Skip to content

Data rights and lineage in PROV-O terms, with one deterministic print gate

Decided in 0.7.0 (content). Recorded in 0.12.0.

Context and problem statement

A consumer prints generated items (cards, notes, briefs) built from sources with different rights: public statistics, the owner's own data, a vendor's personal-use licence, pages scraped without a licence. Who may see an item depends on the most restrictive thing it was made from, and whether its claims cite evidence decides whether it may be printed at all. Without a shared shape, each surface re-derives this, and a lineage reader has to learn each surface's vocabulary.

Decision drivers

  • One rule, pure and deterministic, that says "print" or names exactly one refusal.
  • Vocabulary a lineage reader already knows.
  • No policy engine to deploy for a single-owner library.

Considered options

  1. Ad-hoc flags per surface.
  2. A policy engine evaluating ODRL policies.
  3. Small frozen types named in W3C PROV-O terms, ODRL's words for the rules, and one pure gate.

Decision outcome

Chosen option 3, in content:

  • ContentItem is a prov:Entity, Activity a prov:Activity (trace, role, pack@sha, model, git sha); uses is prov:used, as_of is prov:generatedAtTime. An unrecorded field is absent, never faked.
  • DataClass holds rights (public < owned < licensed < scraped), private and link_required. An item's class is the join of what it uses: the most restrictive tier, private if any input is.
  • CEILING maps each audience to the tiers it may see.
  • print_gate(item, audience) returns ok, or one named refusal in a fixed order: rights, unclaimed (the body was never decomposed into claims), uncited, unsourced, link. unclaimed is separate so a body nobody checked is never reported as uncited by guess.

Consequences

  • Good: lineage exports as PROV-JSON with no translation layer.
  • Good: a refusal is a name a gauge or a reader can count.
  • Bad: ODRL supplies words, not enforcement; the gate is as strong as the classes callers assign.