Third-Party and Open-Weight Model Provenance Sheet

How to use this document

• What it is for. This sheet documents, for each third-party AI model the organization uses through an API or hosts on its infrastructure, where it comes from, under which license it is used, which legal or contractual restrictions may apply to it because of its origin, where data is processed under each hosting mode, whether the artifact is intact and how it behaves in testing. It is especially useful for open-weight models, which have no vendor to answer a due diligence questionnaire.
• Who should own it. Information security completes and keeps the sheet; legal completes the license and restrictions sections; the system owner provides the use case and signs the decision. Enter the function in [RESPONSIBLE FUNCTION] and the person in [OWNER].
• Decisions to take first. Decide from which risk level the sheet is required for models used through an API (recommended: Medium and High), which tools you will use to verify hashes and scan files, which reference model you will use to compare behaviors and who approves under each approval lane.
• One sheet per artifact. Complete one sheet for each combination of model, version and variant (for example, a different quantization or fine-tune is another sheet). Record its identifier in the matching entry of the "AI System Inventory".
• Fields to complete. Replace the bracketed fields. Where the sheet asks for the content of a license or a law, copy it from the current official source: this template does not reproduce licenses or laws.
• Review. Ask legal or your advisor to review Sections 4, 5 and 9, and the data protection officer (or the country's equivalent role) to review the residency matrix in Section 6.
• Approval and filing. The decision in Section 10 is taken by whoever the approval lane for the risk level designates. File the sheet with the evidence and review it again at the frequency of its risk level.

Module selection

Profile or level	Sections to keep
P2 Uses third-party models through APIs or RAG	Sections 1 to 6, 8, 10 and 11
P3 Hosts models, including open-weight models	All sections except 9, if it does not fine-tune models
P4 Trains or fine-tunes models	Also Section 9
Basic and Standard levels	Rows with no level marker
Reinforced level	All rows, including [REINFORCED LEVEL] rows

Delete the sections and rows that do not apply to your organization.

Third-Party and Open-Weight Model Provenance Sheet

Field	Value
Organization	[LEGAL ENTITY NAME]
Function responsible for the sheet	[RESPONSIBLE FUNCTION]
Sheet owner	[OWNER]
Sheet identifier	[COMPLETE]
Identifier in the "AI System Inventory"	[COMPLETE]
AI system or systems that use the model	[AI SYSTEM]
Date opened	[DATE]
Next review	[DATE]

1. Purpose and when the sheet is completed

1.1 [ORGANIZATION NAME] completes this sheet:

a) before hosting any third-party model on its owned or rented infrastructure, including open-weight models of any origin, even for tests with real data;
b) before the pilot of a Medium- or High-risk system that uses a third-party model through an API;
c) whenever the model's version, quantization, fine-tuning or hosting mode changes.

1.2 The sheet does not replace the "AI Vendor Due Diligence Checklist" when the model is contracted from a vendor: it adds the information about the model itself. When there is no contracted vendor (downloaded open weights), this sheet is the self-assessment that replaces that due diligence.

1.3 No model covered by 1.1 goes into pilot or production without an "Approve" or "Approve with conditions" decision in Section 10.

2. Model identification

Field	Value
Commercial or technical name of the model	[COMPLETE]
Family and exact version	[COMPLETE]
Variant (full, quantized and format, adapter, fine-tuned, merged)	[COMPLETE]
Hosting mode (SaaS, developer API, self-hosted, third-party cloud)	[COMPLETE]
Deployment profile of the system (P1 to P5)	[COMPLETE]
Risk level of the system that uses it (Low, Medium, High)	[COMPLETE]
System owner	[COMPLETE]
Intended purpose of the model's use	[COMPLETE]
Data it will receive (categories; personal, sensitive or confidential data?)	[COMPLETE]