Chapter 6 of 7
Don't believe in magic. Measure it.
We'd rather you test us on your own paperwork than trust our adjectives. Here's how, and where your data lives while you do.
Two kinds of error, and why we prefer the second
A wrong value marked as confirmed
The only error that could slip through unnoticed. The whole design exists to make it rare: two models from different vendors would have to make the same mistake independently, in both of their passes, and the value would still have to pass check digits and sums.
A correct value left blank
The field stays empty as disputed, although it could have been read. That's a cost, not a risk: you see it in the status and a person fills it in. The system deliberately turns the first kind of error into the second.
Measure it on your own documents
We publish no accuracy rates until we've measured them on reference sets. Documents differ too much: a clean e-invoice is not a forty-page payroll scan. During a test we can run a benchmark on your own reference documents and measure three things: how often each model misreads a field, how often it repeats the same mistake in its second pass, and how often two vendors' models make the identical mistake. Together they tell you how often a wrong value could get confirmed on your paperwork.
Want it stricter? The rule has dials: more passes, three agreeing models instead of two, a tighter tolerance for amounts. Stricter settings mean fewer false confirmations and more blank fields.
Where your data lives
It stays in the EU
Documents are stored and read in the EU, in the Frankfurt region. Reading uses EU model endpoints only: a setting that points outside the EU is rejected in the system's code, and a deployment outside the EU fails before it starts.
Your own fishbowl
Your organization comes only from your sign-in; no field in a request can point anywhere else. Another organization's documents simply don't exist for your account. Users see their own uploads; your administrator sees the whole organization. Two-step sign-in with an authenticator app is available.
Goldfish memory, on purpose
System logs keep numbers and errors, never document content or model answers. Working files vanish after 7 days. Originals and result sets are encrypted and versioned.
No training on your files
We don't train or fine-tune models on your documents. The system remembers your approved templates, not the content of what you drop. Before any document with personal data goes in, we sign a data processing agreement with you.