Demos / Documents
Reports in as PDFs. Checked figures out.
A test of reading water-test lab certificates into a table with code, checking every value against written limits, and writing a report where each finding cites its rule. Anything the code is unsure of goes to a person instead of through.
- Data
- US Geological Survey water-quality results (public domain), set out by us as lab certificates
- Model
- None. Python with pdfplumber. No OCR.
- Recorded
- 18 November 2026
Recorded test, not client work. There is no staff app or film for this one yet; the viewer below was drawn for this page from the recorded run.
The result
No wrong value got through.
50 held-out certificates in four layouts. A simple line-pattern reader got 330 of 477 values right on the same certificates; this reader got 477 of 477. Reading took 0.021 seconds per certificate on one desktop machine, file refusals included.
Where it falls short
- We set out the certificates ourselves from real measurements. Reports from real labs vary more, and each new layout needs its own check.
- Scanned and cut-off certificates were not read. They were sent to a person.
- Missing units, repeated entries and implausible values still need a person.
One certificate from the run
The page beside the table read from it.

| Measured | Value | Unit | Limit | Finding |
|---|---|---|---|---|
| Calcium | 20.8 | mg/l | none | No limit in this table |
| Magnesium | 3.69 | mg/l | none | No limit in this table |
| Sodium | 4.52 | mg/l | none | No limit in this table |
| Potassium | 1.01 | mg/l | none | No limit in this table |
| Iron | 2390 | ug/l | 0.3 | Over limit |
| Manganese | 65.5 | ug/l | 0.05 | Over limit |
| Copper | 0.98 | ug/l | 1.0 | Within limit |
| Zinc | <2 | ug/l | 5 | Within limit |
| Arsenic | 0.44 | ug/l | 0.010 | Within limit |
| Lead | 0.090 | ug/l | none | No limit in this table |
| Fluoride | 0.08 | mg/l | 4.0 | Within limit |
| Nitrate | 0.050 | mg/l as N | 10 | Within limit |
| pH | 7.9 | std units | 6.5:8.5 | Within limit |
The report it wrote
- Iron: 2390 ug/l. Over limit. Rule: 40 CFR 143.3 2025 edition
- Manganese: 65.5 ug/l. Over limit. Rule: 40 CFR 143.3 2025 edition
- Copper: 0.98 ug/l. Within limit. Rule: 40 CFR 143.3 2025 edition
- Zinc: <2 ug/l. Within limit. Rule: 40 CFR 143.3 2025 edition
- Arsenic: 0.44 ug/l. Within limit. Rule: 40 CFR 141.62(b)(16) 2025 edition
- Fluoride: 0.08 mg/l. Within limit. Rule: 40 CFR 141.62(b)(1) 2025 edition
- Nitrate: 0.050 mg/l as N. Within limit. Rule: 40 CFR 141.62(b)(7) 2025 edition
- pH: 7.9 std units. Within limit. Rule: 40 CFR 143.3 2025 edition
- Hardness: 67.13302 mg/l as CaCO3. Derived value. Rule: EPA 2021 MSGP Appendix J page J-2
Sentences are filled from approved templates by code. No AI writes the report.
Nothing hidden
All 52 values sent to a person.
Each with the reason. The answer key is shown beside what the reader returned, including values in files it could not read.
Filter by the reason a value was held back.
For the engineer
Why code and not a model.
A wrong number that looks right is worse than no number. Code reads each name, value, unit and limit and records where on the page it found them. Units, ranges and totals must agree, and a rules table decides each finding, so the same input gives the same finding every time.
What this test does not show
- Invoices, scanned pages or handwriting. Scans were refused and sent to a person.
- Certificates from real labs, which vary more than the four layouts we set out.
- A person reviewing the held-back values. That step was not run.
Let's talk
Have documents that get retyped?
Tell us what the documents are, roughly how many arrive, and where the figures need to end up.
You will get a reply by email from Matthew, the engineer who would do the work.
Your enquiry goes straight to our inbox. We use your details only to reply to you. Privacy