Northstar Payer Ops Determination

Northstar Payer-Ops Structured Determination

This skill turns a Northstar Health Plan payer-operations work item into one correct JSON object that conforms to the supplied answer_template.json. It is domain-generic: the same workflow covers prior-authorization determinations, pharmacy appeals, claim repricing, peer-to-peer summaries, and finance margin queues. Apply the process below; do not memorize or reuse values from any specific prior case.

What you receive

Each task stages three inputs (paths are relative to the task root):

• input/prompt.txt — the business request in plain language.
• input/payloads/task_context.json — structured context: the target business ID, requester role, reporting date/period, work type, environment pointer (<TASK_ENV_BASE_URL>), and any task-specific parameters (thresholds, queue row IDs, source tables, appeal windows, etc.).
• input/payloads/answer_template.json — the binding output contract: required top-level keys, per-field types/enums/ordering/precision, the basis_audit sub-contract, and the additional-fields rule.

The template is the contract. Every required key, enum choice, list ordering, numeric precision, date format, and null rule in it is binding.

Environment access — network only

Resolve access from /work/environment_access.md (the sole sanctioned source). Do not invent endpoints, base URLs, or credentials.

• The <TASK_ENV_BASE_URL> placeholder in task_context.json maps to the base URL recorded in environment_access.md.
• SQL access: POST /sql/query to the base URL with header Authorization: Bearer <token from environment_access.md>.
• Open business endpoints (no auth), as listed in environment_access.md: GET /portal, GET /api/tables, GET /api/cases, GET /api/cases/{case_id}, GET /api/policies, GET /api/policies/{policy_id}, GET /api/documents/{document_id}, GET /api/rate-schedules, GET /api/appeals.
• Prefer REST endpoints to traverse the obvious entities; use POST /sql/query for joins or facts the REST surface does not expose (drug-trial records, assistance-screen facts, margin rows, line-level detail, etc.).

Hard rule: reach the environment ONLY over its HTTP/SQL interface. Do not open, cat, sqlite3, grep, or otherwise inspect environment source files, generated data files, SQLite/database files, manifests, setup scripts, or construction files. The environment is a black box. Derive every value from live network responses, never from local files or memory.

Workflow

1. Load inputs. Read prompt.txt, task_context.json, and answer_template.json. Identify the target business ID, reporting date or period, requester role, work type, and any task-context parameters (thresholds, queue row IDs, source tables, appeal windows).
2. Internalize the template. Note required top-level keys, each field's type/enum/ordering/precision, the basis_audit contract, and whether extra fields are allowed. You will emit exactly this shape.
3. Resolve environment access from environment_access.md.
4. Discover the schema. GET /api/tables first, so you know what entities and tables exist before querying.
5. Fetch and traverse. Pull the target record, then related entities: member/plan context, requested service/procedure lines, applicable policy and criteria, clinical/evidence documents, the authorization record, and any work-type-specific facts (drug trials, assistance screen, rate schedules, P2P event, margin rows). Use REST where it fits; fall back to SQL for the rest.
6. Reconcile via source-precedence. Apply the domain's precedence rule to resolve conflicts (current vs stale; payer appeal vs manufacturer assistance; effective benchmark vs stale schedule; new P2P info vs prior record; margin threshold vs charge sensitivity). Decide which records control the result and which are exceptions/gaps/excluded.
7. Evaluate criteria. Map each required criterion ID to its review result enum from the reconciled evidence.
8. Classify documents/records. Evidence relied on vs excluded (stale/superseded/inapplicable).
9. Compute the result. Determination/disposition, route/owner, next action, dates, and numeric outputs per work type, using task-context parameters and the template's precision rules.
10. Assemble basis_audit. Per references/basis_audit_contract.md.
11. Emit. Exactly one JSON object conforming to the template.

The basis_audit contract (summary)

Every answer includes a basis_audit object with four required keys:

• source_precedence — one of six fixed precedence rules (see references/basis_audit_contract.md). Pick the rule that governs this work type.
• controlling_record_ids — environment record IDs that directly control the result, in operational evidence order (complete list).
• exception_record_ids — gap/exception records that explain exclusions, denials, missing information, or route priority; criteria/route gaps before stale/excluded records (complete list).
• precedence_record_order — the controlling and exception records in source-precedence order, highest priority first. This is a curated, de-duplicated trail of the records that drive the precedence decision, not a plain union of the two lists above: omit records that are mere operands or supporting evidence, and list a record once even if it appears in both roles.

Full semantics, the six rules, and worked ordering logic live in references/basis_audit_contract.md.

Output contract invariants

• Emit a single JSON object. No markdown, no prose, no comments outside the JSON.
• Enum values must exactly match the template's choices.
• Order every list per its template rule (ascending document_id, ascending CPT, claim-line order, alphabetical, the task-context queue_row_ids order, the template's choices order, operational packet order, etc.).
• Currency in USD rounded to two decimals; ratios to the precision stated (typically 4 decimals); service units as integers.
• Dates as YYYY-MM-DD ISO calendar dates; periods as YYYY-MM. Use null (not "") for absent modifiers and for deadlines that do not apply.
• Do not add fields beyond the template unless it explicitly permits additional properties.

Pitfalls

• Reaching for local env/db files instead of the network — never.
• Copying values from memory or any prior case — derive every value live.
• Treating precedence_record_order as a plain union — it is curated and reordered by the precedence rule (see references).
• Conflating evidence documents with controlling record IDs — they overlap but serve different fields; populate each per its own rule.
• Forgetting to explicitly exclude a stale/superseded source (it belongs in excluded/exception, not evidence/controlling).
• Ignoring a task-context parameter (threshold, queue row IDs, appeal window) — these drive numeric and date results.

References

• references/basis_audit_contract.md — the four basis_audit keys, the six source-precedence rules with semantics and work-type guidance, and the ordering / curation / de-duplication logic.
• references/environment_workflow.md — the detailed data-gathering workflow, the endpoint-to-entity traversal map, per-work-type notes, and formatting invariants.