Field-Level Accuracy
The metric that matches the stakes — was *this field* right, not how many characters were.
Field-level accuracy is the fraction of extracted fields whose final, normalized values are exactly correct — the invoice total right or wrong, the date of birth right or wrong — measured per field type across a ground-truth test set. It is the metric that speaks the business's language: downstream systems and decisions consume field values whole, and a value that is 90% correct at the character level (one digit wrong in an account number) is 100% wrong at the field level, which is the level where payments route and claims pay.
Its diagnostic power comes from granularity. Reported per field type, it exposes the heterogeneity that aggregates bury: the printed header fields at 99.5%, the handwritten totals at 91%, the multi-line address at 96% — each its own automation decision, threshold, and improvement target. Reported per document type and channel, it locates the problems: this vendor's format, that branch's scanner. Crossed with confidence, it yields the calibration curves that make routing rational. The definitional work is the same as all extraction benchmarking — normalization rules stated (else "1,500.00" versus "1500" corrupts the measurement), empty-field semantics defined (correctly extracting "nothing present" counts, and matters for optional fields), and multi-value fields given explicit scoring.
Field-level accuracy also frames the economics of automation honestly. A document is straight-through processable only if every consequential field clears its bar, so document-level automation rates compound from field-level rates — twenty fields at 99% each yield roughly 82% of documents with all twenty right, a compounding that surprises stakeholders calibrated on single headline numbers. That arithmetic is why per-field routing outperforms per-document routing, why the weakest critical field dominates the system's effective performance, and why improvement effort ranked by field-level impact — accuracy gain times field criticality times volume — is how mature teams choose what to fix next.
Not 'how good is the model' but 'how good is it on our documents, per field, against ground truth.'
The OCR world's batting average — how many characters the model got wrong, per hundred it should have read.
One number balancing two sins — missing what's there and inventing what isn't.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo