Clinic Protocol
Overview
This skill covers structured protocol-driven clinical decision support against a synthetic clinic REST API. The solver reads a target case, retrieves protocol rules, applies those rules to clinical observations and patient context, and returns a single JSON object with controlled values.
Core Workflow
Every task follows the same five-step pattern:
Step 1: Read the answer template
The prompt directs you to input/payloads/answer_template.json. Read it first. The template defines every required top-level key, allowed enum values, field types, numeric precision rules, and output constraints. The template is the authoritative shape for your answer — never add extra keys or use prose where an enum is required.
Step 2: Retrieve the case and protocol
Use GET /api/cases/{case_id} at TASK_ENV_BASE_URL. This endpoint returns a composite object containing the case header, patient demographics, observations, findings, imaging studies, medications, allergies, problem list, care-registry entry, and SDOH records in a single response. This is almost always the only endpoint you need for clinical data.
If the task references a specific protocol, retrieve it with GET /api/protocols/{protocol_id}. The protocol body contains the clinical decision rules: thresholds, trigger values, controlled LOINC/RXNORM codes, exclusion criteria, and escalation conditions. Apply protocol rules literally — protocols are deterministic, not advisory.
Step 3: Extract and filter clinical facts
From the composite case response, extract:
- Observations: Filter by
status: "final"unless the template explicitly accepts other statuses. Protocol rules always usestatus: "final". Sort by effective_time when ordering matters. - Findings: Free-text clinical facts (chief complaint, symptom presence/absence, social context). Treat negated findings (
"absent") the same as absent values. - Allergies: Check
status: "active"only. Map allergens to the medication class avoidance list in the template. - Imaging: Check
status: "final"for authoritative impressions. - Medications: Count active medications for polypharmacy triggers.
- Problems: Active problems inform risk tiering and program eligibility.
- Care registry: When present, use risk_score and program_hint as inputs to routing logic.
- SDOH: SDoH domains and severity levels drive referral and outreach decisions.
Step 4: Apply the protocol rules
Protocol documents define decision boundaries as explicit thresholds and trigger lists. Apply them mechanically:
- For numeric thresholds: compare the observation value to the protocol threshold (e.g., potassium < 3.0 triggers the urgent branch).
- For trigger lists: check whether findings/observations match any item in the protocol's trigger set.
- For dose rules: compute using the protocol's stated formula (e.g., mEq per 0.1 mmol/L below target, rounded to nearest step).
- For status filtering: treat preliminary, canceled, and entered-in-error observations as excluded unless the template explicitly says otherwise.
- For ordering: protocols specify ascending effective_time, then observation_id as tiebreaker.
Step 5: Produce the JSON answer
Build a single JSON object with exactly the required top-level keys from the template. Follow these rules:
- Use controlled enum values from the template's allowed_values lists. Do not invent new enum values.
- Use
nullonly where the template permits it (e.g., medication fields when no medication is recommended). - Booleans never have quotes — they are JSON
true/false, not strings. - Include evidence_ids as an array of identifiers for the case and the key observations/imaging used.
- For list fields, ordering is not semantically scored unless the template explicitly states otherwise. Use a stable, reproducible order.
- Do not wrap the JSON in markdown, code fences, or explanatory text.
- Return only the JSON object.
API Conventions
- Base URL: Always
${TASK_ENV_BASE_URL}as provided in the prompt. - Composite case endpoint:
GET /api/cases/{case_id}— use this as the primary data source. - No authentication: The task environment is credential-free.
- Distractor data: The API returns distractor records mixed with target records. Always filter by the target case_id and patient_id.
See references/api-reference.md for the full endpoint listing and response shapes.
Protocol-Specific Guidance
See references/protocol-solver.md for detailed decision logic for each protocol type. Read it when the template reveals the protocol domain but the clinical reasoning is not obvious from the protocol endpoint alone.
Common Pitfalls
- Allergy-blind medication plans: Always check active allergies before selecting an antibiotic or medication class. Map allergen strings to the avoid_allergens enum values in the template.
- Sodium observations in potassium windows: When the target code is
K, observations with codeNAare near-target distractors — exclude them. - Wrong-patient observations: The case endpoint may return observations with a different patient_id. Filter strictly by the target patient_id.
- Preliminary observations: Never use preliminary or canceled observations as the clinical basis for a protocol gate or numeric anchor unless the template explicitly permits it.
- Distractor records: Generic distractor cases use IDs like
CASE-D####and can be ignored. Target cases use semantic IDs likeCASE-RESP-102. - Safety checks: When the template requires safety-check booleans, verify each condition against the clinical data. Set
trueonly when the evidence supports the constraint.