Auditable LLM Extraction of Financial Data from SEC 10-K Filings with Inline XBRL Validation

Item 8 of an SEC 10-K filing contains the audited financial statements and notes that form the financial core of a U.S. public company's annual report. Turning this material into reliable structured data is difficult because each reported number must be tied to the financial line it describes, the fiscal period it covers, the unit and scale stated in the filing, and conventions such as negative values in parentheses or per-share amounts. We present a decomposed, schema-constrained pipeline in which a language model reads each financial statement and note directly from the filing, while deterministic code turns the result into auditable records with normalized values, periods, units, and provenance. The system then compares those records with the issuer's inline XBRL facts. When that comparison reveals a missing or inconsistent value, it directs the model back to the relevant source passage for a limited re-reading and extraction attempt. Across paired baseline-versus-repair runs for eight open-weight models (8B to 70B, three families) on eight Item 8 sections from different 10-K filings, this process increases value coverage of issuer-tagged XBRL facts from 47.7\% to 53.0\% (${\sim}11\%$ relative) and improves six of the eight models. These results show why issuer-provided machine-readable evidence can complement language-model reading: it exposes omissions that would otherwise be invisible and helps produce financial records that can be checked against the source filing.