License data for training and derivation
This scenario shows why ENTITY treats rights as the scarce economic object instead of pretending dataset bytes disappear when licensed. The same governed object can support different entitlements for inspection, query, training, derivation, redistribution and commercialization.
Scenario
A dataset controller wants to permit one organization to train a model and produce internal derived outputs, while prohibiting raw redistribution and requiring a separate commercial right before derived outputs may be sold externally.
1 · Ingest the source without inventing ownership
Bring the dataset or repository under governed state with explicit source, controller, rights-holder/evidence and provenance references. BTDU can preserve exact objects and lineage, but topology and ingestion do not manufacture legal title.
2 · Model rights separately
| Right | Example policy |
|---|---|
INSPECT | May view metadata/schema or bounded samples. |
QUERY | May execute governed queries without receiving full raw corpus. |
TRAIN | May use the governed corpus in an approved model-training workflow. |
DERIVE | May create embeddings, features, model states or other derived outputs under stated conditions. |
REDISTRIBUTE | May redistribute source/derived material only if explicitly granted. |
COMMERCIALIZE | May commercially exploit permitted outputs only if explicitly granted. |
A TRAIN entitlement does not automatically imply REDISTRIBUTE or COMMERCIALIZE. Keep each economic/legal capability machine-readable and independently testable.
3 · Create the instrument/entitlement terms
Define the object reference, issuer, permitted actions, purpose, duration, use caps, geographic/jurisdiction conditions, derivation rules, transferability and any participation/revenue terms that were actually agreed.
underlying_object: DATASET_OBJECT_REF rights: [QUERY, TRAIN, DERIVE] forbidden: [REDISTRIBUTE, COMMERCIALIZE] purpose: model-development expiry: explicit usage_limit: explicit or unlimited by contract jurisdiction: explicit if applicable revenue_participation: only if actually agreed
4 · Record training as governed usage
When training begins, preserve the acting Entity, entitlement reference, source-object references, purpose, model/workflow identity, start/end/checkpoint evidence and resulting derivation identifiers.
5 · Carry lineage into derived outputs
The derived model or artifact should link back to the contributing governed information and the rights under which the derivation occurred. This lineage is not itself a royalty rule. If economic participation is intended, store the agreed rule separately.
6 · Test downstream use
Before an output is redistributed, sublicensed or commercialized, evaluate the requested action against the actual entitlement and any downstream policy. A model existing is not proof that all commercial rights exist.
7 · Preserve economic consequence separately
If a valid commercial right later exists, ENTITY can connect usage, downstream transaction, revenue rule, obligation and settlement evidence. Do not convert training provenance into an automatic invoice or royalty.
Negative tests
- TRAIN exists but COMMERCIALIZE does not → deny commercial use.
- Purpose-bound entitlement used for a different purpose → deny or require new authorization.
- Expired entitlement → deny new governed usage.
- Model output exists but source-right evidence is missing → do not infer clean rights.
- Revenue participation not agreed → do not create a retroactive obligation from provenance alone.
What this scenario demonstrates
ENTITY can make data economically legible without requiring the bytes themselves to be scarce. The scarce/governed object is the bounded right, entitlement, capacity, duration or economic interest connected to the information and its lineage.