Proof & benchmarks
The numbers, and where each one comes from.
Two corpora — one we did not build. Every figure below is measured on held-out data: documents set aside and never used to build the extractor. The measurement rules were fixed before the first run, and nothing here is rounded up.
normalisation spec 4250bb3146941889
On degraded photos, the free alternative fabricates a value on 35.9% of the fields that do not exist — each presented exactly as confidently as a correct one. Ledger’s is 0.0%.
35.9%
free OCR + rules
0.0%
Ledger
hard-photos
Degraded photographs — rotated, blurred, contrast-shifted, 70–110 dpi. 64 held-out documents, calibrated on 30.
hard-photos · n=64 held-out · cal=30 · spec 4250bb31
| Metric | Free OCR + rules | Ledger |
|---|---|---|
| Field exact match HEAD_TO_HEAD.md | 49.1% | 98.7% |
| False-invention rate HEAD_TO_HEAD.md | 35.9% | 0.0% |
| Field coverage @ risk ≤ 0.05 HEAD_TO_HEAD.md · certified | impossible | 99.8% |
| Document straight-through HEAD_TO_HEAD.md · certified | impossible | 95.3% |
| Observed selective risk (held-out) HEAD_TO_HEAD.md · held-out | — | 0.0112 |
| Silent failures (accepted-but-wrong docs) HEAD_TO_HEAD.md · held-out | — | 0 |
| Cost / 1,000 documents HEAD_TO_HEAD.md · measured | $0.00 | $22.05 |
CORD-v2
A public third-party corpus we did not build — 87 held-out Indonesian receipts, calibrated on 71 disjoint documents.
CORD-v2 · n=87 held-out · cal=71 · spec 4250bb31
| Metric | Free OCR + rules | Ledger |
|---|---|---|
| Scalar field exact match HEAD_TO_HEAD_CORD.md · set-both-attempt | 67.0% | 97.3% |
| False-invention rate (scalar) HEAD_TO_HEAD_CORD.md | 1.9% | 1.4% |
| Field coverage @ risk ≤ 0.05 cord-test · certified tier1_practical | impossible | 99.6% |
| Document straight-through cord-test · certified tier1_practical | impossible | 97.7% |
| Observed selective risk (held-out) cord-test · held-out | — | 0.0250 |
| Cost / 1,000 documents cord-test · measured, not projected | $0.00 | $49.31 |
The honest limit
Line items are measured and delivered — but not yet certified.
Measured on CORD
- Cell-level exact match
- 94.9%
- Whole rows entirely correct
- 84.2%
Line items (tables) are measured and delivered as reviewed output — never straight-through. They are not yet certified on CORD: the best threshold accepts 57 rows with 4 wrong against the required tail.
We state this plainly. Hiding it would be the one thing the product exists to prevent.
CORD cell-level · not certified
The one benchmark that counts is on your documents.
A pilot produces an accuracy report on a held-out slice of your own data — reproducible on your machine.