; still deserves a real referee. the 2 major comments →

arxiv 2607.25637 v1 pith:OITHM3QF submitted 2026-07-28 cs.DL cs.AIcs.SI

F(AI)2R: Who Did What, and Who Checked? Verifiable AI Provenance as an Executable Skill

Florian Krebs This is my paper
classification cs.DL cs.AIcs.SI
keywords AI provenancePROV-Overification ladderhuman oversightexecutable skillresearch integrityFAIRmachine-checkable audit
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to show that AI-assisted research can keep a record of itself that later humans and machines can actually check. It defines aiprov, an extension of the PROV-O provenance vocabulary, in which every activity, claim, source, prompt, and agent is logged, and it packages the method as a skill an AI agent operates: the agent records its own work, never fabricates telemetry, and cannot grant the top verification rungs. Two invariants make the record auditable: no claim exists without a generating activity and an attributed agent, and a seven-rung verification ladder reserves the top two rungs for humans. The paper's own production is the case study — 244 activities, 8 claims, 75 sources, 34 human-granted rungs — and the repository is offered as the paper, the process, and the method in one verified artifact. If the approach works, the debate about AI in research stops being about assurances and starts being about records.

Core claim

Provenance can be made the load-bearing constraint of AI-in-the-loop research, not a post-hoc description. aiprov records who did what, from which sources, under which instructions, and with which verification state; prompts are treated as plans, transcripts as parts of the artefact. Two invariants carry the audit: no parentless claim (every claim has a generating activity and an attributed agent), and a seven-rung verification ladder whose top two rungs only humans may grant, with AI-granted promotions there treated as hard failures. The paper demonstrates the scheme on itself: the graph records 244 activities, 8 claims, 75 sources, 34 human-granted rungs, including mishaps and mid-experime

What carries the argument

The verification ladder and its two invariants, carried by aiprov's PROV-O extension with three agent classes (human, AI, tool) and four activity classes (authoring, audit, build, repair). The ladder runs from unverified to human-read; promotions must strictly climb, every promotion is logged as an audit activity, rungs 5 and 6 are human-only, and the conformance validator turns any violation into a hard CI failure. The no-parentless-claim invariant makes every claim traceable to a generating activity and an attributed agent. Commit binding, through a two-commit discipline in which the artefact is committed before the graph entry that names it, gives tamper evidence in both directions.

Load-bearing premise

The whole scheme rests on the agent's own self-reported telemetry: a dishonest or malfunctioning agent could simply not log a session, or log a false one, and nothing in the repository would catch it until an independent witness is added.

What would settle it

Give an AI agent the published skill with instructions to fabricate a short session, logging claims and activities for work it did not do, then push the resulting graph through the validator: if it passes all seven checks and CI stays green, the audit guarantee fails exactly where it claims to hold. A softer check the paper itself proposes is to have an independent model or sampled humans re-verify the eight ai-confirmed claims; widespread disagreement would show that the AI-grantable ceiling is too generous.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any push that violates the invariants — an AI-granted human rung, a parentless claim, a missing grantor — fails validation, making conformance a build-time gate rather than an author's promise.
  • An auditor with a copy of the repository can independently re-run the validator, re-hash vendored artefacts, and re-resolve recorded DOIs; these checks require no trust in the authors' honesty.
  • The same ladder and invariants transfer beyond papers to data pipelines, code, CAD models, and operational decisions in manufacturing, where the human-only rungs can map to release sign-off.
  • Because every rung has a recorded grantor and content hashes are bound to commits, silently swapping an artefact or backdating a record is possible only by rewriting published history, which makes dishonest work loud rather than silent.
  • In the case study, the method's overhead was measurable: roughly one tenth of requests, tokens, and cost went to bookkeeping, with no request mixing bookkeeping and substantive work, giving a concrete price for auditability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the method spreads, the most valuable extension would be provider-signed inference logs — the paper's dual-witness ask — because until then the graph's audit value is capped by the writer's honesty; early adopters should treat it as tamper-evident rather than tamper-proof.
  • The logged novelty-sweep mechanism could generalize into a standard way to make 'no prior work' claims falsifiable: every such claim would carry its recorded search horizon, query terms, and registries, so a sceptic could widen and re-run the sweep.
  • A cheap portability test would be to run the skill in a second domain, such as the ultrasonic-welding data analysis the paper sketches, and compare rung distributions, overhead, and whether the human-only gate actually changes behaviour; the paper itself names this as future work.
  • All eight recorded claims sit at ai-confirmed and were granted by the same agent that wrote the citing sentences, so an immediate and useful bias audit would be independent re-verification of the same sources by a different model or by sampled humans, quantifying self-confirmation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. This paper generalizes the author's earlier F(AI)2R experiment into aiprov, a PROV-O extension for AI-in-the-loop research provenance, and packages it as an executable Agent-Skills bundle. The model defines three agent classes, four activity classes, claims and sources with seven verification rungs, a human-only top of the ladder, and a seven-query conformance validator; the skill bootstraps from an ORCID identifier, scaffolds CI that gates every push, publishes a live preview, and derives an EU AI Act transparency disclosure. The paper is presented as its own case study: the public repository records 244 activities, 8 claims (all at ai-confirmed), 75 sources, 222 commits, and a full transcript, with 34 human-granted rungs. The authors are explicit that this is a demonstration, one operator, one domain, one session, and that several claimed properties are future work.

Significance. If the claims are read at the level the paper ultimately scopes, this is a timely and concrete contribution: it shows how AI-work provenance can be turned into a version-controlled, CI-enforced artefact with content hashes, commit bindings, and an executable validator. The strengths are real: the validator checks are named SPARQL queries; Figure 3 shows the tool refusing AI-granted human rungs and flagging parentless claims; the repository, graph, transcripts, and code are public; the cost, token volumes, and incidents are reported rather than hidden; and the one-witness limitation is discussed explicitly. The originality claim over PROV-AGENT is narrow and honestly bounded. The main weakness is that the strongest advertised property, 'verifiable' / 'machine-checkably auditable', exceeds what the mechanism actually establishes: the validator checks internal consistency of a graph authored by the agent being audited, not the truthfulness or completeness of the underlying telemetry. The paper says this in Section VII, but the abstract and title do not, and one of the two stated invariants is not enforced as hard failure by the seven checks.

major comments (2)
  1. [§IV, 'The validator, exhaustively' and 'Invariant II'] Invariant II is stated as 'Promotions must strictly climb and are themselves logged as audit activities, so every rung a claim or source holds has a recorded grantor.' However, the seven SPARQL checks listed immediately afterward contain no strict-climb check, and the check most directly tied to a missing grantor is described as 'a warning' that flags a rung edited directly into the graph, not a hard failure. A direct edit could place a source at the top of the ladder with a recorded human name but no verifiable promotion chain, and the validator would not block the build. Since the CI gate is the mechanism that makes the invariants 'operational rather than aspirational', this gap is load-bearing. Please either add a hard check that every human-only rung is reachable only through a chain of logged promotion activities with strictly increasing rung positions, or clearly re-scope 'invarian
  2. [Abstract and §VII, 'Self-report honesty'] The title and abstract promise 'Verifiable AI Provenance' and a 'machine-readable audit pass' over every artefact, but Section VII says 'The honesty of self-reported telemetry is not' verifiable and 'the record has one witness.' This is the central assumption of the audit model: all seven validator checks, the commit bindings, the content hashes, and the rung grants operate on a graph written by the agent being audited, so an omitted session, a fabricated activity, or a false token count would leave the graph conformant. The paper's honest discussion of this limitation does not fully neutralize the unqualified wording of the abstract and title. Please either (a) re-scope the headline claim to 'internal consistency checks plus independently re-checkable hashes, DOIs, and validated rung grants', making the one-witness caveat visible in the abstract, or (b) add a concrete adversarial demons
minor comments (4)
  1. [§VI, 'The graph as evaluation object'] The sentence 'The operator reports having checked all sources and claims, and the ladder records which of those checks they chose to formalize as grants' is ambiguous about where that report lives. If it is only in the session transcript, say so; if it is an additional self-report not in the graph, that distinction matters for the auditability discussion.
  2. [§IV, 'The validator, exhaustively'] The statement that 'four checks fail hard' should be reconciled with the preceding list: three invariant-I checks, two invariant-II checks, and two access-gate checks are described, with three of the seven called warnings. A tabular or explicitly labeled list of which checks are hard versus warnings would remove ambiguity.
  3. [§VI, 'Worked examples'] The claim that 'every cited source was checked at the time of writing' is presented as fact, but the supporting evidence is in promotion notes and the transcript. Since the whole audit rests on self-report, consider marking this sentence as 'recorded as' rather than 'was', to keep the epistemic level consistent with Section VII.
  4. [References] Several references are to preprints, submitted works, or non-archival specifications ([26], [40], [45], [43]). This is acceptable for an arXiv paper, but for a journal version please mark their status explicitly in the reference list.

Circularity Check

0 steps flagged

No circularity: the aiprov derivation is definitional and code-enforced; the self-report gap is an acknowledged limitation, not a circular step.

full rationale

The paper's derivation chain is definitional and enforced by code, not derived from its own outputs. The aiprov vocabulary (agents, activities, invariants, ladder rungs) is specified in Section IV, and the conformance validator is seven concrete SPARQL checks whose behavior is demonstrated with injected violations in Fig. 3. The self-application in Section VI is presented as a demonstration ("this is a demonstration, not a study"), not as evidence that the method works beyond its own definitions. The central auditability claim is explicitly qualified: Section VII states "What the auditor can and cannot verify. Hashes, commits, resolving DOIs, rung grants, and the validator's verdicts are checkable by anyone with the repository. The honesty of self-reported telemetry is not; that limit is where the discussion begins." This is a recognized validity gap, not a hidden circular reduction. The dual-witness section further acknowledges "nothing stops a dishonest agent from underreporting usage, omitting sessions, or inventing activities," showing the paper does not launder self-report into proof. The five self-citations (5 of 75 sources) are transparent lineage claims (Section VII, "Transparent self-citation") and are not load-bearing for the method's definitions; the ladder, invariants, and validator are independently defined and tested within this paper. No fitted parameter is renamed as a prediction, and no uniqueness theorem from the authors' prior work is invoked to force the design. Therefore, while the overall audit value depends on an unverified assumption about writer honesty, that is a limitation the paper itself flags, not a circularity in its derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

No new physical entities are invented. The free parameters are measurement-rule choices for the self-report statistics. The axioms are all clearly stated in the text; the single-witness trust assumption is the most fragile. The paper itself names the double-witness fix, but it is future work.

free parameters (2)
  • Request-gap threshold (30 minutes) for active-session clustering = 30 minutes
    The paper uses a fixed 30-minute gap to cluster request timestamps into 'active session time.' Neither this threshold nor its sensitivity is defended. It directly affects the 12.7h active-time measurement.
  • Word-count exclusions for direction volume = excludes tool results, hook messages, skill loads, pasted documents over 500 words
    The 2,112-word direction count depends on exclusion rules that are disclosed in the text but not justified against alternatives. A different exclusion cutoff would change the headline leverage numbers. However, they are stated with the number, so a sceptic can recompute; this is a weak free parameter.
axioms (3)
  • domain assumption PROV-O is a suitable substrate for per-activity collaboration provenance (no parentless claim, human-only rungs).
    The paper assumes that extending PROV-O is the right way to model AI provenance; it cites W3C PROV-O and PROV-AGENT. This is a design choice, not a logical theorem. It is reasonable but load-bearing: the entire graph model stands on PROV-O's expressiveness.
  • domain assumption The session record's provider-reported usage blocks are trustworthy.
    The token counts and 'cost computed at list price' rely on the provider's per-request usage blocks. The paper itself flags the single-witness problem: 'The honesty of self-reported telemetry is not [verifiable].' This is the paper's own stated weakest assumption.
  • domain assumption Git history + sha256 hashes constitute tamper evidence.
    The commit-binding argument assumes rewriting published git history is loud and expensive. Under typical GitHub hosting this is true, but it is not mathematically guaranteed (force-push rewrite is possible; the paper acknowledges 'neither is silent,' not 'neither is possible').

pith-pipeline@v1.3.0-alltime-deepseek · 18975 in / 7905 out tokens · 67372 ms · 2026-08-01T01:47:28.240251+00:00 · methodology

0 comments
read the original abstract

F(AI)2R is FAIR research with AI in the loop, twice: an AI-assisted authoring pass and a machine-readable audit pass over every artefact. AI systems now draft, refactor, and verify research artefacts, yet their contributions are rarely recorded in a form a later human or machine can audit. Building on the original F(AI)2R experiment, we generalize its provenance model beyond scholarly writing into aiprov, a PROV-O extension covering any AI-in-the-loop artefact, and we package the method as an executable skill that an AI agent operates itself: setup asks the human operator for their ORCID ID, resolves their identity from the public registry, and scaffolds continuous integration that gates every push on graph conformance and publishes the current build of this very paper. The paper is its own case study. Every activity, claim, and source in its production is recorded in the repository's provenance graph under two invariants: no parentless claim, and verification rungs that only humans may grant.

Figures

Figures reproduced from arXiv: 2607.25637 by Florian Krebs.

Figure 1
Figure 1. Figure 1: Core aiprov vocabulary as an extension of PROV-O. Every class sub [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The verification ladder. Every promotion is logged as an audit activity; [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Success and failure, captured from the tooling running against a [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The skill’s operating loop from the single setup question to the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The continuously rebuilt preview of this paper (screenshot, 2026-07-28), republished to a stable URL after every editing round. Left: commit, build [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The dashboard rendered from provenance.ttl (screenshot, 2026- 07-28). Top: header, graph totals, agents, and the start of the activity ledger; the token and cost tiles sum per-activity records (two aggregate backfills, roughly half the computed cost), so the session-record aggregates of Table II are the authoritative figures. Bottom: the literature rung distribution, most sources operator-verified, the top… view at source ↗
Figure 7
Figure 7. Figure 7: Two incidents from this experiment, exactly as the graph represents them (node contents abridged from real triples). Left, a mishap: the agent [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Method overhead as shares of the recorded session and repository, [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 11 canonical work pages

  1. [1]

    The FAIR guiding principles for scientific data management and stewardship,

    M. D. Wilkinson, M. Dumontieret al., “The FAIR guiding principles for scientific data management and stewardship,”Scientific Data, vol. 3, no. 1, Mar. 2016. [Online]. Available: http://dx.doi.org/10.1038/sdata. 2016.18

    doi:10.1038/sdata 2016

  2. [2]

    Opinion paper: “so what if ChatGPT wrote it?

    Y . K. Dwivediet al., “Opinion paper: “so what if ChatGPT wrote it?” multidisciplinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy,”International Journal of Information Management, vol. 71, p. 102642, Aug. 2023. [Online]. Available: http://dx.doi.org/10.1016/j. ijinfomgt.2023.102642

    arXiv 2023

  3. [3]

    ChatGPT: five priorities for research,

    E. A. M. van Dis, J. Bollenet al., “ChatGPT: five priorities for research,”Nature, vol. 614, no. 7947, pp. 224–226, Feb. 2023. [Online]. Available: http://dx.doi.org/10.1038/d41586-023-00288-7

    work page doi:10.1038/d41586-023-00288-7 2023

  4. [4]

    Model cards for model reporting,

    M. Mitchell, S. Wuet al., “Model cards for model reporting,” inProceedings of the Conference on Fairness, Accountability, and Transparency, ser. FAT* ’19. ACM, Jan. 2019, pp. 220–229. [Online]. Available: http://dx.doi.org/10.1145/3287560.3287596

    arXiv 2019

  5. [5]

    Academic research in the 21st century: Maintaining scientific integrity in a climate of perverse incentives and hypercompetition,

    M. A. Edwards and S. Roy, “Academic research in the 21st century: Maintaining scientific integrity in a climate of perverse incentives and hypercompetition,”Environmental Engineering Science, vol. 34, no. 1, pp. 51–61, Jan. 2017. [Online]. Available: http: //dx.doi.org/10.1089/ees.2016.0223

    arXiv 2017

  6. [6]

    The strain on scientific publishing,

    M. A. Hanson, P. G. Barreiro, P. Crosetto, and D. Brockington, “The strain on scientific publishing,”Quantitative Science Studies, vol. 5, no. 4, pp. 823–843, 2024. [Online]. Available: http: //dx.doi.org/10.1162/qss_a_00327

    doi:10.1162/qss_a_00327 2024

  7. [7]

    Obscurity is dead — proprietary by design, open by AI,

    F. Krebs, “Obscurity is dead — proprietary by design, open by AI,” Research repository with evidentiary artifacts. [Online]. Available: https://github.com/noheton/Obscurity-Is-Dead

  8. [8]

    F(AI) 2R: FAIR research with AI in the loop, twice,

    ——, “F(AI) 2R: FAIR research with AI in the loop, twice,” Research repository: manuscript, provenance graph, transcripts. [Online]. Available: https://github.com/noheton/f-ai-r

  9. [9]

    Experimental fork of shepard,

    ——, “Experimental fork of shepard,” Development/research workspace with experimental v2 endpoints. [Online]. Available: https://github.com/ noheton/shepard

  10. [10]

    PROV-O: The PROV ontology,

    T. Lebo, S. Sahoo, and D. McGuinness, “PROV-O: The PROV ontology,” W3C Recommendation, Apr. 2013, accessed 2026-07-24. [Online]. Available: https://www.w3.org/TR/prov-o/

    2013

  11. [11]

    The open provenance model core specification (v1.1),

    L. Moreau, B. Cliffordet al., “The open provenance model core specification (v1.1),”Future Generation Computer Systems, vol. 27, no. 6, pp. 743–756, Jun. 2011. [Online]. Available: http://dx.doi.org/10.1016/j.future.2010.07.005

    work page doi:10.1016/j.future.2010.07.005 2011

  12. [12]

    PA V ontology: provenance, authoring and versioning,

    P. Ciccarese, S. Soiland-Reyes, K. Belhajjame, A. J. Gray, C. Goble, and T. Clark, “PA V ontology: provenance, authoring and versioning,” Journal of Biomedical Semantics, vol. 4, no. 1, p. 37, 2013. [Online]. Available: http://dx.doi.org/10.1186/2041-1480-4-37

    work page doi:10.1186/2041-1480-4-37 2013

  13. [13]

    An overview of provenance and its use cases,

    C. Haupt, “An overview of provenance and its use cases,” RDA Deutschland Tagung 2022; DLR elib 185438, 2022. [Online]. Available: https://elib.dlr.de/185438/

    2022

  14. [14]

    How can we ensure visibility and diversity in research contributions? how the contributor role taxonomy (CRediT) is helping the shift from authorship to contributorship,

    L. Allen, A. O’Connell, and V . Kiermer, “How can we ensure visibility and diversity in research contributions? how the contributor role taxonomy (CRediT) is helping the shift from authorship to contributorship,”Learned Publishing, vol. 32, no. 1, pp. 71–74, Jan

  15. [15]

    ANSI/NISO Z39.104-2022, CRediT, Contributor Roles Taxonomy,

    “ANSI/NISO Z39.104-2022, CRediT, Contributor Roles Taxonomy,”

    2022

  16. [16]

    Datasheets for datasets,

    T. Gebru, J. Morgensternet al., “Datasheets for datasets,” Communications of the ACM, vol. 64, no. 12, pp. 86–92, Nov

  17. [17]

    Packaging research artefacts with RO-Crate,

    S. Soiland-Reyes, P. Seftonet al., “Packaging research artefacts with RO-Crate,”Data Science, vol. 5, no. 2, pp. 97–138, Jan. 2022. [Online]. Available: http://dx.doi.org/10.3233/ds-210053

    doi:10.3233/ds-210053 2022

  18. [18]

    PROV-AGENT: Unified provenance for tracking AI agent interactions in agentic workflows,

    R. Souza, A. Gueroudji, S. DeWitt, D. Rosendo, T. Ghosal, R. Ross, P. Balaprakash, and R. F. Da Silva, “PROV-AGENT: Unified provenance for tracking AI agent interactions in agentic workflows,” in2025 IEEE International Conference on eScience (eScience). IEEE, Sep. 2025, pp. 467–473. [Online]. Available: http://dx.doi.org/10.1109/escience65000.2025.00093

    arXiv 2025

  19. [19]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” arXiv preprint arXiv:1706.03762, 2017. [Online]. Available: https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2017

  20. [20]

    Artificial hallucinations in ChatGPT: Implications in scientific writing,

    H. Alkaissi and S. I. McFarlane, “Artificial hallucinations in ChatGPT: Implications in scientific writing,”Cureus, Feb. 2023. [Online]. Available: http://dx.doi.org/10.7759/cureus.35179

    doi:10.7759/cureus.35179 2023

  21. [21]

    The anatomy of a nanopublication,

    P. Groth, A. Gibson, and J. Velterop, “The anatomy of a nanopublication,”Information Services and Use, vol. 30, no. 1-2, pp. 51–56, Feb. 2010. [Online]. Available: http://dx.doi.org/10.3233/ ISU-2010-0613

    2010

  22. [22]

    Micropublications: a semantic model for claims, evidence, arguments and annotations in biomedical communications,

    T. Clark, P. N. Ciccarese, and C. A. Goble, “Micropublications: a semantic model for claims, evidence, arguments and annotations in biomedical communications,”Journal of Biomedical Semantics, vol. 5, no. 1, p. 28, 2014. [Online]. Available: http://dx.doi.org/10.1186/ 2041-1480-5-28

    2014

  23. [23]

    Content credentials: C2PA technical specification,

    Coalition for Content Provenance and Authenticity, “Content credentials: C2PA technical specification,” Specification v2.4, 2025, accessed 2026- 07-28. [Online]. Available: https://c2pa.org/specifications/

    2025

  24. [24]

    in-toto: Providing farm-to-table guarantees for bits and bytes,

    S. Torres-Arias, H. Afzali, T. K. Kuppusamy, R. Curtmola, and J. Cap- pos, “in-toto: Providing farm-to-table guarantees for bits and bytes,” in28th USENIX Security Symposium, 2019. [Online]. Available: https: //www.usenix.org/conference/usenixsecurity19/presentation/torres-arias

    2019

  25. [25]

    Open collaborative writing with Manubot,

    D. S. Himmelstein, V . Rubinettiet al., “Open collaborative writing with Manubot,”PLOS Computational Biology, vol. 15, no. 6, p. e1007128, Jun. 2019. [Online]. Available: http://dx.doi.org/10.1371/journal.pcbi. 1007128

    doi:10.1371/journal.pcbi 2019

  26. [26]

    An uncertainty-aware provenance framework for enhanced traceability in engineering systems,

    D. Valente, A. Schäfer, E. Tasdemir, R. Hoppe, O. Bertram, and F. Dressel, “An uncertainty-aware provenance framework for enhanced traceability in engineering systems,” DLR elib 224241; submitted to IEEE Aerospace and Electronic Systems Magazine, 2026. [Online]. Available: https://elib.dlr.de/224241/

    2026

  27. [27]

    Better architecture, better software, better research,

    S. Druskat, N. U. Eisty, R. Chisholm, N. P. Chue Hong, R. C. Cocking, M. B. Cohen, M. Felderer, L. Grunske, S. A. Harris, W. Hasselbring, T. Krause, J. Linxweiler, and C. C. Venters, “Better architecture, better software, better research,”Computing in Science & Engineering, vol. 27, no. 2, pp. 45–57, Apr. 2025. [Online]. Available: http://dx.doi.org/10.11...

    arXiv 2025

  28. [28]

    Details of the asset administration shell. Part 1 — the exchange of information between partners in the value chain of Industrie 4.0 (Version 3.0RC02),

    Plattform Industrie 4.0, “Details of the asset administration shell. Part 1 — the exchange of information between partners in the value chain of Industrie 4.0 (Version 3.0RC02),” Federal Ministry for Economic Affairs and Climate Action (BMWK), specification, 2022. [Online]. Available: https://www.plattform-i40.de/IP/Redaktion/EN/Downloads/ Publikation/Det...

    2022

  29. [29]

    Specification of the asset administration shell. Part 1: Metamodel,

    Industrial Digital Twin Association (IDTA), “Specification of the asset administration shell. Part 1: Metamodel,” IDTA-01001. [Online]. Available: https://doi.org/10.62628/idta.01001-3-1-1

    work page doi:10.62628/idta.01001-3-1-1

  30. [30]

    Helmholtz metadata collaboration (HMC)

    Helmholtz Association, “Helmholtz metadata collaboration (HMC).” [Online]. Available: https://helmholtz-metadaten.de/

  31. [31]

    HMC conference 2025 — book of abstracts,

    Helmholtz Metadata Collaboration, “HMC conference 2025 — book of abstracts,” 2025. [Online]. Available: https://helmholtz-metadaten.de/ storage/2298/HMC-Conference_2025_Book-of-Abstracts.pdf

    2025

  32. [32]

    Fake data, paper mills, and their authors: The International Journal of Cancer reacts to this threat to scientific integrity,

    S. Heck, F. Bianchiniet al., “Fake data, paper mills, and their authors: The International Journal of Cancer reacts to this threat to scientific integrity,”International Journal of Cancer, vol. 149, no. 3, pp. 492–493, Apr. 2021. [Online]. Available: http://dx.doi.org/10.1002/ijc.33604

    work page doi:10.1002/ijc.33604 2021

  33. [33]

    GPT-fabricated scientific papers on Google Scholar: Key features, spread, and implications for preempting evidence manipulation,

    J. Haider, K. R. Söderström, B. Ekström, and M. Rödl, “GPT-fabricated scientific papers on Google Scholar: Key features, spread, and implications for preempting evidence manipulation,”Harvard Kennedy School Misinformation Review, Sep. 2024. [Online]. Available: http://dx.doi.org/10.37016/mr-2020-156

    work page doi:10.37016/mr-2020-156 2024

  34. [34]

    Fabrication and errors in the bibliographic citations generated by ChatGPT,

    W. H. Walters and E. I. Wilder, “Fabrication and errors in the bibliographic citations generated by ChatGPT,”Scientific Reports, vol. 13, no. 1, Sep. 2023. [Online]. Available: http://dx.doi.org/10.1038/ s41598-023-41032-5

    2023

  35. [35]

    AI models collapse when trained on recursively generated data,

    I. Shumailov, Z. Shumaylovet al., “AI models collapse when trained on recursively generated data,”Nature, vol. 631, no. 8022, pp. 755–759, Jul. 2024. [Online]. Available: http://dx.doi.org/10.1038/ s41586-024-07566-y

    2024

  36. [36]

    Bröder, G

    J. Bröder, G. Preuß, F. D’Mello, S. Fathalla, V . Hofmann, and S. Sandfeld,The Helmholtz Knowledge Graph: Driving the Transition Towards a FAIR Data Ecosystem in the Helmholtz Association. Springer Nature Switzerland, 2025, pp. 183–187. [Online]. Available: http://dx.doi.org/10.1007/978-3-031-78952-6_23

    work page doi:10.1007/978-3-031-78952-6_23 2025

  37. [37]

    Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act),

    European Parliament and Council of the European Union, “Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act),” Official Journal of the European Union, L series, 2024, accessed 2026-07-24. [Online]. Available: https://eur-lex.europa.eu/eli/reg/2024/ 1689/oj

    2024

  38. [38]

    Investigation of power and amplitude control in continuous ultrasonic welding of unidirectional CFRPs: A comparative study,

    M. Janek, D. Görick, L. Larsen, S. Jarka, and M. Kupke, “Investigation of power and amplitude control in continuous ultrasonic welding of unidirectional CFRPs: A comparative study,”Composites Part A: Applied Science and Manufacturing, vol. 199, p. 109194, Dec

  39. [39]

    Claude API pricing,

    Anthropic, “Claude API pricing,” Developer documentation, 2026, accessed 2026-07-28. [Online]. Available: https://platform.claude.com/ docs/en/about-claude/pricing

    2026

  40. [40]

    Doctors vs. algorithms: Physicians, too, struggle to learn from evidence that contradicts AI suggestions,

    A. Vinas, F. Blanco, and H. Matute, “Doctors vs. algorithms: Physicians, too, struggle to learn from evidence that contradicts AI suggestions,” PLOS Digital Health, vol. 5, no. 7, p. e0001490, Jul. 2026. [Online]. Available: http://dx.doi.org/10.1371/journal.pdig.0001490

    work page doi:10.1371/journal.pdig.0001490 2026

  41. [41]

    Recommendation of the Council on Artificial Intelligence (OECD/LEGAL-0449),

    OECD, “Recommendation of the Council on Artificial Intelligence (OECD/LEGAL-0449),” Adopted May 2019, amended May 2024, 2024, accessed 2026-07-27. [Online]. Available: https://oecd.ai/en/ ai-principles

    2019

  42. [42]

    AI incidents and hazards monitor (AIM),

    OECD.AI Policy Observatory, “AI incidents and hazards monitor (AIM),” 2024, accessed 2026-07-27. [Online]. Available: https: //oecd.ai/en/incidents-methodology

    2024

  43. [43]

    Agent skills: An open format for extending AI agent capabilities,

    “Agent skills: An open format for extending AI agent capabilities,” https: //agentskills.io/, 2025, specification and adopter list; originally developed by Anthropic and released as an open standard

    2025

  44. [44]

    EASA artificial intelligence concept paper issue 2: Guidance for level 1 & 2 machine-learning applications,

    European Union Aviation Safety Agency, “EASA artificial intelligence concept paper issue 2: Guidance for level 1 & 2 machine-learning applications,” https://www.easa.europa.eu/en/document-library/ general-publications/easa-artificial-intelligence-concept-paper-issue-2, Mar. 2024

    2024

  45. [45]

    Homogenizing effect of a large language model (LLM) on creative diversity: An empirical comparison of human and ChatGPT writing, preprint,

    K. Moon, A. Green, and K. Kushlev, “Homogenizing effect of a large language model (LLM) on creative diversity: An empirical comparison of human and ChatGPT writing, preprint,” Aug. 2024. [Online]. Available: http://dx.doi.org/10.31234/osf.io/8p9wu

    work page doi:10.31234/osf.io/8p9wu 2024

  46. [46]

    Copyright and artificial intelligence, part 2: Copyrightability,

    U.S. Copyright Office, “Copyright and artificial intelligence, part 2: Copyrightability,” Report of the Register of Copyrights, https://www. copyright.gov/ai/, Jan. 2025

    2025

  47. [47]

    Gesetz über Urheberrecht und ver- wandte Schutzrechte (UrhG), § 2 Geschützte Werke,

    Bundesrepublik Deutschland, “Gesetz über Urheberrecht und ver- wandte Schutzrechte (UrhG), § 2 Geschützte Werke,” https://www. gesetze-im-internet.de/urhg/__2.html

  48. [48]

    shepard: storage for heterogeneous product and research data,

    T. Haase, R. Glück, P. Kaufmann, and M. Willmeroth, “shepard: storage for heterogeneous product and research data,” Zenodo, v5.1.2, DLR, Dec. 2025. [Online]. Available: https://doi.org/10.5281/zenodo. 17897485

    doi:10.5281/zenodo 2025

  49. [49]

    Project MEMAS: Integrated data management for additive manufacturing enabling high- fidelity modeling,

    N. Unger, P. Kamble, M. Vinot, and R. Glück, “Project MEMAS: Integrated data management for additive manufacturing enabling high- fidelity modeling,” HMC Conference 2024, 2024. [Online]. Available: https://elib.dlr.de/208404/

    2024

  50. [2019]

    Available: http://dx.doi.org/10.1002/leap.1210

    [Online]. Available: http://dx.doi.org/10.1002/leap.1210

    doi:10.1002/leap.1210

  51. [2021]

    Available: http://dx.doi.org/10.1145/3458723

    [Online]. Available: http://dx.doi.org/10.1145/3458723

    doi:10.1145/3458723

  52. [2022]

    Available: http://dx.doi.org/10.3789/ansi.niso.z39

    [Online]. Available: http://dx.doi.org/10.3789/ansi.niso.z39. 104-2022

    work page doi:10.3789/ansi.niso.z39 2022

  53. [2025]

    Available: http://dx.doi.org/10.1016/j.compositesa.2025

    [Online]. Available: http://dx.doi.org/10.1016/j.compositesa.2025. 109194

    work page doi:10.1016/j.compositesa.2025 2025