Extraction Eval Suite
Runs the invoice extraction model against 20 fixed test invoices — clean, relabeled, unusual date formats, missing fields, noisy scans, a skewed scan, and 2 native PDFs — and scores the result field by field against known-correct values.
Calls the AI model once per fixture — this uses real AI credits and takes about a minute.
Live Correction Rate
From real usage of the Upload Invoice page — how often a human changed an AI-extracted field before saving. This is a production signal, separate from the synthetic fixtures above.