How it decides
Two separate reads of every invoice, one by AI and one by plain code. Plain rules do the maths and the checks. AI never decides, and nothing clears on one read.
Measured on 88 generated invoices (48 held-out)
Model: gemini-3.5-flash-lite · Evaluated 2026-11-06
AI read 512 of 517 fields (99%) on 48 held-out generated invoices. Checks caught 24 of 24 planted errors across all 88 invoices, with 0 false clears.
Real invoices will score lower than generated ones. Zero false clears is the hard safety rule.
What this isn't
It's a demo, not a client system. The test invoices are generated, so real invoices will score differently. A real setup would be tested on your own files first.