Use separate contract schedules even if a commercial package combines several uses. The table is a proposed issue list for commercial and legal review, not a statement that one legal rule applies to every use.
| Dimension | Broad training | Fine-tuning/adaptation | Retrieval/grounding | Evaluation |
|---|---|---|---|---|
| Buyer objective | Develop general model capability | Improve a defined task or domain | Supply information at answer time | Measure a specified capability |
| Owner's key boundary | Models, versions and downstream recipients | Named models/tasks and reuse | Queries, display, caching and updates | Test set integrity and reuse restrictions |
| Evidence to request | Corpus specification and training-use description | Baseline, task and measurable acceptance criteria | Query demand and attribution/display examples | Test purpose and contamination controls |
| Possible billing unit to negotiate | Defined corpus/term or staged licence | Defined project/model scope | Subscription, request, displayed use or agreed value measure | Test-set licence or evaluation project |
| Exit question | What survives in trained models and retained copies? | Can adapted models continue to be used? | What access, caches and exports stop? | What results and test material may remain? |
Billing units above are alternatives to evaluate, not claims about prevailing prices. “Training rights” should not silently become permission for unlimited redistribution or competing products. Conversely, a retrieval agreement needs more than a prohibition on training: define display length, retention, refreshes, audit evidence and the treatment of derivative datasets.
Public examples show why these distinctions matter. The New York Times–Amazon announcement includes both display-related uses and foundation-model training. Snowflake describes Cortex Knowledge Extensions as shared search services. Neither arrangement should be reduced to a generic licence to “use content for AI.” New York Times–Amazon licensing announcement, 29 May 2025 Snowflake Cortex Knowledge Extensions overview