Model Licences, Contracts, and Permitted Use
What exactly does the recorded agreement permit? A tabletop audit of E1–E4 across twelve actions, a licence-comparison matrix from Apache 2.0 to provider terms, and five hypothetical cases that stay hypothetical.
Thirty-five chapters treated permission as a gate: a record, a status, a reviewer, a date. This chapter opens the gate and reads what is inside — because “the teacher is available” never answered the question the engineer must ask:
What exactly does the recorded agreement permit, and what must be escalated before a distillation run begins?
The answer is always action-specific. A source may permit evaluation but forbid redistribution, allow fine-tuning but forbid competing-model development, or grant everything except commercial use. The chapter builds that discipline into a licence-comparison matrix, applies it to every E1–E4 study across twelve actions, and works five hypothetical cases that are labelled as hypothetical — because the most dangerous sentence in this territory is “it was publicly downloadable, so we assumed”.
This chapter is educational, not legal advice. It records what documents say, with versions and retrieval dates, and escalates everything else. A technical audit system cannot guarantee legal compliance; neither can this chapter. What both can do is make the decision inspectable — and stop the project from treating silence as permission. The chapter proceeds in five steps: the seven distinctions; the licence matrix with a worked comparison table; the twelve-action audit of the book’s own studies; five labelled hypotheticals plus a worked blocked record; and the reading method, escalation list, and limitations.
The distinctions that do the work
Seven distinctions carry the chapter; everything else is application.
Open-source software means code under an OSI-approved licence with source available and the freedoms to use, study, modify, and share. The repository’s own code lives here under Apache 2.0 (L-001) — and that licence covers the repository’s code, not any third-party teacher, dataset, or output the code might touch.
Open-source AI, under the Open Source AI Definition, version 1.0 (October 2024; consulted 16 November 2026), is stricter than “weights on a server”. The definition requires four freedoms — use, study, modify, share — backed by three components under OSI-approved terms: data information sufficient to build a substantially equivalent system, complete code for training and operation, and the parameters themselves. A release missing training-data information fails the definition however freely its weights download.
Open weights is the weaker claim: parameters published for download, whatever the terms. Many “open” models are open-weight, not open-source AI — the weights are there, the training data is not, and the licence restricts what follows. Confusing the two is the central failure mode this chapter exists to prevent.
Proprietary and gated models add access controls to licence restrictions: manual approval, acceptable-use policies, user thresholds. Gating is a contract event as well as a download friction — access under stated terms, with the terms surviving the download.
Research-only and non-commercial terms permit experimentation while forbidding deployment for revenue. The boundary is the use, not the user: a company researcher training on research-only weights for a commercial roadmap is outside the grant.
Derivative-model and output conditions govern what comes out of the model, not just the model itself. Some licences expressly treat distillation — via intermediate representations or synthetic outputs — as creating a derivative, which inherits the base licence’s conditions. Others attach naming, notice, or scale conditions to models trained on outputs.
Copyright versus contract is the last distinction and the most load-bearing. Copyright (and neighbouring rights, database rights, TDM exceptions) sets what the law reserves by default; the licence or service contract sets what the parties agreed on top. Either can block an action the other seems to allow, they operate in different jurisdictions under different rules, and this chapter records findings under each without conflating them. Chapters 37–39 take copyright, privacy, and the auditable pipeline in turn; this chapter establishes the reading method.
The licence matrix
The comparison below uses operative texts and the register’s dated findings (all retrieved 12 November 2026 unless stated). Summaries are pointers; the clause governs. Read across a row for one licence, down a column for one question:
| Licence | Derivatives | Redistribution | Commercial use | Competing-model training | Notice/attribution |
|---|---|---|---|---|---|
| Apache 2.0 (L-001) | permitted, conditions | permitted, conditions | permitted under terms | not addressed (see provider terms) | licence + notices required |
| MIT (L-002) | permitted | permitted | permitted | not addressed | copyright + permission notice |
| OpenRAIL (L-003) | permitted, restrictions travel | permitted, restrictions travel | restricted by use terms | restricted by use terms | terms + notices required |
| Gemma (L-004) | distillation expressly derivative | conditional | restricted | restricted via derivative terms | terms + notices + use restrictions |
| Llama 3.2 (L-005) | conditional (naming, scale) | conditional | conditional (threshold) | acceptable-use constrained | naming + notice required |
| API terms (L-015–L-019) | outputs owned, uses restricted | restricted | restricted | prohibited (narrow exceptions) | compliance + attribution duties |
Apache License 2.0 (L-001) — version 2.0, January 2004. Permits use, reproduction, and derivative works including commercial use, subject to licence-and-notice conditions; grants a patent licence; grants no trademark rights. The Foundation’s licensing FAQ (consulted alongside the text, explanatory only) confirms the reusable design: modified code may be distributed, even under a different licence with acknowledgement of the original, with no duty to contribute changes back — while the marks (“Apache”, project names) stay restricted. Operative for the repository’s own code. It does not by itself license any third-party model, dataset, or output.
MIT License (L-002) — permissive use, modification, and distribution subject to retaining the copyright and permission notice. No model-specific use restrictions, no patent grant, no trademark grant. Applies where the model card declares it — and only there: a card’s licence tag must be checked against the linked terms rather than trusted (Chapter 17’s discipline).
OpenRAIL and the RAIL family (L-003) — open access with use-based restrictions that downstream distributions must adopt. Not an OSI open-source licence. The specific model licence text governs, not the explainer: restrictions travel with derivatives.
Gemma Terms of Use (L-004) — last modified 1 April 2026. Defines Model Derivatives to include distillation using intermediate representations or synthetic outputs; distribution of a derivative must carry the terms, notices, and use restrictions; Google claims no rights in Outputs. The critical point for this book: training a student on Gemma outputs is derivative-making under the express terms, not a grey area — and if that student is ever distributed, it must carry the terms, notices, and use restrictions.
Llama 3.2 Community License (L-005) — released 25 September 2024. Permits use, modification, and distribution with conditions: a model trained on Llama outputs and then distributed carries naming duties (“Llama” prefix, “Built with Llama” display), on top of notice and acceptable-use duties and a 700-million-monthly-user threshold. Not an OSI open-source licence. The naming condition is the memorable one because it is unusual — a distributed student that must announce its teacher — but the acceptable-use policy is the broader constraint on deployment.
Permissive model cards: Qwen3-0.6B (L-006) — last modified 26 July 2025, declares Apache 2.0, not gated. The clean case: a small text-generation model whose card declares a permissive licence with no access friction. Even here the discipline holds — the card declares, the licence text governs, and any use beyond the licence (trademark use, patent questions) needs its own reading.
Gated cards: Gemma 3 1B IT (L-013) and Llama 3.2 1B Instruct (L-014) — both gated manual access, carrying the Gemma and Llama licences respectively. Gating means the terms were accepted at access: the project cannot claim it never saw them.
Provider terms (L-015, L-016, L-017, L-019). Four dated findings from the register, each action-specific. OpenAI’s consumer terms (updated 16 January 2026) prohibit using output to develop competing models and prohibit automated extraction. OpenAI’s Services Agreement (updated 1 December 2025, effective 1 January 2026) prohibits competing-model development in section 3.3(e), with a narrow Permitted Exception — models primarily for categorising, classifying, or organising data, undistributed and not commercially available, plus fine-tuning OpenAI models — that does not cover a generative replacement student. Anthropic’s Commercial Terms (effective 17 June 2025) prohibit accessing the services to build a competing product or service, including training competing models, except with Anthropic’s express approval; the recorded terms contain no other competing-model exception.
Google’s Gemini API Additional Terms (effective 23 March 2026) prohibit developing competing models and reverse engineering or extraction, with compliance and attribution duties and paid/unpaid data-use differences (paid services required for EEA, Swiss, and UK API clients). Customer ownership of outputs appears alongside these restrictions — owning the text does not license the competing use. None of these terms governs another provider’s product, another region’s agreement, or a negotiated contract: the record states the document, the date, and the action, and stops there.
The E1–E4 tabletop audit
The audit (full record in planning/36-rights-audit.md, machine
check in examples/common/rights_audit.py) covers twelve actions
against the book’s own studies. Its premise is unusually clean:
every teacher, dataset, and output in E1–E4 is self-created, every
one of the 25 rights manifests records VERIFIED_PERMITTED for
its study’s action, and no third-party model, API, dataset, or
personal data appears anywhere. The findings:
Permitted under recorded conditions (actions 1–7). Downloading or accessing the teacher (no external access occurred), running teacher inference, storing teacher outputs, generating synthetic examples, filtering or annotating them, training the student, and retaining intermediate datasets — each covered by the study manifests with local-only, recorded-artefact conditions. The retention row carries a forward pointer: versioned manifests and protocols are kept, raw checkpoints are git-ignored by policy, and a tested erasure procedure across caches, checkpoints, and distilled copies does not exist yet — Chapter 39’s subject.
Blocked pending determination (actions 8–12). Redistributing
student weights, redistributing training examples, commercial use,
publication (this book: REQUIRES_PERMISSION — independent review
and publisher sign-off outstanding by policy, all chapters still
drafts), and compliance with a future deletion or rights change
(UNKNOWN_PENDING_REVIEW — no lineage-linked erasure pipeline
tested). The verdict of the audit is therefore split by design:
training was permitted; leaving the repository was never
determined. No redistribution, commercial use, or publication is
authorised by this audit — stated here, in the audit record, and
in the matrix validator, which refuses any permitted-without-basis
row.
| Actions | Status | Record basis |
|---|---|---|
| 1–7. Access, inference, storage, generation, annotation, training, retention | PERMITTED | 25 self-created manifests, local-only conditions |
| 8. Redistribute weights | BLOCKED (no basis) | — |
| 9. Redistribute examples | BLOCKED (no basis) | — |
| 10. Commercial use | BLOCKED (no basis) | — |
| 11. Publication | REQUIRES_PERMISSION (review + sign-off outstanding) | policy |
| 12. Future deletion/rights change | UNKNOWN_PENDING_REVIEW (no tested erasure pipeline) | — |
Table 1. The twelve audited actions: training was permitted; leaving
the repository was never determined. From planning/36-rights-audit.md.
Two properties of the audit deserve emphasis because they
generalise. First, the existing records cover exact actions, not
programmes: RM-101 permits train_and_evaluate_e1, and nothing
about redistribution follows from it. Permission does not radiate.
Second, the blocked rows are blocked by absence, not by
prohibition: UNKNOWN_PENDING_REVIEW means no determination was
made, and the gate treats that as a stop — the policy’s second
principle, applied.
Five hypothetical cases
Each case is a demonstration — the book ran none of them — and each ends with the escalation it would trigger.
Case A — permissively licensed local model. A 0.6B-parameter instruction model whose card declares Apache 2.0 with no gating (L-006). Training a student on its outputs: the licence text permits derivatives with notice conditions; the card must be checked against the linked terms at retrieval; the teacher’s licence tag is not a dataset licence for any evaluation data. Redistribution of the student carries the Apache notice duties. Escalation: confirm the card-to-licence link at the retrieval date; record the version; no specialist review required for the licence question as stated — any trademark or patent question goes to counsel.
Case B — open-weight model with restrictions. A 1B instruction model behind gated access carrying use-restricted terms (L-013 or L-014 pattern). The weights download; the terms travel. Distillation creates a derivative (expressly, under Gemma terms; conditionally, under Llama terms with naming and scale duties). Open weights are not open-source AI: the OSAID data, code, and parameter requirements fail on the missing training-data information alone, before any restriction is read. Escalation: qualified review before training begins — the restrictions are custom, the naming and threshold conditions operational, and the gated access proves notice.
Case C — API teacher. A commercial chat API whose terms
prohibit competing-model development (L-016, L-017, or L-019
pattern). Storing outputs for analysis: outside the competing-model
clauses as recorded, though still subject to the rest of the service
terms. Training a replacement student on them: prohibited
under each recorded clause, with only the narrowest of exceptions
(the L-016 categorisation carve-out, which a generative student
does not meet, and L-017’s route of express approval from the
provider). Owning the outputs changes nothing. Escalation:
do not begin; seek express written permission or a different
teacher. This is the VERIFIED_PROHIBITED shape: a recorded
document forbidding the specific action.
Case D — mixed-origin synthetic dataset. Synthetic rows generated partly by a repository script (clean) and partly by a third-party model under output restrictions (encumbered), then mixed into one training file. The safe working assumption is that the mixture carries the strictest condition of any contributor — and, worse, the provenance of each row must survive filtering, shuffling, and retraining, or the restriction becomes unenforceable and the whole file suspect. Escalation: per-source records before mixing; row-level provenance or refusal to mix; qualified review of the output terms. This case is Chapter 37’s problem in miniature, which is why the chapter ends there.
Case E — rights change after training. A provider updates its terms mid-programme to prohibit an action the project already performed, or a dataset’s rightholder reserves TDM rights. The audit’s action-12 row fires: every derived artefact — weights, adapters, quantised copies, retained datasets — must be located through lineage and re-evaluated, retained under the new terms or erased. The book’s current answer (versioned manifests, ignored checkpoints, no tested erasure pipeline) is honestly insufficient, and the audit records it as such. Escalation: author decision with specialist review; Chapters 37–39 build the machinery — and until they do, the exposure stands open, stated, not hidden.
A worked rights record and the documented blocked case
The contract requires a complete rights record and a documented blocked case. Here is both: a record for a proposed run the book did not perform — distilling a student from a commercial chat API — written in the repository’s schema and refused by its gate.
Record RM-DEM-01 (demonstration, not minted): source, commercial chat API, version dated terms; action,
train_student_on_api_outputs; document, provider services agreement with the competing-model clause (L-016 pattern, section 3.3(e)); jurisdiction IE/EU; clause quoted: outputs may not be used to develop models competing with the provider’s services; conditions: none available that cover a generative replacement student; evidence: the dated terms page; reviewer: named author; status:VERIFIED_PROHIBITEDfor this action. Decision: blocked. The run does not start, the outputs are not collected, and the record is kept — a refusal with a pointer is evidence; a silent reroute to another endpoint is evasion.
Contrast the book’s own RM-101: source, self-created spiral
dataset/teacher/student; action, train_and_evaluate_e1;
document, study protocol with the repository’s Apache-2.0 terms;
status VERIFIED_PERMITTED under local-only conditions. The
decisive difference between the two records is the operative
clause for the stated action, and the gate treats them
oppositely.
That is the whole method: same schema, same reviewer discipline, opposite outcomes, both inspectable. Jurisdiction rides along on every record because copyright exceptions, database rights, and contract enforcement differ across borders: an IE/EU determination does not travel to another regime, a TDM reservation valid where made may not exist where relied upon, and any cross-jurisdiction action re-enters the method at step one rather than inheriting a domestic permission. The details belong to Chapters 37–39; the rule — re-determine, do not inherit — belongs here.
Reading method and escalation
The method, reduced to a card: identify the source with version;
state the action; locate the primary document with effective and
retrieval dates; quote or paraphrase the operative clause with its
reference; record jurisdiction and conditions; determine the
status — and if the document does not resolve the action, record
UNKNOWN_PENDING_REVIEW or REQUIRES_PERMISSION, never
permitted. A named human reviewer signs; ambiguity goes to a
specialist. The run manifest references the record; the gate
refuses anything else. Escalation triggers follow the policy:
competing-model questions, redistribution or commercialisation,
custom or ambiguous licences, personal or special-category data,
TDM reservations, unclear AI Act roles, third-party synthetic
data, conflicting licence tags, cross-jurisdiction actions — each
a stop, each with an owner.
flowchart TD
S[identify source + version] --> A[state action + jurisdiction]
A --> D[locate primary document + dates]
D --> C[quote operative clause]
C --> G{action resolved?}
G -->|yes| R[record PERMITTED/PROHIBITED with evidence + reviewer]
G -->|no| U[record UNKNOWN or REQUIRES_PERMISSION]
U --> E[escalate to specialist]
Figure 1. The reading method as a gate: every path ends in a recorded status or an escalation — never in silent permission.
What the method separates, the chapter keeps separated: the exact legal text from the licence’s requirements, the requirements from repository policy, policy from technical controls (a gate refuses; it does not interpret), controls from the author’s approvals, approvals from unresolved questions, and all of it from legal advice — which this chapter, the audit, and the code all decline to give. The repository’s Apache 2.0 licence governs the repository’s manuscript and code; it grants no rights in any third-party teacher, dataset, output, or quotation, and the book never implies otherwise.
Limitations
A tabletop audit of self-created studies is the easiest possible case — no third parties, no personal data, no negotiated terms — and its blocked rows still outnumber a reader’s comfort. Nothing here applies a finding to a different action, product, region, or agreement version; provider terms change, model cards relist, and every retrieval date in the register is a snapshot. The OSAID consultation (version 1.0, this batch) is recorded with its date and is not part of the D-FP-03 register. No case law or negotiated-agreement evidence was examined (the methods-map gap, stated). Ambiguous questions are flagged for a qualified lawyer, not resolved. No independent review.
Preparing for copyright, privacy, and the pipeline
Permission to train was the question of this chapter. Ownership of training data is the question of Chapter 37; personal data and regulated obligations, of Chapter 38; and the pipeline that makes every answer above reproducible and erasable, of Chapter 39. The audit’s blocked rows are their starting backlog — eight through twelve, named, owned, waiting.
References
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H. and Crawford, K. (2021). Datasheets for Datasets. Communications of the ACM 64(12). arXiv:1803.09010. https://arxiv.org/abs/1803.09010
Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D. and Gebru, T. (2019). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT 2019).* https://doi.org/10.1145/3287560.3287596
Legal sources
Apache License, Version 2.0. Version 2.0, January 2004. https://www.apache.org/licenses/LICENSE-2.0 (retrieved 12 November 2026). Cited as L-001 for the use, derivative-work, distribution, patent, and notice conditions governing the repository’s own code; not legal advice.
MIT License. SPDX MIT. https://opensource.org/license/mit (retrieved 12 November 2026). Cited as L-002 for the permissive use and notice conditions where a model card declares MIT; not legal advice.
OpenRAIL and RAIL licence family explainer. Dated 31 August 2022. https://huggingface.co/blog/open_rail (retrieved 12 November 2026). Cited as L-003 for the use-based restrictions travelling with derivatives, and for the not-OSI-open-source distinction; the specific model licence governs; not legal advice.
Google Gemma Terms of Use. Last modified 1 April 2026. https://ai.google.dev/gemma/terms (retrieved 12 November 2026). Cited as L-004 for the Model Derivative treatment of distillation via intermediate representations or synthetic outputs; not legal advice.
Llama 3.2 Community License Agreement. Version released 25 September 2024. https://www.llama.com/llama3_2/license/ (retrieved 12 November 2026). Cited as L-005 for the naming, notice, acceptable-use, and scale conditions on derivatives; not legal advice.
Qwen3-0.6B model card. Last modified 26 July 2025. https://huggingface.co/Qwen/Qwen3-0.6B (retrieved 12 November 2026). Cited as L-006 for the Apache-2.0 declaration in the permissive-local-model case; not legal advice.
Gemma 3 1B IT model card. Last modified 4 April 2025. https://huggingface.co/google/gemma-3-1b-it (retrieved 12 November 2026). Cited as L-013 for the gated-access and Gemma-licence context in the open-weight-with-restrictions case; not legal advice.
Llama 3.2 1B Instruct model card. Last modified 24 October 2024. https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct (retrieved 12 November 2026). Cited as L-014 for the gated-access and Llama-licence context in the open-weight-with-restrictions case; not legal advice.
OpenAI Europe Terms of Use. Updated 16 January 2026. https://openai.com/policies/terms-of-use/ (retrieved 12 November 2026). Cited as L-015 for the competing-model and automated extraction prohibitions in the API-teacher case; not legal advice.
OpenAI Services Agreement. Updated 1 December 2025, effective 1 January 2026. https://openai.com/policies/services-agreement/ (retrieved 12 November 2026). Cited as L-016 for section 3.3(e), the narrow Permitted Exception, and customer ownership of output in the API-teacher case; not legal advice.
Anthropic Commercial Terms of Service. Effective 17 June 2025. https://www.anthropic.com/legal/commercial-terms (retrieved 12 November 2026). Cited as L-017 for the section D.4 competing-product and competing-model prohibition in the API-teacher case; not legal advice.
Gemini API Additional Terms of Service. Effective 23 March 2026. https://ai.google.dev/gemini-api/terms (retrieved 12 November 2026). Cited as L-019 for the competing-model, reverse-engineering, attribution, and paid-service conditions in the API-teacher case; not legal advice.
Open Source AI Definition, version 1.0. Published October 2024 by the Open Source Initiative. https://opensource.org/ai/open-source-ai-definition (consulted 16 November 2026; not part of the D-FP-03 legal register). Cited for the four freedoms and the data-information, code, and parameters requirements distinguishing open-source AI from open weights; not legal advice.