the assay report
987‰98.7% of every human-checkable field was read correctly — measured over 1,061 documents, computed live from the register, never asserted.
- documents processed
- 1,061
- fields auto-accepted
- 61.1%
- documents a person saw
- 92.8%
- mean cost / document
- $0.000022
- mean reading time
- 8.7s
- register events
- 13,024
distinct documents with a completed reading
fields the gate cleared ÷ all extracted fields
documents with ≥1 held field ÷ all processed
power-metered GPU energy × $0.11/kWh
mean over all recorded readings
append-only, hash-chained, verifiable
The scale run, measured
One recorded batch, 2026-07-05: 1,001 documents on both cards in 76 minutes, zero failures; batch accuracy 98.77% (5,873/5,946). Full method: docs/METRICS.md in the repository.
RTX 3080 Ti
0documents / hour
8.7s per document · 0.2763 Wh · $0.0000304 energy each
RTX 4070
0documents / hour
7.8s per document · 0.1268 Wh · $0.0000139 energy each
Accuracy by field
matched ground-truth verdicts ÷ all measurable verdicts, per field, whole corpus — weakest first, published.
invoice number
93.6% · n=932
buyer
98.9% · n=659
total
99.5% · n=849
subtotal
99.6% · n=719
date
99.7% · n=1039
tax
99.8% · n=405
supplier
99.9% · n=724
currency
100.0% · n=975
The invoice number is the weakest field — on some templates the reader keeps the printed label glued to the number. Real behaviour, recorded, and the target of the calibration stage.
Does the grade predict correctness?
accuracy split by routing — if grading works, hallmarked fields must be right more often than held ones. They are.
hallmarked (auto-passed)
99.9% · n=4328
held for a person
96.3% · n=1974
Accuracy over runs
60-document proving run · 2026-07-0398.3%
1,001-document scale run · 2026-07-0598.8%
held at scale across all 50 layouts, 20 of them never seen before the run