Audit Report Checker

Problem

Audit reports often contain errors spread across reconciliations, totals, note-to-statement consistency, and formatting: balance sheet identities not holding, note detail totals mismatching statement values, missing row or column sums, typos, page-number, company-name, or unit issues. Manual review is slow, while keyword-based tools can produce systematic false positives. These checks often involve consolidated versus parent-company scopes, duplicated statement titles, cross-table note details, and irregular column layouts, which rules alone struggle to cover.

How it works

• Semantic mapping: Claude identifies consolidated and parent-company statements, note sections, and column meanings, then generates manifest.json and note_map.json instead of relying on hardcoded formats.
• Deterministic arithmetic: all additions and subtractions are checked with scripts/calculator.py, supporting thousands separators, parenthesized negatives, ten-thousand units, and tolerances.
• Layered checks: code-side checks cover L1 statement identities, --scan row/column totals, page numbers, and company names; semantic review covers note-to-statement reconciliation, special table structures, and false positives.
• Traceable output: results include page, context, and source_file, exported as Excel and Markdown for easy source verification.

Boundaries

• Best for Word/PDF audit reports; scanned PDFs and image-based statements may require cloud OCR, with data-upload considerations.
• Read-only review, no source modifications; deep checks consume more tokens and final conclusions should still be confirmed by an auditor.
• Note-only documents skip statement reconciliations and focus on note-internal checks and text formatting.