Some extract failures are format mismatches; others lack any source for a required field. How should handling differ?
Select an answer to reveal the explanation.
Short Explanation
Fix the smudged date with a retry; if the deed was never in the envelope, call a clerk. Format mismatches retry; missing sources escalate.
Full Explanation
Extract failures fall into different remediation classes. Format mismatches—bad date spellings, enum casing, lightly malformed amounts—are often retryable with the document, failed JSON, and validator errors. Missing-source cases—required fields that never appear in the packet—are not fixed by more prompting and should escalate to humans or document retrieval. Architects who retry format mismatches and escalate missing-source cases protect both compute budget and data integrity for municipal intake. Escalating every format typo to humans and never retrying floods clerks with noise the model can usually correct on a second pass. Retrying missing-source cases forever until a field appears invites fabrication and never supplies the absent deed or attachment. Treating format and missing-source failures as identical always either over-escalates cheap fixes or under-escalates impossible ones. Exam caveat: classify failures by recoverability before choosing retry versus escalate; schema error codes alone may not encode that distinction unless you design them to. Operational check: tag validator outcomes as format versus missing_source, cap retries on format class, auto-escalate missing_source with a clerk checklist of which source document is absent, and measure false escalations and fabricated fills separately.