Cedar Ridge Intake Coordination

Cedar Ridge Intake Coordination — Task Solving Skill

What these tasks are

Each task targets one batch (a roster, referral batch, transfer batch, or program) in the Cedar Ridge Intake Coordination Portal and asks you to produce one JSON object that follows a provided answer_template.json. The portal is a shared, read-only store of intake, referral, transfer, chart, and program data. The portal base URL is given in the task prompt as <TASK_ENV_BASE_URL>; the per-task environment_access.md documents the available endpoints (treat that doc as the source of truth for endpoint mechanics — do not hardcode URLs).

The work is deterministic reconciliation + rule application, not free-form narration. Every output field is a controlled enum, a code from a fixed set, an ID, a count, or a date. There is no prose.

Universal method (apply to every task)

1. Read the prompt and answer_template.json together. The template is the contract: it lists required top-level keys, per-item required keys, allowed enum values, list orderings, and count-key shapes. Print the template's allowed-values lists and keep them next to you while deriving. Most errors come from inventing a value not in the allowed set, omitting a required key, or wrong list ordering.
2. Identify the batch identifier and the entity set the batch spans (patient IDs, referral IDs, transfer IDs, or program candidate IDs). Pull the full list for that batch first, then fetch each entity's detail.
3. Gather all related records per entity. A single entity (e.g. a patient) typically fans out to coverage, PBM, pharmacies, lifestyle, clinical_history, chart_artifacts, and its own referrals/transfers/rosters. Fetch the detail record for every member; do not rely on list-level summaries. Cross-check with the read-only SQL endpoint (it accepts a SQL string) when you need to reconcile across tables (duplicates, shared insurance, capacity by date).
4. Derive each output field by a stated rule (see references/derivation_rules.md). Write the rule down before assigning values so the logic is consistent across all items — ad-hoc per-item judgments drift and produce inconsistent cohort counts.
5. Assemble counts/summaries from the per-item results, not by re-deriving. Every summary count is a rollup of the per-item fields; if a per-item field flips, recompute the affected counts.
6. Validate against the template before submitting: every required key present, every value in the allowed set, every list in the specified order (ascending ID unless noted), count objects have all required keys (including zero counts), integers are integers. Output JSON only — no surrounding prose.

Data model (what to pull and from where)

See references/data_model.md for the full entity/field map. Key relationships:

• patients are the hub. A patient detail bundles: patient demographics, coverage (insurance), pbm (prescription benefit), pharmacies (preferred + network), lifestyle, clinical_history, chart_artifacts, and the patient's own referrals, transfers, rosters, program_candidates.
• referrals carry icd10_code, diagnosis_description, referral_reason, records_received, imaging_received, auth_required/auth_status, appointment_scheduled, insurance_id, urgency, service_line, batch_id.
• icd_codes map a code to chapter, service_family, laterality, description. Use service_family to detect clinical-code discrepancies (code's service family vs. referral's service line) and laterality/description for narrative/laterality checks.
• transfer_requests carry requested_start_date, modality, days_requested, chair_window, transportation. facility_capacity is keyed by (location_id, date, modality) with open_chairs; the requested start date may have no capacity row (treat as 0 / unavailable).
• documents are keyed by transfer_id (and sometimes referral_id); each has doc_type, status (final/draft), finalized (0/1), received_date. Packet completeness and freshness are derived from these.
• program_candidates carry consent_status, preferred_outreach, adherence_score, target_condition. Pair with the patient's clinical_history (chronic conditions) and chart_artifacts to determine eligibility.

Core derivation rules (transferable across batches)

These are the rules that consistently drive correct field values. Full detail and edge cases in references/derivation_rules.md.

Status fields are conjunctive — a status is "valid" only when every sub-condition holds.

• Insurance is valid only if coverage is active, in-date on the requested service date, and the requested service line appears in coverage.service_lines. A code on a referral whose ICD service_family differs from the referral service_line is a clinical-code discrepancy.
• Prescription (PBM) is valid only if active=1, status=approved, formulary_status=covered, and the PBM policy_number matches the coverage policy_number. Otherwise the specific failure maps to pbm_invalid / pbm_policy_mismatch / pbm_missing.
• Pharmacy status follows the preference_rank=1 pharmacy's network_status.

Missing vs. stale documents are different.

• A required document is missing if its doc_type is absent from the packet or it is present but not finalized (status != final / finalized != 1). Drafts count as missing for completeness.
• A document is stale only among the freshness-limited types, comparing received_date to the relevant service/start date against the type's freshness limit. Only finalized documents are staleness-checked (a draft is already "missing").
• Some "required" documents are satisfied by a non-document field (e.g. transportation is the transfer's transportation arrangement, not a doc_type). A null arrangement = missing.

Eligibility ≠ enrollment action.

• For program panels, eligible reflects clinical criteria (correct target condition + the required diagnoses present). Consent, chart, and artifact problems are enrollment barriers that drive enrollment_status (enroll / hold / reject), not the eligible boolean. A patient can be clinically eligible yet rejected (e.g. consent declined) or held (consent missing).

Capacity is summed across locations for the exact requested date.

• open_chairs_total = sum of open_chairs across all locations for that date+modality. If no capacity row exists for the requested start date, it is 0 / unavailable.

Cohort summaries roll up per-item decisions. Re-derive every count object from the final per-item list; include all required count keys even when zero.

Output discipline (non-negotiable)

• Return a single JSON object — no prose, no markdown fences, no commentary outside the JSON.
• Use only values from the template's allowed sets; never invent codes or enum values.
• Preserve exact key names and include every required key (templates mark required keys/lists).
• Order lists as the template specifies: usually ascending by ID; some arrays are explicitly "unordered sets" (still emit them sorted for determinism). Count objects must contain all required count keys, including those whose value is 0.
• IDs are uppercase exactly as the portal returns them (P001, REF0010, TR0001, etc.).
• Dates are YYYY-MM-DD. Integers are integers, never strings.

Task archetypes

The batch identifier signals the archetype; each has its own field set in references/derivation_rules.md:

• New-patient access verification (roster, e.g. NPI-JUN-01): per patient, insurance / prescription / pharmacy status, lifestyle & overall risk, registration status, blocked reason codes, plus a cohort summary.
• Referral readiness audit (e.g. ORTHO-JUN-01): per-referral readiness, issue codes, ICD discrepancies, duplicate groups, shared-insurance anomalies, blocker sets, ready-to-schedule list, action plan, and multi-axis summary counts.
• Transfer packet review (e.g. DIAL-WINTER-01): per-transfer packet completeness, missing/stale docs, requested-start capacity feasibility, intake decision, next-contact owner/route, cohort summary.
• Program enrollment panel (e.g. DMHTN-2026A): per-candidate eligibility, enrollment disposition, reason codes, follow-up cadence, missing chart artifacts, outreach channel, initial monitoring package, summary counts.
• Referral-to-chart activation (e.g. PULM-JUN-02): readiness by referral, clinical-code discrepancy referrals, blocker sets, duplicate handling, ready-referral chart needs, correspondence queue, priority order.

For each, build the per-item records first, then the cross-item structures (duplicates, shared-insurance, blocker sets, summaries), then validate the whole object against the template.