The customer problem
A claims team processes documents: eligibility, coverage rules, adjudication, appeal. The backlog is real, but the constraint is defensibility. Every automated decision must trace to the document span and the policy clause that produced it, and anything ambiguous must reach a human with the evidence assembled.
Architecture shape
Extraction produces claims with document-anchored evidence. Policy-as-code gates decide what auto-adjudicates and what routes to review. Reviews are durable pauses with evidence-rich payloads. An append-only event log records every step, model decision and human action, keyed for replay.
The decisions that matter
- The model extracts; policy decides. Regulated decisions cannot ride on sampled text.
- Evidence is the schema, not an afterthought: a decision without a document anchor is not representable in the system.
- The audit trail is a product requirement, so the event log is designed first, not logged incidentally.
Where Course 001 uses this case
Evidence-backed state (Module 02), human control (Module 07), and the production gauntlet’s auditability checks (Module 15) run their scenarios here.