An e-invoice carries many details twice: as a structured field and as free text. Software that reads only the structured fields misses a detail that appears in the text alone. We measured 34 test invoices from the XRechnung test suite, ten questions per invoice, 340 verdicts in total.
Test bench
Each row is a test invoice, each column a detail. The fill shows how sure Jev is that the free text contains the detail. The dot in a cell shows that the structured field is filled. The replay runs the measurement at the response times we recorded.
Test bench · field against text
Replay of the measurement from 22/09/2026no live requestModel jev-latest (jev-1.13.0)
34 requests, 340 verdicts, 11.3 s of recorded response time.
Arrow keys move between cells, Enter opens the detail panel.
Select a cell in the grid. The question, the verdict, the structured field and the free text of the invoice then appear here.
The numbers
Each dot is one verdict. Most sit close to 0 or close to 1. The hatched band between the thresholds is the set a person would look at. Its size is in the figures below.
| Field filled | Field empty | |
|---|---|---|
| Text yes | 8 | 0 |
| Text no | 16 | 10 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 1 | 4 |
| Text no | 0 | 29 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 0 |
| Text no | 3 | 31 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 0 |
| Text no | 8 | 26 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 0 |
| Text no | 3 | 30 |
1 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 3 | 0 |
| Text no | 6 | 25 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 0 |
| Text no | 10 | 24 |
0 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 0 |
| Text no | 15 | 17 |
2 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 0 | 2 |
| Text no | 11 | 13 |
8 uncertain
| Field filled | Field empty | |
|---|---|---|
| Text yes | 2 | 3 |
| Text no | 7 | 19 |
3 uncertain
Findings
We count only cells in the class "text only" that a second language model checked one by one against the wording of the invoice. With 34 test invoices these are small numbers. They show that it happens, not how often it happens in real invoices.
4of 34 test invoices
In 4 of 34 test invoices the text offers an early payment discount that appears nowhere in machine-readable form. EN 16931 has no dedicated field for it. XRechnung provides an encoding with #SKONTO# in the payment terms text BT-20, and that encoding is missing here.
„Bei Zahlungen binnen 14 Tagen, 2% Skonto“
1of 34 test invoices
In 1 of 34 test invoices the text states a service period, yet neither the invoice nor any line item carries a period with start and end date (BG-14 / BG-26).
„Bearbeitungszeitraum: Dezember 2014/ 5.01.-09.01.2015“
Refuted by the cross-check and not counted: 1.
1of 34 test invoices
In 1 of 34 test invoices the text says when goods were delivered or the service was performed, at least to the month. The delivery date BT-72 is empty.
„Die Lieferung erfolgte im Monat 06/2016.“
Refuted by the cross-check and not counted: 2.
Counted: 6 confirmed cells. Refuted and not counted: 3. Not cross-checked: 0.
Setup
The model judges the text and nothing else. What a field contains and what follows from it is decided by code that you can read and change.
01Code
A program checks in the XML file whether the structured field is filled, for example BT-9 for the payment due date. That needs no model. Where the check covers more than one field, the column says so, for example BG-14 / BG-26 for a period at invoice or line level.
02Model
Jev receives the free text only: notes, payment terms, line item texts and the reason for a VAT exemption. The ten yes or no questions go out in one request. For each question a probability between 0 and 1 comes back.
03Code
The comparison and the thresholds live in code. Moving the thresholds changes the classification, not the verdicts. No new request is needed for that.
{
"state": {
"zahlungsbedingungen": [
"Bei Zahlungen binnen 14 Tagen, 2% Skonto"
]
},
"model": "jev-latest",
"questions": {
"skonto": {
"type": "noul",
"instructions": "Does the text offer a cash discount (Skonto) for early payment?"
}
}
}{
"answers": {
"skonto": {
"noul": 0.99
}
},
"usage": {
"input_tokens": 734
}
}export function klasse(p: number, feldBelegt: boolean, unten: number, oben: number): Klasse {
if (p > unten && p < oben) return 'unsicher';
const textSagtEs = p >= oben;
if (textSagtEs && !feldBelegt) return 'nurText';
if (!textSagtEs && feldBelegt) return 'nurFeld';
return 'stimmig';
}Limits
This page shows a measurement on test data. Here is what it supports and what it does not.
The invoices come from a test suite. They are built to exercise rules and say nothing about how often a detail appears only in the text of real invoices. Some test cases are variants of the same invoice with the same free text, so a finding in several rows can go back to the same wording.
The questions to Jev are in English, the invoice texts in German. We did not measure German questions against them.
Each verdict comes from a single run. In an earlier measurement on the same test invoices, with different questions, a second run differed by at most 0.09.
At the default thresholds a second language model read every cell in the classes "text only" and "uncertain" against the wording of the invoice. A third, independent check then overturned three cells, which the detail panel marks as refuted. No person has reviewed this cross-check yet. A month in which goods were delivered counts as a delivery time, not as a service period. Consistent cells were not checked. If you move the thresholds, the detail panel shows which cells are unchecked.
For invoices with more than twelve line items, Jev saw the texts of the first twelve only. Anything further down was not part of the request.
The page plays back stored results. Your browser sends no request to a model.
The test invoices come from the XRechnung test suite by KoSIT, published under the Apache-2.0 licence. The price per million input tokens is listed in the TypeSafe model overview.
Bring a sample. Together we look at which details reach your accounting today only through free text and how that can be checked.
Schedule a conversation