The Governance Layer · Clause Library · Instrument 03
Model provenance representations.
An institution deploying a model it did not build is in the position of a producer who has not cleared chain of title. The work functions, the release is scheduled, and the question of what rights actually run through to the party bearing the consequence has not been asked in writing by anyone. In film and television that question has a settled answer, an established documentary practice, and an insurance market that will not bind without it. In model deployment it has none of the three.
This is a set of representations rather than a covenant. Covenants promise future conduct; representations state facts at a point in time, are qualified by a disclosure schedule, are brought down at closing or at each renewal, and are backed by indemnity. That structure is the right one here, because provenance is a question about what is already true.
It is the third instrument in the clause library, and it serves MARQUE Identity — the question names chain of title for the model itself as largely unestablishable today — together with the third-party dependency dimension of KYA.
What the language derives from.
- Chain of title practice. Motion picture and television financing has for decades required documented chain of title as a condition of errors-and-omissions coverage and of distribution. The structure — trace every contributing right from origin to the party who will be sued, and schedule every gap — is directly transferable, and it is the closest mature analogue to model provenance that exists in commercial practice.
- Provider disclosure obligations. The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) imposes obligations on providers of general-purpose AI models from August 2, 2025, including technical documentation, information to be made available to downstream providers, a policy to comply with EU copyright law, and a sufficiently detailed publicly available summary of the content used for training. Where a provider is subject to these obligations, the deploying institution's representations can be grounded in disclosures the provider is already required to make.
- Third-party risk management expectations. The Interagency Guidance on Third-Party Relationships: Risk Management, issued by the Federal Reserve, FDIC and OCC in June 2023, sets out due diligence, contracting and monitoring expectations across the life cycle of a third-party relationship. The deploying institution remains responsible for compliance whether or not it built the capability.
- Contractual provisions for technology dependencies. Article 30 of the EU Digital Operational Resilience Act (Regulation (EU) 2022/2554, applicable from January 17, 2025) prescribes terms including audit and access rights, incident reporting, and exit arrangements. Those are the rights an institution needs in order to be able to give the representations below at all.
- Component disclosure. The software bill of materials practice established for federal software procurement following Executive Order 14028 (May 2021) supplies the drafting model for an enumerated dependency schedule. The unresolved question is what the equivalent enumeration contains for a model, and the schedule below is one answer.
Form of representations.
The Representing Party is the institution deploying the System. Each representation is given as of [date] and, where a bringdown applies, as of each subsequent [date]. Exceptions are set out in the Disclosure Schedule and, where scheduled, qualify the representation.
1. Identity and version.
The System is identified at Schedule A, which states for each model in use: the provider; the model name and version identifier; the date the version was placed into production; and the date of any prior version in use during the [period] preceding the date of these representations. No model is in production use that is not listed at Schedule A.
2. The contractual chain.
The Representing Party holds a current written agreement with each provider listed at Schedule A, or with an intermediary through whom access is obtained, and Schedule B identifies each agreement in the chain from the model provider to the Representing Party. To the Representing Party's knowledge, no agreement in that chain has been terminated, suspended, or materially amended in a manner that would restrict the use described at Schedule A.
3. Rights in outputs.
Under the agreements identified at Schedule B, the Representing Party's position as to ownership of and rights in the outputs of the System is as stated at Schedule C, and the provider asserts no rights in those outputs except as stated there. The Representing Party makes no representation as to the copyrightability of any output.
4. Use of inputs and outputs by the provider.
Under the agreements identified at Schedule B, the provider is [permitted / not permitted] to use inputs submitted by the Representing Party or outputs generated for it for the training or improvement of any model, and any exceptions are stated at Schedule C. Where the position differs between products, tiers, or endpoints in use, each is stated separately.
5. Training corpus.
The disclosures made by each provider as to the content used to train the model are identified at Schedule D, together with the date of each disclosure. The Representing Party has no knowledge of any fact indicating that the content used to train a model listed at Schedule A was obtained or used in a manner giving rise to a claim that would materially affect the Representing Party's use of the System. The Representing Party has not independently verified any provider's disclosure and does not represent its completeness or accuracy.
6. Adaptation by the Representing Party.
Schedule E identifies each fine-tuning, adaptation, distillation, or continued training operation performed by or for the Representing Party, the data used in each, and the basis on which the Representing Party holds rights in that data sufficient for the use made of it. The Representing Party has the right to use each data set for the purpose for which it was used.
7. Retrieval corpus.
Schedule F identifies each source from which the System retrieves content at inference time, the basis on which the Representing Party accesses each source, and whether the terms governing that source permit the use made of retrieved content. Where a source is accessed under terms that do not expressly address retrieval by an automated system, that fact is stated at Schedule F rather than resolved.
8. Dependencies.
Schedule G enumerates each third-party component on which the System depends in production, including orchestration frameworks, vector stores, evaluation and observability services, and any upstream autonomous system whose output the System consumes. For each, the schedule states the provider, the version, and whether the Representing Party holds audit rights.
9. Audit and reproduction rights.
The Representing Party holds, under the agreements at Schedule B, the rights stated at Schedule H to obtain documentation, to receive notice of material change to a model, to receive incident information, and to obtain assistance in reconstructing a past output. Where the Representing Party does not hold a right necessary to satisfy an obligation it has assumed elsewhere, that gap is identified at Schedule H.
10. Change and deprecation.
Schedule I identifies each notice received from a provider during the [period] preceding these representations of a model change, deprecation, retirement, or material behavioral modification affecting a model at Schedule A, and the Representing Party's response to each.
11. Incidents.
Schedule J identifies each adverse outcome involving the System during the [period] preceding these representations, its classification, the remediation undertaken, and whether the matter is closed. The absence of an entry means no such outcome occurred, not that none was recorded.
12. Compliance with provider terms.
The Representing Party's use of each model listed at Schedule A complies in all material respects with the terms of the agreements at Schedule B, and no provider has notified the Representing Party of any asserted breach during the [period] preceding these representations.
Drafting notes.
On knowledge qualifiers, and why they are the finding. Several representations above are qualified by knowledge, and one — the training corpus — expressly disclaims verification. That is not defensive drafting; it is an accurate statement of the current market. No deploying institution can independently verify what a frontier model was trained on. The useful consequence is diagnostic: the set of representations an institution cannot give unqualified is a precise map of what its providers decline to disclose, and assembling that map across providers is a publishable finding in its own right.
On Schedule C and the separation of ownership from copyrightability. Representation 3 states the contractual position and expressly declines to represent copyrightability. Those are different questions, and conflating them creates a representation the Representing Party cannot support. The United States Copyright Office has maintained a human authorship requirement in its registration guidance published March 16, 2023, and the boundary continues to develop; a contractual allocation of whatever rights exist is the furthest a party can responsibly go.
On Representation 7, the least drafted and most exposed. Retrieval pulls third-party content into the System at inference time, and the terms governing most sources were written before automated retrieval by an inference system was contemplated. Almost no institution has examined the rights position of its retrieval corpus. The form requires the question to be answered on a schedule rather than assumed, and where the terms are silent it requires the silence to be recorded rather than construed. Silence recorded is a disclosed risk; silence construed is an unbudgeted one.
On Representation 9 and the gap disclosure. The instruction to identify rights the Representing Party does not hold is unusual in a representation set and is the most valuable line in this instrument. An institution that has covenanted elsewhere to reconstruct a past decision, but holds no right to obtain the provider's assistance in doing so, has assumed an obligation it cannot perform. Surfacing that at the schedule level is cheaper than discovering it during an examination.
On the bringdown interval. Model versions change on the provider's schedule, not the institution's. A representation set brought down annually will be materially stale for most of the year. The practical interval is tied to change notice under Representation 10 rather than to the calendar.
The negative space.
- Indemnity scope. Several providers now offer indemnities in respect of third-party intellectual property claims arising from outputs. Whether those indemnities survive fine-tuning, retrieval augmentation, or use in an autonomous loop has no settled answer, and this form does not supply one.
- Provenance across agent boundaries. Where the System consumes the output of another institution's autonomous system, the provenance chain leaves the Representing Party's knowledge entirely at that boundary, and no instrument establishes what may be relied on.
- Verification. There is no available mechanism by which a deploying institution can verify a provider's training disclosures. Until one exists, every representation in this area is a representation about disclosures received rather than about facts.
- Materiality thresholds. No market practice establishes what constitutes a material model change for notice purposes, which means Representation 10 currently depends on each provider's own definition.
Relationship to the framework.
This instrument serves MARQUE Identity and supplies the documentary basis for the KYA third-party dependency dimension. An institution holding a completed representation set with populated schedules, brought down on a stated interval, together with the agreements at Schedule B, is evidenced on that dimension rather than documented. Schedules A, G and I also support the agent inventory dimension, and Schedule J the incident dimension.
Completing Schedule C and Schedule H across an institution's providers produces, as a by-product, the comparative tabulation of model provider terms that the market currently lacks.
This is an editorial form with commentary, published in the tradition of model agreements and form books. It is not legal advice, not a compliance determination, not a certification, and not a representation that any provision here is enforceable in any jurisdiction. It should be reviewed by counsel and adapted before use.