You are building a structured data extraction system using Claude.

Option A uses the limited review capacity as a risk-ranking mechanism rather than spending it uniformly. JSON Schema validation can prove that an output has the expected types and structure, but it cannot detect a semantically misplaced value such as a duration stored as an ingredient quantity. Field-level confidence can identify the exact parts of a record that warrant inspection, provided those scores are calibrated against labeled examples rather than trusted at face value. The labeled validation set must mirror the production distribution and include edge cases; Anthropic’s evaluation guidance explicitly recommends task-specific evaluations, representative distributions, measurable criteria, and automated grading where possible. Calibration then converts raw confidence into an empirically tested estimate of error risk, allowing reviewers to inspect the lowest-confidence records or fields first. Option C, random sampling, is useful for unbiased quality measurement and discovering new failure modes, but it does not maximize error detection within a fixed 20 percent review budget. Options B and D target narrow proxies—layout anomalies and missing required values—while the observed errors pass structural validation and may occur in ordinary-looking documents. The operational design should combine calibrated risk-based review with a smaller random audit sample to detect drift and overconfidence.

================