OCR & extraction accuracy benchmark #2

Description

We need to know whether OCR and field/clause extraction are actually accurate, not just that they run without erroring.

Proposed Solution

Build a labeled dataset of sample documents with ground-truth text/fields, and a benchmark that scores OCR and extraction output against it.

Acceptance Criteria

• Labeled dataset of sample documents with ground truth exists
• Benchmark scores OCR text accuracy against ground truth
• Benchmark scores field/clause extraction accuracy against ground truth
• Results are reproducible and comparable across runs