Build an application that audits business reports, presentations, and analytical summaries for internal inconsistencies before they are shared.
Users can provide report text, tables, supporting CSV files, and a collection of source documents. The tool should extract checkable claims and compare them with the supplied evidence.
Core capabilities:
Claim Extraction: Identify numerical claims, dates, percentages, named entities, comparisons, and statements about trends.
Arithmetic Verification: Recalculate totals, percentage changes, averages, ratios, and subtotals whenever the necessary source values exist.
Cross-Section Comparison: Detect when an executive summary contradicts a chart, table, appendix, or another section.
Source Traceability: Link each extracted claim to its supporting passage, table cell, or source record where possible.
Unsupported Claim Detection: Flag claims for which no supplied evidence was found without automatically labeling them false.
Version Comparison: Identify changed numbers or claims between report versions and show whether the supporting evidence also changed.
Review Queue: Let a human reviewer confirm, reject, or revise each finding.
Display findings with severity, evidence, calculation details, and a suggested correction. Export an audit report in CSV or JSON.
Do not fabricate citations, source passages, or verification results. Distinguish mathematical errors, direct contradictions, unsupported statements, and ambiguous wording. Treat AI-generated findings as review suggestions, and never mark a claim verified solely because another AI-generated statement agrees with it.
0 Comments