> ## Documentation Index > Fetch the complete documentation index at: https://docs.verial.ai/llms.txt > Use this file to discover all available pages before exploring further. # Prior Auth Testing > Define and run a prior authorization benchmark. ## User Prompt > "Create a benchmark that tests whether my agent can submit a prior auth for an MRI." ## Tool Call Sequence **Call:** `benchmarks` ```json theme={null} { "action": "create", "name": "Prior Auth Test", "environmentId": "env_clinic_01", "timeout": 300 } ``` **Response:** ```json theme={null} { "id": "bm_pa_01", "name": "Prior Auth Test", "environmentId": "env_clinic_01", "timeout": 300, "tasks": [], "createdAt": "2026-04-06T11:00:00Z" } ``` **What the agent learns:** The benchmark is created and linked to the environment. The 300-second timeout gives the agent-under-test five minutes to complete each task. **Call:** `tasks` ```json theme={null} { "action": "create", "benchmarkId": "bm_pa_01", "name": "Submit PA for MRI", "instruction": "Submit a prior authorization request for a lumbar spine MRI (CPT 72148) for the patient. Use the payer portal to file the request and confirm submission." } ``` **Response:** ```json theme={null} { "id": "task_mri_01", "benchmarkId": "bm_pa_01", "name": "Submit PA for MRI", "instruction": "Submit a prior authorization request for a lumbar spine MRI (CPT 72148) for the patient. Use the payer portal to file the request and confirm submission.", "criteria": [], "createdAt": "2026-04-06T11:00:01Z" } ``` **What the agent learns:** The task defines what the agent-under-test should do. The instruction is specific about the procedure type and CPT code so criteria can verify exact behavior. The agent creates two criteria: one to verify submission and one to verify the correct CPT code. **Call 1:** `criteria` ```json theme={null} { "action": "create", "taskId": "task_mri_01", "name": "pa-submitted", "assertion": { "type": "portal-state-match", "spec": { "table": "prior_auth_requests", "where": { "patient_id": "pat_01" } } }, "weight": 1.0 } ``` **Response:** ```json theme={null} { "id": "crit_pa_01", "taskId": "task_mri_01", "name": "pa-submitted", "weight": 1.0, "createdAt": "2026-04-06T11:00:02Z" } ``` **Call 2:** `criteria` ```json theme={null} { "action": "create", "taskId": "task_mri_01", "name": "correct-cpt", "assertion": { "type": "portal-state-match", "spec": { "table": "prior_auth_requests", "fields": { "cpt_code": "72148" } } }, "weight": 0.5 } ``` **Response:** ```json theme={null} { "id": "crit_cpt_01", "taskId": "task_mri_01", "name": "correct-cpt", "weight": 0.5, "createdAt": "2026-04-06T11:00:03Z" } ``` **What the agent learns:** The first criterion (weight 1.0) is the critical check: did the agent actually submit a PA? The second criterion (weight 0.5) verifies correctness of the CPT code. Weighted scoring means the submission check counts twice as much as the code check. **Call:** `benchmark_runs` ```json theme={null} { "action": "create", "benchmarkId": "bm_pa_01" } ``` **Response:** ```json theme={null} { "id": "run_01", "benchmarkId": "bm_pa_01", "status": "Running", "score": null, "verdict": null, "createdAt": "2026-04-06T11:00:04Z" } ``` **What the agent learns:** The run is in progress. The agent-under-test is now executing against the live playground. The agent should poll for completion. The agent polls the run, then drills into task-level and criterion-level results. **Call 1:** `benchmark_runs` ```json theme={null} { "action": "get", "id": "run_01" } ``` **Response:** ```json theme={null} { "id": "run_01", "benchmarkId": "bm_pa_01", "status": "Completed", "score": 0.87, "verdict": "Pass", "completedAt": "2026-04-06T11:03:42Z" } ``` **Call 2:** `task-runs` ```json theme={null} { "action": "list", "runId": "run_01" } ``` **Response:** ```json theme={null} { "items": [ { "id": "tr_01", "runId": "run_01", "taskId": "task_mri_01", "name": "Submit PA for MRI", "status": "Completed", "score": 0.87 } ] } ``` **Call 3:** `criterion-runs` ```json theme={null} { "action": "list", "taskRunId": "tr_01" } ``` **Response:** ```json theme={null} { "items": [ { "id": "cr_01", "criterionId": "crit_pa_01", "name": "pa-submitted", "result": "Pass", "score": 1.0, "reasoning": "The agent navigated to the BlueCross portal and submitted a prior auth request for the patient. The submission confirmation was recorded in the interaction log." }, { "id": "cr_02", "criterionId": "crit_cpt_01", "name": "correct-cpt", "result": "Partial", "score": 0.6, "reasoning": "The prior auth request included CPT code 72148, but the agent initially entered 72141 and had to correct it. The final submission was correct." } ] } ``` **What the agent learns:** The run passed with a score of 0.87. The PA was submitted successfully (full marks), but the CPT code check scored 0.6 because the agent needed a correction attempt. The `reasoning` field explains exactly what happened. ## Agent Response > Your prior auth benchmark completed with a **Pass** verdict and a score of **0.87**. > > | Criterion | Result | Score | Detail | > | ------------ | ------- | ----- | ------------------------------------------------------------------------ | > | pa-submitted | Pass | 1.0 | PA was submitted through the payer portal | > | correct-cpt | Partial | 0.6 | CPT 72148 was correct in the final submission, but required a correction | > > The agent successfully submitted the prior authorization, though it initially entered the wrong CPT code before self-correcting. You might want to improve your agent's CPT code lookup logic. ## Summary | Metric | Value | | ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Tool calls | **7** | | Tools used | `benchmarks`, `tasks`, `criteria`, `benchmark_runs`, `task-runs`, `criterion-runs` | | Key technique | Define the benchmark structure (benchmark, task, criteria) before starting the run. Use weighted criteria to distinguish critical checks from secondary validations. Poll the run, then drill into task-runs and criterion-runs for detailed results. |