The workflows we've actually recorded, compiled, and verified.

Task

Create exactly one fictional patient (name, DOB, sex, address, phone, and a reserved example.invalid email) and save it.

Recording / bundle

benchmark/openemr_local — recorded browser demonstration, compiled to a bundle. Baseline SHA-256 8d2901490a0a6a6044e94b6a8a1663436b7dacedda4f2fe1fb8c48405165011d; raw results in the git-ignored local directory results-model-free-corrected5-20260716/. View benchmark →

Policy

Model-free run: healthy compiled replay makes 0 model calls at $0 model cost. A separate small-N paid-agent run produced 0/6 correct writes and 6/6 missing_write outcomes, with 0/6 over-halt, 0/6 silent incorrect, and $0.8901/run. Because it used a separately provisioned baseline, full_matrix_complete and publication_ready remain false.

Identity contract

Two distinct OAuth clients. The actor holds openid api:oemr user/patient.crus; the oracle holds openid api:oemr user/patient.rs (read-only, cannot create/update/delete). The oracle is never the writer.

Effect oracle

Separately authenticated read-only REST patient readback compared field-for-field with direct SQL from patient_data, plus an exact before/after table-delta inventory and an unchanged-digest check on every non-target patient_data row. Exactly one new patient_data row is required; pixels and actor self-report never establish success.

Known limits

• The 6 paid-agent trials used a separate baseline and all ended missing-write; this is not a matched three-arm comparison.
• Synthetic PHI-shaped data only (all values fictional); not regulated, clean-machine, or design-partner evidence.
• Not customer-proven and not a broad reliability claim.

A separate historical field showcase on the shared OpenEMR public demo ran compiled 20/20 vs. computer-use agent 10/10 (2026-07-08); that is a field result on a daily-resetting shared instance, not a CI-reproducible benchmark.

.venv/bin/python scripts/openemr_local_demo.py run --profile initial --model-free --bundle benchmark/openemr_local/out/bundle --out benchmark/openemr_local/out/model-free-initial-YYYYMMDD

Task

Create exactly one synthetic Loan Application — product “OpenAdapt Synthetic Term Loan”, amount 125000, 18 repayment periods, and the declared contact fields — and save one application.

Recording / bundle

benchmark/frappe_lending — recording-live-valid11, compiled to a bundle. Baseline SHA-256 7fd6c965f6b7a11f54e451cdc73fdf65f88d9883dc5f8eb5b2055b3cd4be8b83; raw results in the git-ignored local directory results-model-free-corrected3-20260716/. View benchmark →

Policy

Model-free run: 0 model calls at $0 model cost. Frappe exposes a good REST API, so a real deployment should prefer the API arm; the browser arm isolates the value of compiled replay when a UI path is required. A separate paid-agent run completed 6/6 correct writes (5/6 clean), with 1/6 post-write cost-cap over-halt, 0/6 silent incorrect, and $0.4240/run; its baseline was not matched to the model-free subset, so publication_ready is false.

Identity contract

A read-only REST fixture user (openadapt.oracle@example.invalid) with a custom read-only Loan Application permission — distinct from the UI/API writer, so it cannot substitute a different Customer.

Effect oracle

Read-only REST field readback plus direct MariaDB read-back and an exact per-table row-count contract: “tabLoan Application: +1” and every other table +0. A new subscriber table fails closed until reviewed. Pixels and self-report never establish success.

Known limits

• The paid-agent arm used a separately provisioned baseline; not a matched three-arm comparison.
• Frappe is API-rich and is NOT a legacy Windows/Citrix LOS; this is not evidence for a proprietary system such as Encompass or Calyx.
• Not customer-proven and not a broad reliability claim.

.venv/bin/python scripts/frappe_lending_demo.py run --profile initial --model-free --bundle benchmark/frappe_lending/out/bundle --out benchmark/frappe_lending/out/model-free-initial-YYYYMMDD

Task

Enter exactly one synthetic health-facility claim (insuree number, claim number, diagnosis A000, explanation note, and service A1) and save it into status “Entered”.

Recording / bundle

benchmark/openimis_claims — recorded claim OA000003, then three compiled replays (OA231353, OA231429, OA231458). Evidence commit 3276ad2b537c558211a5a357fd7ac1e19f0a029e (openadapt-flow #141). View benchmark →

Policy

Model-free compiled replays: 0 model calls. A separate 3-trial paid-agent run was 3/3 correct, with 0/3 over-halt, 0/3 silent incorrect, and $0.4793/run. This remains a reference/demo environment: no matched timing matrix or publication protocol.

Identity contract

The openIMIS demo actor credential (Admin) with a fixed health facility (VIHOS001) and claim-administrator context as unmeasured setup. Weaker than the OpenEMR/Frappe entries: verification is by direct SQL rather than a separate read-only oracle client.

Effect oracle

Direct SQL read of the openIMIS claims database requiring exactly one non-voided claim row per run, in status “Entered”, matching the demonstrated insuree number and health facility. A duplicate or missing row fails the run loudly; pixels never establish success.

Known limits

• openIMIS is AGPL-3.0 — the reference environment stays repository-only and is never shipped in a packaged artifact.
• Reference/demo only: the 3 paid-agent trials were separate, with no matched timing matrix; not adjudication automation.
• Not a customer deployment, not Windows/RDP/Citrix, and not customer-proven. OpenAdapt is not affiliated with the openIMIS Initiative.

.venv/bin/python scripts/openimis_claims_demo.py replay --bundle benchmark/openimis_claims/out/bundle --headed