Clinic Protocol Decision Support

Clinic protocol-bound decision support (structured JSON)

What this family of tasks looks like

Each task gives you:

1. A prompt naming exactly one target case id and the clinical deliverable (an assessment, disposition, plan, routing, or a retrieval/gate result).
2. An answer template (usually input/payloads/answer_template.json) that is a strict contract: required top-level keys, enums (allowed_values), types, numeric precision, ordering rules, required constant values, and null rules.
3. A separately-provided environment-access file listing the runtime base URL and the allowed read-only endpoints for this run.

Your job: read the clinic data, apply the applicable clinic protocol, and return exactly one JSON object that conforms to the template — no markdown, no prose, no extra keys.

Do not mutate the environment or "place orders." Everything is read-only (GET).

Workflow

1. Parse the task

• Extract the target case id verbatim. The environment contains distractor cases and distractor records; only the named case is in scope.
• Identify the domain from the prompt/summary so you know which protocol governs it.
• Note any explicit constants the template pins (e.g. a fixed task_id, case_id).

2. Treat the answer template as a contract (read it before touching data)

• List every required_top_level_keys entry — output all of them, and only them.
• For each field record: type, allowed_values (enums), precision, ordering rules, required_value/expected_constant, and where null is permitted.
• Objects usually must be fully correct to earn credit — plan to get every sub-field right, not just most.

3. Pull the data (read-only)

• Read the run's environment-access file for the base URL and allowed endpoints; do not hardcode a URL — it is provided per run.
• Prefer the case-detail endpoint (GET .../api/cases/{case_id}). It returns a consolidated bundle: case, patient, allergies, problems, medications, observations, imaging, findings, care_registry, sdoh. The findings list often spells out the decision-relevant facts (current time, thresholds, disclosed barriers) with their source_id.
• List protocols (GET .../api/protocols), find the protocol_id whose title matches the domain, and fetch its body (GET .../api/protocols/{protocol_id}). The body holds the thresholds, controlled codes, formulas, and ordering rules you must apply — read them at runtime; do not assume fixed numbers.
• If an endpoint returns 401/needs a credential you weren't given, skip it and use the open GET endpoints. The case bundle is normally sufficient.

4. Filter to the authoritative, in-scope record set

Distractors are deliberate. Before using any observation/result, require all of:

• Status is authoritative — protocols list authoritative_statuses (typically final). Drop preliminary, entered-in-error, canceled.
• Code matches the target exactly (e.g. a serum code vs a whole-blood code vs a different analyte are different codes). Use the protocol's controlled_codes.
• Patient equals the target patient. Records for another patient are not part of the review — exclude them from every list (matched and "excluded distractors").
• Time window (if any) is respected: inclusive start, exclusive end.
• When you need "the latest/current" value, take the latest eligible record by effective_time after applying the filters above.

See references/retrieval_and_protocol_rules.md for the detailed retrieval rules.

5. Apply the protocol deterministically

• Evaluate escalation/urgent branches first. Protocols encode explicit trigger sets (vital thresholds, critical values, high-risk conditions, symptoms). Check each trigger against the filtered data. If none fire, take the routine/outpatient branch. The disposition, imaging, and urgent-action fields follow from this.
• Let the protocol's own "support" criteria pick the label. When a protocol lists criteria that support a particular assessment/tier and the case meets them, choose that label. Matching the protocol's language to the right enum is the single highest-value decision — get the primary categorical right first.
• Apply formulas exactly (e.g. dose = increments-below-target × per-increment amount, then round as specified), using the eligible latest value as input.
• Be allergy-aware: use only active allergies; avoid the implicated drug classes; pick a regimen compatible with them; list avoided classes as a set.
• Tier by threshold: map scores/values to tiers using the protocol's cutoffs.

6. Build and check the output

Emit one JSON object. Then verify it against the template contract and the output-conformance rules below.

Output-conformance rules (these move the score)

• Exactly the required keys, no extras, no prose, valid JSON. Honor pinned constants and use only allowed_values for enums.
• Numeric precision: respect stated decimals (one/two places) and integer units (whole hours/days/count). Format composite strings exactly as the template shows (e.g. a blood pressure as "systolic/diastolic").
• Canonical clock times: when a follow-up is described qualitatively (e.g. "next morning"), schedule it on the correct calendar day at a canonical morning hour (08:00) with a trailing Z. The exact hour is graded — don't invent an odd time.
• Sets are scored as exact sets — be evidence-tight, not inclusive. For list/enum fields (red flags, recommended tests, referrals, priority problems, escalation conditions), include only items with clear support in the data or protocol. Over-including weakly-supported items (e.g. a mild screening score → a specialty referral) loses points. Prefer minimal-but-complete, protocol-literal sets. When a field asks for absent/negative findings, list the full set of serious findings that are documented or clinically confirmed not present.
• Required structured minima can be true by convention. Some booleans a compliant plan requires (e.g. a person-centered care plan requiring a member-stated priority) should be true even if the protocol's bulleted list doesn't spell them out. Don't set a required-by-domain flag false just because the protocol text omits it — and don't overwrite a well-reasoned structured object with a guess.
• Provenance splits: separate objective chart-derived facts from member-disclosed facts per the field's allowed_values; include only facts actually present (omit a fact key when its source datum is absent).
• Safety-check booleans: set each true when your answer upholds the condition (no contraindicated drug prescribed; no false claim of normal imaging / clear exam / an absent symptom). Make the rest of the answer consistent with them.
• Evidence/id lists: include the decision-relevant identifiers (target case id plus the observation/imaging/protocol ids you relied on) and honor any stated ordering (case first, or descending relevance, or effective_time ascending). These carry little scoring weight — spend your effort on the decision fields.

Priorities when time is limited

1. Identifiers + pinned constants (free points — verify the patient id from the bundle).
2. The primary categorical assessment/tier/disposition/gate (highest weight).
3. Numeric anchors and formula outputs at the required precision.
4. Fully-correct nested objects (medication order, follow-up, contraindications).
5. Tight, protocol-literal sets.
6. Evidence/id lists last.

See references/domain_playbooks.md for per-domain checklists (respiratory, head injury, electrolyte repletion, care-management routing, observation window).