Northstar Payer Ops

Northstar Payer-Operations Deliverable Builder

What this skill is for

Northstar payer-operations tasks all share the same shape:

• A task_context.json naming a target business ID, a requester role, a reporting date, and environment access.
• An answer_template.json that fixes the exact output JSON (required keys, enums, ordering rules, precision, null handling).
• A shared read-only environment you must query for the real facts.
• The instruction "return JSON only" conforming to the template.

Your job is to pull the real records from the environment, apply the operating rule for the task's archetype, and emit one JSON object that exactly matches the template. Never invent facts, IDs, amounts, or dates — every value must trace to a record you retrieved.

Five recurring archetypes exist, each keyed to one source_precedence value; a sixth precedence value exists for appeal-deadline-driven mixed tasks. See the mapping table below and reference/playbook.md for the per-archetype procedure.

Step 1 — Read access config (never hardcode)

Read environment_access.md (and the task's task_context.json) at run time for:

• Base URL — the GDPEVO_ENV_BASE_URL value, substituted for <TASK_ENV_BASE_URL>.
• Bearer token — sent as Authorization: Bearer <token>.
• Allowed endpoints — only use endpoints the file lists.

These are environment-scoped and may change between task groups, so always resolve them from the file rather than assuming a prior value.

Step 2 — Understand the environment (SQL is the workhorse)

• POST /sql/query, JSON body {"sql": "<SELECT ...>"}, with the bearer header. Only SELECT, WITH, and PRAGMA table_info are allowed (writes are rejected with {"error":"invalid_sql"}). Missing/aborted auth → {"error":"unauthorized"}; a bad table/column → {"error":"sql_error","message":...}. Success → {"columns":[...], "rows":[{...}], "row_count":N, "limited":bool, "max_rows":500}. Results cap at 500 rows — filter by the target ID so you never rely on truncated data.
• Discover the schema with GET /api/tables (returns every table with column names, types, PK, not-null). A cached copy is in reference/schema.md, but re-fetch to confirm — table shapes can differ per environment.
• Business REST endpoints (GET /api/cases, /api/cases/{id}, /api/policies, /api/policies/{id}, /api/documents/{id}, /api/rate-schedules, /api/appeals, /portal) return the same data pre-joined and are convenient for spot checks; SQL is better for precise joins and filters.

Guardrails: Do not read environment source files, generated data files, SQLite files, manifests, or setup scripts directly — several prompts forbid it. Get every fact through the SQL/REST endpoints only. Access is read-only; do not attempt writes.

Step 3 — Identify the archetype and pull its records

Resolve the target ID in cases (or claims/appeals when the ID is a claim/appeal), read request_type / service_domain / current_stage, then follow the matching row:

Full per-archetype procedures, formulas, and field mappings are in reference/playbook.md. Read the one that applies before computing.

Step 4 — Apply the operating rules (cross-archetype invariants)

• Current-over-stale. documents.is_current = 1 are the evidence documents you rely on; is_current = 0 (e.g. stale_export, legacy source systems) are excluded and become exception records — never controlling evidence.
• Effective-dated benchmarks/policies. When repricing or applying a rate, select the record matching payer + plan_type + service_domain + cpt_code + modifier whose [effective_start, effective_end] window contains the service date. Reject any schedule whose window ended before the service date (a stale/legacy schedule) and any "distractor" schedule that does not match the claim's CPT/plan. allowed_amount is per unit — multiply by line units.
• Criteria come from the data, filtered to the template's keys. case_criteria.result already holds met/not_met/partial/unclear. Emit only the criterion IDs the answer template's required_keys list; ignore informational criteria (approval_required = 0) unless the template asks for them. When a required criterion is not met, policy_criteria.result_if_missing (pend/deny/uphold) drives the disposition/route.
• Documented vs insufficient evidence. drug_trials.documented = 1 (with a real fill record / letter) counts as a documented failure; documented = 0 or "referenced without fill record" is undocumented/insufficient.
• Assistance after payer path. Manufacturer-assistance readiness (assistance_screen) is secondary to the payer appeal; assistance gaps rank after appeal-evidence gaps.
• Margin math. total_cost = variable_cost + fixed_cost_allocated; margin = net_revenue - total_cost; revenue_to_cost_ratio = net_revenue / total_cost; below_threshold = ratio < threshold (threshold from task_context, e.g. 1.2); gap_to_120pct = threshold * total_cost - net_revenue for the top below-threshold row.
• Deadlines. Use the plan's stated window (e.g. 180-day internal appeal) counted in calendar days from the relevant determination/denial date; use null when no deadline applies.

Step 5 — Build basis_audit (required in every answer)

Every template requires a basis_audit object with these four keys:

• source_precedence — the enum value for this archetype (table above).
• controlling_record_ids — the environment record IDs that directly decide the result (e.g. the current documents, criteria rows, chosen benchmark, auth record, P2P event, or the below-threshold margin row), in operational evidence order.
• exception_record_ids — the gap/exclusion records: unmet-criteria or route gaps first, then stale/excluded/decoy records (stale documents, rejected legacy benchmarks, undocumented trials, charge-sensitive-only rows).
• precedence_record_order — the controlling and exception records merged into one list, highest priority first under the source_precedence rule.

Use the real record IDs from the environment (document_id, criterion_id, benchmark_id, auth_id, line_id, month_id, appeal_id, p2p_id, etc.), not invented ones.

Step 6 — Enforce the output contract, then self-check

• JSON only. No markdown, comments, or prose outside the object. Include every required_top_level_field. Honor additional_fields_allowed / additional_properties.
• Enums. Every enum value must be an exact allowed choice — no synonyms or casing drift.
• Ordering. Follow each field's ordering rule precisely (ascending document_id, ascending CPT, alphabetical medication/segment, claim-line order, the queue_row_ids order from task_context, choices-order for factor lists, etc.).
• Precision & types. Currency → 2 decimals (dollars); ratios → the stated precision (e.g. 4 decimals); units/counts → integers; dates → YYYY-MM-DD (periods YYYY-MM).
• Null vs empty. Absent modifier → null, not "". Empty list only when genuinely none. Absent deadline → null.
• Self-check before returning: all required keys present; enums valid; every ID and amount traceable to a retrieved record; stale/decoy records excluded and recorded as exceptions; ordering and precision applied; basis_audit complete.

See reference/pitfalls.md for the planted decoys to watch for.