Synthetic Clinic Protocol Answer

Synthetic-clinic protocol answer

What this task family is

You are given, per task:

• input/prompt.txt — names one target case id (e.g. CASE-...) and, in prose, which clinical questions the answer must resolve.
• input/payloads/answer_template.json — the exact output contract: the required top-level keys, each field's type, the closed enum allowed_values, numeric precision, ordering rules, where null is permitted, and any fixed constants (often task_id, case_id).
• A separate environment-access file (e.g. environment_access.md) — the clinic runtime base URL, credentials (usually none), and the allowed endpoints for this run.

Your job is to return exactly one JSON object that conforms to the template, with every value derived from the clinic record by applying the governing protocol. No prose, no markdown, no extra top-level keys.

The correct answer is always derivable from the target patient's authoritative records plus the protocol. Do not guess, and do not invent findings the chart does not support.

Procedure

1. Read the contract first

Parse answer_template.json completely before touching the network. Note, for every field: the exact key name, type, the closed set of allowed_values (you may only emit values from these sets), precision (integer vs. N-decimal), ordering rules ("normalized as a set" vs. "sorted by ..."), nullability, and any required_value / expected_constant. This is the ground truth for shape; the environment supplies the content.

Set the fixed constants directly: task_id = the task folder id (e.g. train_001 / the current task id), case_id = the id from prompt.txt.

2. Load environment access

Read the environment-access file for the base URL and the allowed endpoint list. Only call endpoints on that list. Do not mutate anything — every task is read-only (no orders, no POSTs that change state). If a POST /api/query endpoint is listed but rejects calls for a missing "clinic token" and no token is provided, treat it as unavailable and use the GET resource endpoints, which are sufficient.

3. Pull the case bundle

GET /api/cases/{case_id} returns a bundle for the target case:

• case — patient_id, case_type, service_date, status, summary.
• findings — list of {finding_key, finding_value, source_id}; the structured clinical narrative (vitals ranges, symptoms present/absent, current_time, allergy constraints, outreach posture, etc.).
• observations — FHIR-like: observation_id, code, display, status, effective_time, value_number/value_text, unit, interpretation, patient_id.
• imaging, medications, allergies, problems, sdoh, care_registry (any may be null/empty).

Fetch GET /api/patients/{patient_id} (and, if useful, the topical list endpoints such as GET /api/observations) for demographics and fuller context.

4. Select and read the governing protocol

GET /api/protocols lists the protocols; pick the one whose title/scope matches the case's case_type, then GET /api/protocols/{protocol_id} and read its body. The protocol body — not your prior knowledge — is authoritative for thresholds, controlled_codes, authoritative_statuses, follow-up timing, ordering, dose rules, escalation triggers, and referral criteria. See reference/api_and_protocols.md for the case_type → protocol map and how each protocol's fields feed each answer section.

Read protocol constants at runtime; never hardcode them. Applying the protocol you fetched is what makes the answer correct and keeps the skill robust if the environment changes.

5. Derive each field, mapping to the template's controlled vocabulary

Work through the template's keys. For each, apply the protocol rule to the case data, then translate the raw clinical fact into the template's enum. The protocol's internal code names and the template's allowed_values are different vocabularies — map deliberately; emit only values the template allows. Example shape (not an answer): a saturation range in findings crosses a protocol threshold → choose the corresponding template red-flag / risk enum, not the protocol's own string.

6. Provenance and distractor discipline

This is where most errors happen. The environment is seeded with decoys.

• Target patient only. Use records whose patient_id matches the case's patient. Silently drop any wrong-patient record — it is not evidence and is not even a listed distractor.
• Authoritative status only. Treat only statuses in the protocol's authoritative_statuses (typically final) as satisfying gates. Exclude preliminary / entered-in-error / canceled for authoritative claims.
• Right code, right window. Match the target observation code; honor date windows as the protocol/template state them (commonly from inclusive, to exclusive).
• Latest-wins. When several qualifying same-code results exist, select per the protocol (usually the latest final effective_time in the window).
• Where the template asks for excluded/distractor ids, list the same-patient records that were relevant to review but disqualified by date, code, or status — sorted as the template directs.

7. Evidence ids

evidence_ids (or source_provenance) must be the real resource ids you actually relied on — e.g. observation_ids, imaging_id, renal/eGFR observation ids, the case_id, protocol id — in the order the template requests (often case id first, then clinical sources; or descending relevance). Never fabricate ids.

8. Safety-check booleans

Safety-check fields assert you did not put an unsupported claim in the answer. Set each to reflect the chart, e.g.: do not claim a normal chest X-ray when imaging shows consolidation; do not recommend a drug class the patient has an active allergy to; do not assert loss-of-consciousness / vomiting / photophobia the record marks absent. The boolean is true when your answer honors that constraint.

9. Format exactly

• Emit only the required top-level keys — no extras, no prose.
• Enums: only allowed_values. Lists that are "normalized as a set": no duplicates, order-insensitive. Lists with an ordering rule: sort exactly.
• Numbers: honor precision (integer hours/days; N-decimal values). Timestamps: ISO-8601 UTC with trailing Z.
• null only where the field spec permits it; otherwise supply a value.
• Output one JSON object and nothing else.

Final checklist

• Every required top-level key present; no extra keys.
• Every enum value is in the template's allowed_values.
• Constants (task_id, case_id, patient_id) correct.
• Findings, observations, imaging cross-checked against protocol thresholds — decision derived, not guessed.
• Only target-patient, authoritative-status, in-window, right-code records used; distractors excluded correctly.
• evidence_ids are real ids you used, ordered per the template.
• Safety checks honor the chart (no fabricated/normal/contraindicated claims).
• Numeric precision, timestamp format, nullability, ordering all match.
• Output is a single JSON object with no surrounding text.