Cedar Ridge Intake Audit

Cedar Ridge Intake Audit

This skill completes Cedar Ridge intake-coordination audit tasks. Each task gives a prompt plus an input/payloads/answer_template.json that defines the exact JSON shape required, and asks you to query the Cedar Ridge Intake Coordination Portal and return one JSON object, nothing else.

The portal holds the source of truth; the template holds the output contract. Your job is: read the prompt → identify the work family → pull portal records → apply the decision rules → emit template-conformant JSON.

Operating procedure (every task)

1. Read the prompt. Identify the work family and the target identifier:
   • Patient access verification → a roster id (e.g. NPI-*) + patient list + service line.
   • Referral readiness audit → a batch id (e.g. ORTHO-*, PULM-*) of referrals.
   • Dialysis transfer review → a transfer batch id (e.g. DIAL-*).
   • Chronic-care enrollment panel → a program code (e.g. DMHTN-*).
   • Referral-to-chart activation → a batch id of referrals.
2. Read answer_template.json end to end. It is the contract: required top-level keys, required fixed values (e.g. task_id, batch_id, roster_id, program_code), per-field enums, list orderings, and required count keys. Every value you emit must come from the template's allowed values; every list must use the template's ordering; every required key must be present.
3. Resolve the base URL from environment_access.md (the prompt's <TASK_ENV_BASE_URL> placeholder). Use only the endpoints listed there. Do not hit /health or any reset/reseed endpoint.
4. Pull the data. Prefer the read-only SQL endpoint (POST /query) for bulk/relational queries; use the REST endpoints for per-record detail. See reference/schema.md for the full table map and request formats.
5. Apply the decision rules for the work family (see reference/decision_rules.md). Compute each field from the raw records. Watch for distractor records that belong to a different service line/batch/program — only act on records relevant to the target identifier.
6. Assemble the JSON exactly to the template: fixed required values verbatim, enums only from allowed sets, lists ordered as specified, arrays that the template calls "unordered set" sorted for determinism (typically ascending by id or alphabetical by code — match the template's stated ordering where given).
7. Emit JSON only. No prose, no markdown fences, no commentary. A single JSON object.

Portal access (essentials)

• Base URL: from environment_access.md (e.g. http://task-env:9013/). No auth.
• GET / — landing page (lists available batches/programs; useful for discovery).
• GET /patients, GET /patients/{patient_id} — patient identity + assembled record.
• GET /referrals, GET /referrals/{referral_id} — referral rows (supports ?batch_id=).
• GET /transfers, GET /transfers/{transfer_id} — transfer requests (supports ?batch_id=).
• GET /documents — packet documents (filter client-side by referral_id/transfer_id/patient_id).
• GET /chart/{patient_id} — assembled chart (active_problems, meds_allergies, recent_vitals_labs, chart_artifacts, clinical_history, patient).
• GET /programs/{program_code}/candidates — program candidate list.
• GET /icd/{code} — ICD-10 metadata (description, chapter, service_family, laterality).
• GET /pharmacies — pharmacy directory (network_status).
• POST /query — read-only SQL. Body {"sql":"SELECT ..."}. Returns {"columns","rows","row_count","truncated"}. Use single-quoted SQL string literals (e.g. WHERE batch_id='<batch_id>') so inner quotes do not break the JSON payload. Only SELECT. If truncated is true, re-query with a tighter filter or LIMIT paging.

Full schema and column meanings: reference/schema.md.

Output discipline (non-negotiable)

• One JSON object. No text before or after. No code fences.
• Fixed required values (task_id, batch_id, roster_id, program_code, requested_service_date source, etc.) must match the template exactly — copy them from the prompt/template, not from memory.
• Enums only. If a field has allowed_values, never invent a value outside that set. Unknown/missing → use the template's unknown/none/null/not_applicable option where one exists.
• Ordering. Lists whose ordering is specified (e.g. "ascending by referral_id", "alphabetical by code") must be sorted that way. "Unordered set" arrays should be sorted deterministically (ascending id / alphabetical) for reproducibility.
• Counts are integers. Summary count objects must include every required key (even zero counts) with integer values.
• Dates are YYYY-MM-DD strings.
• IDs are uppercase strings exactly as the portal returns them.

Work families and where the rules live

Each family has its own decision logic in reference/decision_rules.md. Summary:

• Patient access verification (NPI-* roster, primary care intake): per patient compute insurance_status, prescription_status, pharmacy_status, lifestyle_risk, overall_risk, registration_status, blocked_reason_codes, plus a cohort summary. Driven by coverage, pbm, patient_pharmacy+pharmacies, lifestyle, clinical_history, patients, and the roster's service_line/requested_service_date.
• Referral readiness audit (ORTHO-* etc.): per referral compute readiness_status, issue_codes, priority_tier; plus batch-level icd_discrepancies, duplicate_groups, shared_insurance_anomalies, blocker_sets, ready_to_schedule, action_plan, and summary counts. Driven by referrals, icd_codes (for discrepancies), and the referral's own records_received/imaging_received/auth_*/appointment_scheduled flags.
• Dialysis transfer review (DIAL-*): per transfer compute packet completeness, missing_required_documents, stale_documents, requested_start feasibility vs chair capacity, final_intake_decision, next_contact_*, plus cohort summary. Driven by transfer_requests, documents (packet), facility_capacity, patients.
• Chronic-care enrollment panel (DMHTN-*): per candidate compute eligible, enrollment_status, reason_codes, follow_up_cadence, missing_chart_artifacts, outreach_channel, initial_monitoring_package, plus summary counts. Driven by program_candidates, chart_artifacts, clinical_history, patients. Program identity (e.g. DMHTN = diabetes+hypertension) determines the required target condition and diagnosis.
• Referral-to-chart activation (PULM-* etc.): per referral compute readiness_status+blocker_codes; plus clinical_code_discrepancy_referrals, blocker_sets, duplicate_handling (groups vs cleared), ready_referral_chart_needs (chart_action + artifacts_to_create), correspondence_queue, priority_order. Driven by referrals, icd_codes, chart_artifacts, patients.

Important pitfalls observed

• Distractor records. A patient may carry an unrelated referral/transfer/program membership that does not belong to the target batch/roster/program. Filter strictly by the target identifier; ignore notes like "distractor referral".
• missing_records/missing_imaging come from referral flags, not from the documents table (referral readiness audits use referrals.records_received / imaging_received). The documents table is for transfer packets and chart artifacts.
• ICD discrepancy = service-family mismatch. Compare icd_codes.service_family to the referral's service_line; a code from another family (e.g. a cardiology code in a pulmonary batch) is a clinical code discrepancy. narrative_mismatch/laterality_mismatch additionally compare the code description/laterality to the referral narrative.
• Duplicate vs cleared. True duplicates share (patient_id, icd10_code) within the batch. Referrals merely flagged "possible duplicate" but with distinct keys are cleared, not duplicates.
• Freshness limits are clinical constants, not in the portal. Dialysis packet freshness windows must be applied per doc type (see reference/decision_rules.md); they are inferred standards, applied consistently.
• Required values can be hidden in the environment. requested_service_date and service_line for a roster live in intake_rosters, not the prompt — read them from the portal.

When a rule in reference/decision_rules.md is marked (inferred rubric), it was distilled from the training data rather than read verbatim from the portal; apply it consistently so output is reproducible.