Engineering review using AI coding tools
This review was performed in the same implementation session using AI coding tools. It is not an independent human document audit.
Reviewed generated invoice/statement sources, all four rendered pages, the baseline and repaired extractors, Decimal controls, page sequence, declared row counts, provenance, output typing, formula-like text, review queue and failure states. All three workbook sheets were previewed from saved values; this is not native Excel rendering.
Regression assertions compare known PDF values, totals, row counts and provenance, round-trip the saved XLSX, exercise malformed monetary values and failed controls, and reject missing/duplicate/reordered pages. The intentionally ambiguous debit is kept blank and prevents complete debit/net-movement reconciliation. Matching credits or closing balance cannot change that review outcome. Source-derived QA is a control mechanism, not proof of perfect extraction.
Residual limits: two known ruled text-PDF layouts only; no OCR; no universal confidence calibration; no automatic correction from totals; no native Excel/Sheets test; new output folders are recommended to preserve prior snapshots if processing fails. See output/QA_REPORT.md, evidence/tests.json and HANDOFF.md.