epic: regulator-readable audit export (provenance_audit_log schema)

[bokelley]

Problem (commercial, not compliance)

A publisher facing an Art 50 or CalAG inquiry currently responds in weeks — engineering ticket to query impression logs, separate query for provenance claims, manual reconciliation with creative-policy state at serve time, lawyer assembling a narrative. The compliance team's resource constraint is response time under regulatory deadlines.

A provenance_audit_log schema that aggregates, for a given media buy or impression cohort:

• what was claimed (provenance object)
• by whom (declared_by / signed claim)
• with what signature (provenance_signature, when implemented)
• what was verified (governance agent's get_creative_features response)
• what was served (served_disclosure receipt — see #TBD)
• against what creative-policy state at serve time

…cuts the response from weeks to hours. That's the commercial lever: AdCP adoption becomes "cut regulatory response time from N weeks to N hours," which moves the procurement conversation from ad ops to legal/risk, where the larger dollars sit.

Why this is an epic, not a sprint ticket

This is roadmap signaling. The right time to spec it concretely is after the dependencies land:

• Depends on #TBD signed provenance claims — without signatures, the export is just compiled JSON, not legally useful evidence.
• Depends on #TBD served-disclosure receipt — without post-serve evidence, the export is pre-serve intent only.
• Depends on #TBD provenance history — for chain-of-custody audit questions ("who else touched this?").

Strawman shape (informational only — for roadmap publication, not implementation)

{
  "audit_export_id": "audit_xyz...",
  "scope": {
    "media_buy_id": "mb_abc...",
    "time_range": { "start": "2026-04-01", "end": "2026-04-30" },
    "jurisdictions": [{ "country": "DE", "regulation": "eu_ai_act_article_50" }]
  },
  "creatives": [
    {
      "creative_id": "cr_123",
      "provenance": { /* full object */ },
      "provenance_signature": { /* JWS, verified */ },
      "provenance_history": [ /* chain */ ],
      "creative_policy_at_serve": { /* policy snapshot */ },
      "verification": {
        "agent_url": "https://verifier.example.com",
        "feature_id": "ai_detection_v2",
        "result": "confirmed",
        "confidence": 0.92,
        "called_at": "2026-05-12T13:00:00Z"
      },
      "impressions": {
        "total": 142893,
        "with_served_disclosure": 142893,
        "rendered_per_guidance": 142841,
        "discrepancies": 52
      },
      "discrepancy_samples": [
        {
          "impression_id": "imp_...",
          "expected_persistence": "continuous",
          "rendered_duration_ms": 8000,
          "impression_duration_ms": 15000
        }
      ]
    }
  ],
  "export_signed_by": { /* publisher's signature over the entire export */ },
  "exported_at": "2026-05-12T15:00:00Z"
}

The structured-discrepancy section is the audit value: a regulator doesn't want a screenshot, they want a sampleable, queryable record.

Goals for the roadmap entry

1. Publish the strawman publicly now so legal/risk procurement conversations can anticipate it.
2. Block detailed spec on dependency landings — signed claims, served-disclosure receipt, provenance history all need to land first.
3. Recruit a tier-1 publisher and a compliance-team buyer as design partners when spec work starts. This is the surface they'll actually use; design without them is wasted.

Out of scope

• Specifying retention windows or jurisdictional storage requirements. Publisher policy, not protocol.
• Defining the regulator-facing API. The export is a file (signed JSON), not an endpoint. Regulators don't poll us.
• Real-time streaming variants. The use case is "respond to inquiry within hours"; daily-batch is sufficient.

Refs

• #TBD signed provenance claims (dependency)
• #TBD served-disclosure receipt (dependency)
• #TBD provenance history (dependency)
• PR docs(provenance): frame provenance as transport, not compliance #4437 — frames the protocol as transport; the audit export is what makes that transport commercially defensible

Priority

Roadmap (epic). Post-GA. The commercial lever, not the compliance fix.

[bokelley]

Linked issues from GA-readiness review (dependencies):

• RFC: signed provenance claims (JWS over the provenance object) #4441 — RFC: signed provenance claims (dependency — signed evidence is what makes the export legally useful)
• served-disclosure receipt: machine-readable record of what the user actually saw at first exposure #4442 — served-disclosure receipt (dependency — post-serve evidence)
• provenance_history[]: chain of custody for assets that move through intermediaries #4443 — provenance history (dependency — chain-of-custody)
• RFC: unified disclosures[] with type discrimination (AI + sponsored + health + political + financial) #4444 — RFC: unified disclosures[]
• provenance: spec verification-unavailable fallback states (accepted_verifiers all timeout/error) #4446 — verification-unavailable fallback states