[abrichr]

Phase 2 of qualifying the verifier against a commercial EHR through a public developer sandbox; Phase 1 (the terms) merged in #1650. No agent created an account; the founder registered the Epic app.

FINDINGS.md gains a Phase 2 section, and its registration steps now match what Epic asked for (a JWK Set URL).

Phase 2, Epic (benchmark/commercial_ehr/epic/):

• Read path, probed (probe.json): search and read work. There is no AuditEvent and no _history (both 404) and no meta. _lastUpdated is refused on Patient and silently ignored on Observation. Provenance is available only through _revinclude.
• Epic's write behavior:
  • A Patient create answers 201 but drops the phone (email kept).
  • A repeat create of an existing person answers 201 with the old record's Location.
  • A created allergy doesn't appear in search.
  • An Observation create needs an Encounter, which the app has no scope to find.
• Verifier: openadapt_flow.verifier with the fhir record check and its existing smart_backend auth, contract register_patient.yaml. The frame is the search set. The claim names the path: read back through Epic's FHIR search and read.
• Blind round: 13 cases from a separate agent, frozen by sha256 (ff402d6a…c08d) at commit c93069d0d and run once. 12/13 right, none passed as verified. The miss was a false refusal of a clean save beside a look-alike patient. Each value check matched every record with the task's family name, which was a contract defect. The contract now uses one exact_new_set effect. Only the hash and counts are public (blind_summary.json).
• Non-blind rounds: 7/9, then 10/10, then 10/10 after the contract fix. A rerun of the blind file after the fix got 13/13; it is no longer blind.
• Read path, named in RESULTS.md: FHIR search and read only. There is no AuditEvent and no _history. The probe read Provenance through _revinclude, but the verifier doesn't use it, so authorship isn't checked.

Fix found by the sandbox: runtime/effects/fhir.py counted a search's OperationOutcome entry (Epic's empty-search warning) as a record, so a clean create read as a removed record and was refused. A warning or information outcome is now skipped, and an error outcome makes the read unreadable. An OperationOutcome with no search.mode is handled the same way, and ensure_record now treats an error outcome as unreadable. The adversarial review passed with no blocking finding (REVIEW.md); tests cover each case.

Also: a CAPABILITIES.md row, and a regenerated llms-full.txt. Tests run: test_effect_paging, the FHIR, verifier and ensure_record tests, the capabilities map and shipped-copy tests (470 passed).