Clinic Protocol Cds

Clinic Protocol CDS

Answer a single synthetic-clinic decision-support task by grounding every field in the runtime record and the matching protocol, then returning exactly one JSON object that conforms to the task's answer_template.json.

What each task gives you

• input/prompt.txt — names a target case id (e.g. CASE-RESP-102), the case family in prose, and the fields to populate. It references <TASK_ENV_BASE_URL> for the runtime and says the environment access is "listed separately".
• input/payloads/answer_template.json — a schema descriptor (not a fill-in form). It defines required_top_level_keys, per-field type, allowed_values (enums), numeric precision, unit, nullability, ordering rules, and sometimes an expected_constant / required_value for task_id / case_id.
• An environment-access file (e.g. environment_access.md) — gives the base URL as GDPEVO_ENV_BASE_URL=..., the credentials (often none), and the exact allowed endpoints for this run. Use it only to reach the network.

Your output must contain only the JSON object — no markdown, no comments, no prose, no extra top-level keys.

Procedure (case-family agnostic)

1. Read the template first. Extract the ordered required_top_level_keys, every enum's allowed_values, precision/unit, which fields may be null, and all ordering rules. Note any expected_constant/required_value for task_id, case_id. The template is the contract; the output shape is driven entirely by it, not by this doc.

2. Set the identifiers.

   • task_id: use the template's expected_constant/required_value when given; otherwise use the task's own identifier (the input/run/directory id).
   • case_id: the target case id from the prompt (confirm against any template constant).
   • patient_id: read it from the case record — never assume or reuse a value from an example.

3. Pull the case bundle. GET {base}/api/cases/{case_id} returns one bundle with case, patient, findings[], observations[], imaging[], medications[], allergies[], problems[], care_registry, sdoh[]. This is usually all you need. case.case_type selects the protocol (see reference/protocol_playbook.md).

   • The findings[] array (finding_key / finding_value / source_id) is the curated evidence layer — most enum decisions and evidence_ids come from it.
   • For patient-wide retrieval (observation-window tasks especially), also GET {base}/api/observations?patient_id={patient_id} so you catch every eligible and distractor observation, not just those already attached to the bundle. GET {base}/api/patients/{patient_id} gives patient-level allergies, meds, and problems.

4. Fetch and apply the protocol. GET {base}/api/protocols then GET {base}/api/protocols/{protocol_id}. Read it at runtime and use its thresholds/codes/status rules — do not rely on memorized constants; versions change. reference/protocol_playbook.md summarizes the current bodies and the case_type→protocol map as an orientation, but the live protocol wins.

5. Derive each templated field from evidence + protocol. Map findings and final observations onto the template's enums. Choose enum values only from the template's allowed_values. Honor precision, units, nullability, and ordering.

6. Emit one JSON object with exactly the required keys, then run the verification checklist below before returning.

Environment access

• Base URL comes from the environment-access file's GDPEVO_ENV_BASE_URL. Do not hardcode a host; read it per run. Endpoints join as {base}/api/....
• Read-only. Never mutate the environment or place orders. Only the GET endpoints are usable when Credentials: none — POST /api/query requires a clinic token that is not provided and returns invalid or missing clinic token. Everything needed is reachable via the GET endpoints, chiefly GET /api/cases/{case_id}.
• List endpoints (/api/observations, /api/medications, /api/allergies, /api/problems, /api/imaging, /api/care-registry, /api/sdoh) return {count, items, limit, offset} and accept filters like ?patient_id=.
• Helper: scripts/fetch_clinic_case.py <CASE-ID> pulls the bundle, patient-filtered observations, and all protocol bodies into one JSON blob to work from (reads the base URL from GDPEVO_ENV_BASE_URL or --base).

Cross-cutting output rules

• Ground everything; fabricate nothing. Every asserted finding, value, and id must trace to the record. If the record does not support a red flag / symptom / normal result, do not claim it. The safety_checks booleans exist to attest you did not assert unsupported findings (e.g. no false loss-of-consciousness, no "normal CXR" claim when imaging is abnormal) — set each to true only when your answer genuinely honors that constraint, which a grounded answer always does.
• Enums only. For scored/status/action fields, emit a value from the template's allowed_values — never free prose.
• Only final observations satisfy protocol gates. Every protocol lists authoritative_statuses: [final]. Exclude preliminary, entered-in-error, canceled, and any non-final status from matches, latest-value selection, and clinical conclusions.
• Allergy-awareness. Use active allergies only (ignore inactive). Map them to the template's allergen classes for avoid_allergens, and pick a medication/antibiotic strategy that avoids every implicated class.
• Numeric precision & units. Round exactly as the template says (e.g. one decimal for mmol/L, two decimals for a risk probability, integer hours/days/ mEq). Keep the specified unit semantics.
• Timestamps. ISO-8601 UTC with a trailing Z (e.g. 2026-03-20T07:45:00Z). Use the record's current_time finding as the clinical review time when a field needs "now". For date windows: from inclusive, to exclusive (a calendar month → first day 00:00:00Z of the month to first day 00:00:00Z of the next month).
• Ordering. Most lists are sets — "no semantic ordering", emit each value at most once. Where the template specifies a sort (e.g. observations by effective_time ascending then observation_id ascending; evidence with case id first), apply it exactly. When an "excluded" list mixes same-target-code and other-code distractors, group same-target-code exclusions before other-code ones, each still ordered by the template's key.
• Nullability. Use null only where the field spec permits it (e.g. no medication in this setting → null dose/route). Do not invent placeholder strings for absent data.
• evidence_ids. List the stable identifiers you actually used to support the decision — typically the case id plus the specific observation / imaging / registry / protocol source_ids (from findings[].source_id and resource ids), not every record present. Follow any template ordering (case id first when included).
• No extra keys. Include every required key and nothing beyond what the template allows.

Distractor traps to avoid

• Distractor cases CASE-D30xx and unrelated real cases share case types — only operate on the prompt's target case id and its patient.
• Wrong patient / wrong case observations: cross-check patient_id (and case_id where relevant) on every resource.
• Non-final status results (e.g. a preliminary potassium) look eligible but are excluded.
• Wrong code: an observation for a different analyte (sodium when you need potassium K; a blood potassium code vs. the controlled serum K code) is a distractor. Use the protocol's controlled_codes for the exact code.
• Out-of-window dates: a final, correct-code result dated outside the window does not qualify.
• Inactive allergies / resolved problems: filter by status.

Verification checklist (before returning)

• Exactly one JSON object; no prose/markdown/comments; no extra top-level keys.
• All required_top_level_keys present, in a valid shape.
• task_id / case_id match any template constant; patient_id came from the record.
• Every enum value is in the template's allowed_values.
• Numbers meet precision/units; timestamps are ISO-8601 Z; windows are [from, to).
• Only final observations used for gates; correct controlled codes; in-window.
• Active allergies avoided; medication plan consistent with them.
• Lists respect set-vs-ordered rules; no duplicates in sets.
• evidence_ids are real ids actually used; ordering rule applied.
• safety_checks booleans truthfully reflect that no unsupported claim was made.
• Nothing fabricated: every asserted value traces to the retrieved record.