Intake Coordination Audit

Intake Coordination Audit

What these tasks look like

Each task points at a shared read-only clinical intake portal and hands you an answer template (input/payloads/answer_template.json). The prompt names a single target cohort — a roster, a referral batch, a transfer batch, or a program code — and asks for one JSON object that classifies every member of that cohort into controlled-vocabulary fields, plus a cohort/summary block of integer counts. There is no free-form prose in the answer.

The whole task is deterministic: every output value is a function of records in the portal's data model. Your job is to (1) read the template as a contract, (2) pull the cohort's records, (3) apply the classification rules, (4) shape the JSON exactly as ordered, and (5) return JSON only.

The five task families you will see are described, with their confirmed decision rules, in references/decision-rules.md. The underlying data model (record types and the fields that drive each rule) is in references/data-model.md.

Method (applies to every task)

1. Treat the answer template as the spec. Enumerate its required top-level keys, and for every field note: its type, its allowed_values/enum, any required_value/constant (echo these verbatim — e.g. task_id, batch_id, roster_id, program_code), and its declared ordering. Never emit a value outside a field's allowed set.

2. Identify the cohort and its members. The prompt names the roster / batch / program. Load its member records (roster rows, referrals in the batch, transfers in the batch, or the program's candidate list). Include every member the portal returns — no more, no fewer.

3. Pull the joined records for those members. For each member gather the supporting rows the template's fields reference (coverage, PBM, pharmacy, lifestyle, clinical history, ICD metadata, documents, chart artifacts, facility capacity). A read-only query interface over the data model is the fastest way to reconcile these; prefer exact lookups over guesses.

4. Compute each field with the deterministic rules in references/decision-rules.md. Most fields decompose into: detect a set of boolean flags/issue codes from the raw fields, then map that flag set to an enum (status / tier / decision) and to any reason-code list.

5. Aggregate the cohort/summary block by tallying the per-member results you just produced. Keep it internally consistent (e.g. the status counts must sum to the member count; a per-member list's length must match its summary count).

6. Shape and order the JSON. Honor every ordering note (see references/output-format rules below). Return the JSON object only.

Output discipline (get these exactly right — they are cheap points)

• Echo constants verbatim. Any field with a required_value/constant (task_id, batch_id, roster_id, program_code) must match exactly, including case. IDs elsewhere use the portal's casing exactly (uppercase referral/transfer IDs as shown).
• Ordering matters. "ascending by " → sort ascending; "alphabetical by code/doc_type/artifact" → sort by the enum string; a summary cross-tab with a declared order → follow it. When a list is called an unordered set, its membership is scored but order is not — still emit it deterministically (sorted) for reproducibility.
• Enums only. Reason-code / issue-code / blocker-code arrays may only contain the template's allowed_values. Dates are YYYY-MM-DD. Counts are integers.
• Dates and windows use the cohort's own reference date (a roster's requested service date, a transfer's requested start date, a program's as-of date), not today's date, unless the template says otherwise.
• One JSON object, no prose, unless the template explicitly nests a text field.

Cross-cutting modeling notes

• Flags → enum, then enum + flags → decision. Separate detection (which issues exist) from classification (what status/decision that implies). This keeps reason-code lists and the headline enum consistent.
• Reason/issue/blocker code lists are usually the union of every detected issue for that record (coverage + prescription + pharmacy + demographic + risk, or coding + records + auth + duplicate + scheduling). Include a "high risk"/"met criteria" style positive code when the template offers one.
• current / final / finalized beats presence. A record that exists but is a draft, stale, pending, or not-finalized generally does not satisfy a "present/valid/complete" requirement — treat it as missing or invalid and emit the matching reason code.
• Set-membership and cross-references drive many flags: a service line inside a comma list, an ICD code's service_family/chapter versus the batch's specialty, an insurance id shared across different patients, a doc type present and finalized, a capacity row existing for a date.
• Confidence: the detection rules (dates, statuses, set membership, duplicates, capacity sums, wrong-specialty codes) are firm. The exact numeric thresholds and the tie-breaking between adjacent statuses (e.g. hold vs clinical_review, weekly vs biweekly, a freshness limit in days) vary by task — read the template's enum names and the data spread, and prefer the reading the data most cleanly supports.

See references/decision-rules.md for the per-family rule library and references/data-model.md for the record types and fields.