Schema contract tests, not smarter AI models, fix silent extraction pipeline failures

A software engineer inherited an LLM-based data pipeline that extracted fields from vendor emails and PDF invoices and inserted them into a Postgres database. A silent failure occurred at 4 a.m. when the model returned a price as a formatted string instead of a numeric value, causing inserts and retries to fail unnoticed for two days. The root cause was the absence of output validation — the model's responses were never checked against a structured contract before reaching the database. The engineer built a three-gate testing harness using Pydantic to measure what percentage of model outputs correctly parse, validate, and coerce at scale. The key insight is that extraction pipelines fail on formatting rather than logic, and require continuous schema contract testing rather than simply switching to a more capable model.