Fixture set: extraction. Generated by reconcile eval-extraction. This file is committed and regenerated when the extractor or the labeled fixtures change. The fixtures are deterministic synthetic intake forms with hand-written ground-truth labels; there is no real personal data in them.
Field-level precision and recall of the offline PDF extractor. A predicted field is correct when its field name and normalized value match a labeled field in the same document, using the same normalizers the matching pipeline applies, so formatting differences do not count as errors. Unmatched predictions are false positives; unmatched labels are false negatives. This is a REVIEW metric in the metrics ledger, read by a person, and a test also holds the committed fixture at or above the ledger targets so the numbers cannot drift silently.
| Metric | Value | 95% CI |
|---|---|---|
| Documents | 4 | |
| Labeled fields (ground truth) | 17 | |
| Predicted fields | 16 | |
| Precision | 100.0% (16/16) | [80.6%, 100.0%] |
| Recall | 94.1% (16/17) | [73.0%, 99.0%] |
| Field | TP | FP | FN | Precision | Recall |
|---|---|---|---|---|---|
| dob | 2 | 0 | 1 | 100.0% | 66.7% |
| 3 | 0 | 0 | 100.0% | 100.0% | |
| first_name | 4 | 0 | 0 | 100.0% | 100.0% |
| last_name | 4 | 0 | 0 | 100.0% | 100.0% |
| phone | 3 | 0 | 0 | 100.0% | 100.0% |
Ledger targets: precision at least 95.0%, recall at least 90.0%. Observed: precision 100.0%, recall 94.1%. MET.
False negatives (1 here) are shown, not hidden: the fixture set deliberately includes at least one field the deterministic extractor is known not to parse (see the fixture README), so the measurement demonstrably catches a real miss rather than scoring a set the extractor is guaranteed to ace.