Documents as they are actually filed
An extraction model that has only seen clean templates has not seen this corridor's paperwork — the stamp over the field, the third-generation photocopy, the form filled in two hands.
The problem
Document AI trained on clean synthetic templates and Western form layouts degrades sharply on the real thing: rubber stamps landing across the values you need, handwriting interleaved with print, scans of scans, and layouts that were never standardised in the first place.
The corpora that would fix it are held by institutions, not published, and cannot simply be bought. They come with obligations that survive the sale — which is why this line is sourced to a mandate rather than held as stock.
What we put against it
Enterprise and operational document corpora, property and transaction records, and de-identified institutional records — each sourced against a named mandate from the holding institution, not assembled from a standing inventory.
On the institutional line the constraint is hard and it is written into the schema, not only the copy: the holding institution is the only counterparty, de-identification happens under its supervision before anything leaves its control, and account identifiers and authentication material are never received, held or delivered. That line carries no public sample and will not be given one.
What this has to satisfy
| Requirement | How it is met |
|---|---|
| Real degradation, not synthetic noise | Photocopy generations, stamps over fields, mixed hand and print as filed |
| Layouts that were never standardised | Operational forms in the shape the issuing body actually uses |
| De-identification that survives audit | Performed under the holding institution's supervision, before egress |
| A defensible counterparty | The holding institution is the only counterparty; mandate named per corpus |
What you can check first
- Schema published
- The records schema and a transform log are published as samples
- Never received
- Account identifiers and authentication material are never received, held or delivered
- Mandate first
- No corpus is presented as inventory; each is sourced to a named mandate
- Supervised egress
- De-identification completed under the institution's supervision before transfer
Supply lines behind this
A solution is marked live only where a line behind it holds cleared material. Nothing sourced to mandate is presented as inventory.
Fifty delivered hours at production standard, two weeks from spec sign-off.
Other solutions
ASR and speech for languages the benchmarks skip
Word error rates in this corridor are not a modelling problem. They are a data problem, and the data was never collected.
Post-training and evaluation outside the crawl
The web is not a sample of the world. For this corridor it is a sample of what got typed in English by people with a connection — and a model trained on it inherits that gap as confidence.
First-person work, including the part where it goes wrong
Manipulation policies generalise from the variety of hands, tools and failures they have seen. The open egocentric corpora are a fixed process in an industrial interior, filmed until it works.