cedar-ridge-intake-audit

name	cedar-ridge-intake-audit
description	Audit and reconcile a batch, roster, or program against the Cedar Ridge Intake Coordination Portal (REST + read-only SQL), then emit one template-conformant JSON object. Use when a task points at the Cedar Ridge portal and supplies an answer_template.json describing the required output shape.

Cedar Ridge Intake Coordination — Batch Audit & Reconciliation

This skill solves a family of intake-coordination audit tasks against the Cedar Ridge Intake Coordination Portal. Each task scopes work to one identifier — a roster (new-patient access verification), a referral batch (readiness / activation audits), a transfer batch (dialysis packet review), or a program code (chronic-care enrollment panel) — and demands a single JSON object whose shape and controlled vocabularies are fixed by an answer_template.json.

The work is always: read the template → pull the complete scoped record set → reconcile and cross-reference → classify each record with template-controlled values → aggregate cohort summaries → emit one JSON object that matches the template exactly.

These instructions are reusable. They contain no task-specific final values (no patient classifications, no referral readiness, no computed counts). Apply the procedure to whatever scope key and template the task gives you.

When to use

Use this skill when a task:

• Names the Cedar Ridge Intake Coordination Portal and gives a <TASK_ENV_BASE_URL> placeholder, and
• Hands you an input/payloads/answer_template.json (the output contract) and a scope key — a roster_id, batch_id, or program_code.

If either is missing, you do not have a Cedar Ridge intake-audit task — do not use this skill.

Ground rules (read first)

1. Resolve the base URL from environment_access.md only. Do not invent a URL. The portal is on the network at the base URL recorded there (no auth). Use only the endpoints listed in that file. Never hit /health or any reset/reseed endpoint.
2. The template is the contract. Read input/payloads/answer_template.json before you query anything. It defines required top-level keys, per-record fields, every controlled vocabulary (allowed_values), list-ordering rules, and summary count buckets. Every value you emit must come from the template's allowed sets.
3. One JSON object, no prose. The final answer is a single JSON object. No commentary before or after. No extra fields. Use IDs exactly as the portal shows them (uppercase).
4. Read any extra payload. Some tasks include a second payload (e.g. a roster file listing target IDs and a note like "use the roster record in the environment for requested_service_date and service_line"). It tells you where to find data, not the values themselves.

Critical portal gotchas (these cost correctness if missed)

• GET list endpoints silently cap at 100 rows. There is no truncated flag on GET responses, and the count field is the number returned, not the total. A batch with many documents, or the patients directory itself (which exceeds 100 rows), will be silently truncated. Always pass an explicit large limit (e.g. ?limit=1000) or filter by batch_id/patient_id, and cross-check with a SQL SELECT COUNT(*).
• Distractor records are everywhere. The portal mixes many batches/rosters. Out-of-scope records are deliberately present (often literally labelled "distractor"). Filter strictly to your scope key; never include a record whose batch_id/roster_id/program_code does not match.
• The SQL body key is sql, not query. POST /query with {"query": ...} returns {"error":"sql is required"}. Send {"sql": "SELECT ..."}. Check the truncated flag in the SQL response — if true, narrow the WHERE clause or split the query.
• Endpoint ↔ SQL table names differ. /transfers ↔ table transfer_requests; /icd/{code} ↔ icd_codes; /programs/{code}/candidates ↔ program_candidates; /patients/{id} bundles patients+coverage+pbm+patient_pharmacy/pharmacies+lifestyle+clinical_history+chart_artifacts+documents+program_candidates+referrals. See reference/portal_endpoints.md.

Procedure

Phase 1 — Parse the brief and the contract

1. Identify the scope key and type: roster_id (new-patient access), batch_id (referral readiness / activation, or transfer review), or program_code (enrollment panel). Note the literal value.
2. Read answer_template.json. Extract: required top-level keys; the per-record item schema and its allowed_values; list-ordering rules (ascending by id, alphabetical by code, or "unordered set"); the summary object's required count buckets; and any required_value/constant you must echo verbatim (e.g. task_id, batch_id, roster_id, program_code).
3. Read any second payload for IDs or data-location hints.
4. Note which entities the task needs (patients, referrals, transfers, documents, charts, ICD, pharmacies, programs, capacity) — that drives which endpoints you call.

Phase 2 — Pull the complete, scoped record set

1. Fetch the scoped list for your entity, overriding the 100-cap: e.g. GET /referrals?batch_id=<scope>&limit=1000, GET /transfers?batch_id=<scope>&limit=1000, GET /programs/<scope>/candidates. For a roster, fetch roster rows via SQL: SELECT patient_id, requested_service_date, service_line FROM intake_rosters WHERE roster_id='<scope>'.
2. Cross-check completeness: run SELECT COUNT(*) FROM <table> WHERE <scope column>='<scope>' and confirm it equals the number of records you pulled. If a list still looks capped, switch to SQL for that set.
3. For each in-scope record, fetch the detail bundle that carries the per-record assessment data:
   • New-patient access (roster): GET /patients/{id} → coverage (insurance), pbm (prescription benefit), pharmacies (preferred-pharmacy network), lifestyle (lifestyle risk), clinical_history (overall risk), plus emergency_contact_present/address/preferred_contact for contact blockers.
   • Referral audits: GET /referrals/{id} → referral (service_line, urgency, auth_status, records_received, imaging_received, appointment_scheduled, insurance_id, icd10_code, diagnosis_description, referral_reason), patient, icd (chapter, service_family, laterality), documents.
   • Transfer review: GET /transfers/{id} → transfer (requested_start_date, modality), patient, documents (the packet), capacity (chair availability for the requested start).
   • Enrollment panel: GET /programs/<code>/candidates then GET /chart/{id} per candidate → active_problems, recent_vitals_labs, meds_allergies, clinical_history (recent_hospitalization, risk_flags), plus candidate fields consent_status, adherence_score, existing_chart, target_condition, preferred_outreach.
4. When an endpoint omits a field you need, fill it with a SQL join (e.g. , , ).

Phase 3 — Reconcile and cross-reference (the audit layer)

• Coding discrepancies (referral tasks): for each referral, compare the ICD chapter/service_family against the referral's service_line (icd_chapter_mismatch), the ICD description/laterality against the referral narrative (diagnosis_description, referral_reason) (narrative_mismatch), and the ICD laterality against any laterality in the narrative (laterality_mismatch). Use GET /icd/{code} (or the icd_codes table) for the metadata.
• Duplicates: group referrals by patient_id (same patient, same service family). Designate the primary (earliest/lowest id) and recommend consolidation; non-duplicates stay separate.
• Shared-insurance anomalies: group by insurance_id; if one insurance_id maps to multiple distinct patient_ids, flag it; if it maps to the same patient across duplicate referrals, that is legitimate.
• Blockers: check records_received, imaging_received, auth_status (pending/denied/not_submitted), and appointment_scheduled (already-scheduled referrals need review, not re-booking).
• Transfer packet completeness & freshness: assemble the packet from documents for the transfer_id; compare against the required doc set in the template; for doc types the template marks as freshness-sensitive, compare received_date to the per-doc freshness limit and list stale ones with their limit.
• Capacity feasibility: sum open_chairs for the requested_start_date + modality across locations (the /transfers/{id} capacity bundle, or the facility_capacity table). Feasibility combines packet-readiness with capacity availability per the template's feasibility enum.
• Enrollment eligibility: check active DM/HTN diagnosis, consent status, chart currency/active-problems/vitals/labs/meds recency, recent hospitalization/ED and adherence (high-touch triggers), CKD (biweekly monitoring), and target-condition match.

Phase 4 — Classify each record with template-controlled values

Map the evidence to the enum values the template allows — never invent codes:

• Access verification: insurance_status, prescription_status, pharmacy_status, lifestyle_risk, overall_risk, registration_status, and the blocked_reason_codes set.
• Referral audits: readiness_status (ready/blocked/under_review/admin_followup), issue_codes/blocker_codes, priority_tier.
• Transfer review: packet_completeness_status, feasibility, final_intake_decision, next_contact_owner, next_contact_route.
• Enrollment panel: eligible, enrollment_status, reason_codes, follow_up_cadence, missing_chart_artifacts, outreach_channel, initial_monitoring_package. Apply the task's domain decision logic (e.g. overall_risk_high → hold/clinical_review; auth denied or missing records → blocked; consent declined → reject). If evidence doesn't cleanly fit, choose the closest allowed value and keep summary counts consistent with the per-record choices.

Phase 5 — Aggregate cohort summaries

• Derive every summary count from the per-record classifications in Phase 4 — do not recompute independently.
• Buckets must reconcile: the sum of each status/risk/decision bucket equals the total record count. Cross-tabs (e.g. counts_by_urgency_and_status) must sum to the same total. All counts are integers.
• Include exactly the summary keys/buckets the template lists — no more, no less.

Phase 6 — Emit and self-validate

1. Build one JSON object with exactly the template's required top-level keys.
2. Set the scope/identity fields (task_id, batch_id, roster_id, program_code) to the template's required_value/constant, verbatim.
3. Order every list as the template specifies. Treat reason/blocker/issue code arrays as unordered sets; order id lists ascending; order stale-document and artifact lists alphabetically.
4. Use only allowed_values. No extra fields. No prose outside the JSON.
5. Run the reference/output_checklist.md before submitting: enums in-range, required keys present, lists ordered, distractors excluded, summary buckets sum to totals, and the records-reviewed count matches the SQL scope count.

Reference files

• reference/portal_endpoints.md — every endpoint's request/response shape, the SQL endpoint, the REST↔SQL table map, and the 100-cap / distractor / truncated details.
• reference/output_checklist.md — pre-submit validation steps for template conformance.

Scope of this skill

This skill covers the five training archetypes — new-patient access verification, orthopedic referral readiness, dialysis transfer review, chronic-care enrollment panel, and pulmonary referral-to-chart activation — and generalizes to any future Cedar Ridge intake-audit task that supplies a scope key and an answer_template.json. It encodes the procedure and portal behavior only; it does not contain any task's computed answers.