Cedar Ridge Intake Coordination

Cedar Ridge Intake Coordination — reusable operating guide

Every task in this family has the same shape:

1. A short prompt names one work unit — a roster (NPI-…), a referral batch (ORTHO-…, PULM-…, CARD-…, …), a transfer batch (DIAL-…), or a program (DMHTN-…, RENAL-…, COPD-…, CAD-…) — and points at the Cedar Ridge Intake Coordination Portal at <TASK_ENV_BASE_URL>.
2. input/payloads/answer_template.json defines the exact output shape and the controlled vocabulary (allowed enum values, required keys, ordering).
3. You must read portal data, classify each item, roll up cohort counts, and return one JSON object, JSON only, no prose, that conforms to the template.

The template is the contract. The portal is the source of truth. Your job is to join the right tables, apply the classification rules, and reconcile the counts.

1. Golden rules (do these every time)

• Re-read the task's answer_template.json first. Treat its required_*_keys, allowed_values/allowed, required_value/constant, and ordering as authoritative. Never emit a key or enum value not listed there. The controlled vocabulary is self-describing — a code named coverage_expired, missing_imaging, or consent_missing tells you exactly the condition it represents.
• Parameterize by the id in the prompt/template. The work-unit id (roster_id/batch_id/program_code) and any constant task_id come from the prompt and the template's required_value/constant/expected_value fields — copy those verbatim. Do not hardcode ids or dates from prior runs; a held-out task will use a different id (e.g. NPI-JUL-02, ORTHO-JUL-04, DIAL-FALL-03, RENAL-DM-2026B).
• Cover exactly the roster's members. Include a row for every member the portal returns for that work unit, and no others. For rosters, the member list and the requested_service_date + service_line come from the intake_rosters rows for that roster_id (not from the prompt text).
• Ordering matters. Sort every list exactly as the template says (usually ascending by patient_id / referral_id / transfer_id / group_id). Where the template says "unordered set" for a code array, order is not graded but stay consistent; where it says alphabetical (e.g. dialysis missing_required_documents, activation artifacts_to_create), sort the strings.
• Reconcile counts. Every summary/cohort count must equal what you can recompute from the item list (per status, per risk, per urgency, totals). A mismatch between the detail rows and the summary is the most common failure. Cross-tab lists (e.g. counts_by_urgency_and_status) should enumerate the template's key order and may include explicit zero rows if the template's key set is fixed.
• Emit JSON only. No markdown, no commentary, no trailing text. IDs keep the portal's exact casing (uppercase REF0001, P001, etc.).
• Dates are YYYY-MM-DD strings. Compare dates as calendar dates. Use the work unit's own reference date (roster requested_service_date, transfer requested_start_date, program candidate_date/as-of) — see each playbook.

2. Reaching the portal

Base URL is provided per task (environment_access.md → GDPEVO_ENV_BASE_URL, shown in prompts as <TASK_ENV_BASE_URL>). Credentials: none. Everything is read-only.

Two equivalent ways to read; prefer whichever makes the join cleanest. The SQL endpoint is usually fastest for batch joins and exact reconciliation.

• REST (GET): /patients, /patients/{id}, /referrals (filter ?batch_id=&service_line=&limit=), /referrals/{id}, /transfers (?batch_id=), /transfers/{id}, /documents, /chart/{patient_id} (bundles patient + chart_artifacts + clinical_history), /icd/{code}, /pharmacies, /programs/{program_code}/candidates. GET /health returns table row counts — a useful sanity check.
• Read-only SQL: POST /query with body {"sql": "SELECT …"}. Only SELECT / read-only PRAGMA are allowed (writes are rejected). Returns {columns, rows, row_count, truncated}. Use LIMIT and paginate if truncated is true.

Example:

curl -s -X POST "$BASE/query" -H 'Content-Type: application/json' \
  -d '{"sql":"SELECT * FROM referrals WHERE batch_id = ?BATCH? ORDER BY referral_id"}'

The full table schemas, column enums, and per-archetype join recipes are in references/data_model.md. The step-by-step classification logic for each of the four archetypes is in references/task_playbooks.md. Read both before writing output.

3. Identify the archetype, then follow its playbook

Match the prompt/template to one of four archetypes and apply the matching section of references/task_playbooks.md:

Archetype B has two closely related variants — a batch audit (duplicate groups, shared-insurance anomalies, action plan, urgency×status summary) and a referral-to-chart activation file (readiness, chart needs, correspondence queue, priority order). Both read the same referrals/icd_codes data; let the template's top-level keys tell you which outputs to produce.

4. General decision scaffolding (applies across archetypes)

These structural patterns recur; the playbook gives the field-level specifics.

• Readiness / registration / decision status is the worst-severity outcome across that item's blockers. Something with a hard blocker is blocked/rejected/hold; something clean is ready/approved/accept; items needing human judgement are under_review/clinical_review; purely clerical items are admin_followup. Map to the exact enum the template lists.
• A code array is the set of every condition that is true for that item, drawn only from the template's allowed list. Empty list when nothing applies.
• Priority tiers: clinical/urgent + blocked → tier_1_immediate; substantive but non-urgent follow-up → tier_2_short_term; clerical/administrative → tier_3_administrative. Some templates allow null tier for already-ready items — check enum_or_null.
• Missing vs. invalid vs. stale: missing = the row/record does not exist; invalid = it exists but fails its rule; stale = it exists and is otherwise valid but older than its freshness window relative to the reference date.
• Contact routing / outreach channel follows the member's stated preference (patients.preferred_contact or program_candidates.preferred_outreach), and falls back / becomes none when the preferred channel's contact value is absent.

Always finish by recomputing the summary from your own item rows and confirming every template-required count key is present (including zeros).

5. Self-check before returning

• Output is a single JSON object; nothing outside it.
• Top-level constant fields (task_id, batch_id/roster_id/ program_code, etc.) equal the template's required/constant values.
• One item per work-unit member; ids in the template-specified order.
• Every enum/code value appears in the template's allowed list.
• Code/reason arrays contain no duplicates; alphabetized where required.
• Every summary count reconciles with the detail rows; all required count keys present (zeros included).
• Dates are YYYY-MM-DD; comparisons used the correct reference date.