File formats cut LLM workflow accuracy 53.63%

Changing only a document’s file format cut LLM-workflow accuracy by as much as 53.63% in a 48,000-execution study spanning four workflows, four tasks, and four formats. Semantically equivalent inputs produced decision drift in more than 41% of cases, while lightweight mitigations recovered up to 44.21% of that drift. For teams building the verification loops that keep persistent agents honest, CSV, JSON, spreadsheet, and document variants belong in the regression matrix rather than being treated as neutral wrappers.