Parsing silently destroys document structure, so AI answers over case files come back wrong and cannot be traced to the source

Summary

Teams building document Q&A over real files repeatedly trace wrong answers back to extraction, not the model: headings are flattened, table rows are chopped in half, two-column pages are read straight across both columns, and continuations across pages are lost. Benchmarks show this is systemic — one widely used converter turned columns into fake tables so that only ~4% of sentences survived intact. For litigation work this is catastrophic, because the same extraction layer feeds chronologies, contradiction analysis, and AI answers: if a deposition heading or a table row is mangled, the system asserts the wrong fact with full confidence, and the page-pinpoint needed for citation and highlighting is gone. Nothing in the market guarantees structure-preserving, page-anchored extraction tuned for legal documents, and general converters silently degrade rather than warn. The unsolved need is parsing that keeps headings, table integrity, reading order, and bounding boxes so every answer can be pointed back to the exact source passage.