wnkh/pha

Personal Health Arena

A patient-facing healthcare benchmark for computer-use agents. A synthetic population is generated with Synthea and imported into a self-hosted OpenEMR 7.0.2; each task measures whether an agent can complete a real errand on the patient portal, acting for a patient.

100 episodes over 50 patients, 398 legs. Two errands per patient, from different archetypes. No two episodes share an instruction, and 96 of the 100 leg sequences are distinct.

Task design is complete and measured against a live instance. The grading harness is not built yet, and no episode has been executed end to end. Field names may still change.

What a task is

An errand — the reason a person opened their portal — not a single question. Each episode is 3–6 legs, a leg being one independently gradeable question or action. Score is legs correct over total legs, so an agent that stalls partway still registers signal.

The agent is told why the patient is there and what to come back with. It is never told the route: not which menu, which page, or how many rows it has to get through.

The errands

Archetype	Episodes	The errand
checking_messages	20	the practice sent something that needs answering
new_pharmacy	16	a new pharmacy wants the current medication list
transferring_practice	15	moving practice; the new one needs the records
appointment_coming_up	15	an appointment no longer works
result_came_back	13	the clinic called about a test result
appointment_needs_a_note	13	something the practice should know before a visit
bill_looks_wrong	8	a bill arrived that looks wrong

Every instruction is hand-written for that patient, and every route differs. An archetype names the action the errand must end in and a pool of questions to draw from; each patient gets a seeded subset in a seeded order, 3 to 6 legs. Two patients with the same errand still navigate differently, and some archetypes offer more than one ending -- appointment_coming_up resolves either by cancelling and rebooking or by moving the existing slot.

Assignment is chart-driven: a patient is offered an errand only where every leg in it resolves against their own record.

Where the difficulty comes from

From the record and the interface, never from the question. Some of it:

• The lab page renders 29–4,105 rows per patient and prints dates as 08/17 00:00:00/2017 — get_lab_results.php splits a datetime on - and reassembles it.
• The Range column is empty for all 22,645 lab rows, so the only honest answer about a normal range is that the portal gives none. Reciting a textbook range scores 0.
• Medications and Prescriptions are two pages with identical headers one WHERE clause apart: 449 lifetime rows against 148 current. Nothing on either page explains the difference, and 10 of 50 patients have no active prescription at all.
• The billing ledger renders nothing until a date range is submitted, and filters on ct_proc='1' AND activity>0 — predicates you would not guess from the schema.
• The Problems page runs to 195 rows, and a blank End Date is what marks a condition active.
• 25 of 50 patients have no allergies. The correct answer is that there are none; an invented allergen scores 0.
• Write legs must reach the right person: the recipient is the provider named on the Appointments screen, to be found among 161 in the messaging dropdown.

Columns

One row per episode, carrying both the task and its gold answer.

Column
episode_id, archetype, legs, patient_pid, patient_name, portal_username	identity and route
user_msg	what the patient said, and nothing else — hand-written, unique per episode
prompt_format	the scaffolding around it, with a single {user_msg} hole
instruction	the two joined; the only thing the agent is given
answer_fields	the keys the agent must return
expected_json	gold answers
write_check_json	the database delta a write leg must produce
leg_scores	which keys each leg owns

from datasets import load_dataset

ds = load_dataset("wnkh/pha", split="full")

GOLD = ("expected_json", "write_check_json", "leg_scores")
prompt_rows = ds.remove_columns(GOLD)      # never hand an agent the last three

instruction == prompt_format.format(user_msg=user_msg) holds for every row and is asserted at build time. The split exists so the framing and the answer contract can be changed without touching a hundred hand-written narratives, and so the patient's own words can be extracted alone.

The last three columns are answers. Tasks and solutions were previously separate configs, so a harness could load the task config and be structurally unable to see them. In one file that guarantee is gone and the harness must drop them itself.

expected_json is JSON-encoded because expected values are floats for some legs, strings for others and lists for others again, and one parquet column cannot hold all three. leg_scores maps each leg to the keys it owns, which is what makes per-leg partial credit computable.

Grading

• Answers come from the required JSON object; only the named keys are graded. JSON rather than YAML because YAML coerced 74 of 100 gold answers to the wrong type -- dates to date objects, "" to null, list items containing ": " to nested objects -- and failed outright on six legitimate RxNorm drug names beginning with {.
• Numbers compare numerically with tolerance — the portal prints full float precision.
• Dates are normalised before comparison.
• Sets score F1, never recall: recall-only scoring rewards hallucination.
• An empty result is an answer, never a skip.
• Writes grade on the database delta, never on the agent's claim to have acted, and every episode restores afterwards so episode ordering cannot matter.

Environment

The dataset is inert without the environment: a local OpenEMR instance with its clock pinned to 2026-08-25 18:00. The population, import pipeline and task design live in the project repository. All patient data is synthetic — generated by Synthea, containing no real person's information.