Add: Document extraction hallucination — agent reports correct-looking JSON with wrong values

[digitalpromptmarket-beep]

Failure Mode: Silent data extraction errors (Tool Hallucination)

Category: Tool-calling hallucination / Output corruption

Description:
When AI agents extract structured data from documents (invoices, contracts, financial reports), the output often looks perfectly valid — well-formed JSON, reasonable values, correct field names — but contains subtle errors:

• Line item amounts that don't sum to the stated total
• Dates that are plausible but don't match the source document
• Vendor names that are close but not exact (OCR artifacts or LLM "corrections")
• Tax calculations that use the wrong rate

These errors are particularly dangerous because they pass schema validation and human spot-checks. They only surface when someone manually re-checks the math or compares field-by-field against the original.

Real-world impact:

• Financial systems that auto-process extracted invoices can pay wrong amounts
• Compliance workflows that trust extracted dates can miss deadlines
• Audit trails become unreliable when the extracted record doesn't match the source

Mitigation:
Add a mechanical verification step between extraction and downstream action: arithmetic validation, cross-field consistency checks, and source-document comparison. This should be automated (not human review) to scale with agent throughput.

References:

• Vectara Hallucination Leaderboard — shows variation in model accuracy for document summarization
• Common pattern in invoice processing pipelines using Textract, Azure Document Intelligence, or LLM-based extraction

[ofermend]

Thanks for the PR @digitalpromptmarket-beep
I don't think this type of issue should be categorized as a tool hallucination - that category is intended for when a tool is called and responds with a bad output during the query flow.
I think the extraction process is done either by the agent LLM itself or offline / batch during ingest. In the case you're suggesting is the extraction really a tool call, or is it just a separate type of thing?