data-productsUsage Policy

Usage Policy

ORP describes what a dataset is. A trade also needs to say what may be done with it — in a form a rail can enforce automatically and a lawyer can read. v0.3 adds a Layer 6 usage_policy: a profile of W3C ODRL 2.2 restricted to a fixed vocabulary.

usage_policy:
  policy_id: "https://..."
  permissions:
    - permission_id: "p1"
      purpose: advertising_targeting      # from the purpose taxonomy below
      assignee: any                        # 'any' or a list of principal ids / DIDs
      scope: { geo: [DK], media: [ctv], time: { from, to } }
      derivation: internal                 # none | internal | licensed
      duties: [report_usage, delete_on_expiry]
  prohibitions:
    - { purpose: model_training }
    - { purpose: resale }
  expiry: "2027-10-24T00:00:00Z"
  revocable: true
  governing_law: "DK"

Purpose taxonomy (normative)

advertising_targeting, suppression, measurement, analytics, model_training, fine_tuning, evaluation, retrieval, inference, resale, research. Profiles may restrict this list, not extend it — extension goes through the spec.

Evaluation semantics

Given an activation context { purpose, assignee?, geo?, media?, time?, derivation_requested? }, an evaluator returns:

  • deny if any prohibition matches the purpose;
  • deny if no permission matches;
  • allow with the matching permission id otherwise.
⚠️

Unknown purpose or unknown scope key ⇒ deny. The evaluator fails closed. Derivation is ordered none < internal < licensed: a grant permits itself and anything below it.

Reference evaluator

from orp_validator.usage import evaluate_policy
 
d = evaluate_policy(policy, {"purpose": "advertising_targeting", "geo": "DK", "media": "ctv"})
d.decision      # "allow"
d.permission_id # "p1"
d.duties        # ["report_usage", "delete_on_expiry"]
 
evaluate_policy(policy, {"purpose": "model_training"}).decision  # "deny"  (prohibited)
evaluate_policy(policy, {"purpose": "advertising_targeting", "geo": "US"}).decision  # "deny" (scope)

A rail calls this at activation time to decide whether a concrete use is allowed. When the data product is a derivative, the rail evaluates the whole ancestry — see Lineage.

Next steps

  • Lineage — inherited constraints across derivations
  • Usage Records — prove what was actually done
Trust TiersLineage