What Thaw sees when it reads, and what it never sees

Thaw reads documents with a large language model from a provider named in our terms. Here is what goes across, what comes back, and the rules around it. If a product is going to read your passports, it should be able to answer this precisely.

What is sent

The document. A PDF is sent as a document; a photo as an image; a text file as text. Files over 30 MB or 100 pages are sent as their extracted text instead.

The text layer, when the PDF has one, as a cross-check for Thaw. It is told the file itself is authoritative.

The registry summary. The twelve categories, the kinds of paper, and the curated types with their fields. This is the same for everyone and sits behind a prompt-cache breakpoint, so it is processed once per cache window rather than per document.

The household glossary. The people, pets, vehicles, properties and businesses already known in your family or business, with their aliases and identifying attributes (a plate, an address). This is what turns “Dana” on a prescription into the right person. It is per family and also cached.

Hints. The file name, and if the document came from an import, whatever the previous system called it: notebook, tags, correspondent, date. Thaw is told these may be wrong and to confirm from the page.

What is never sent

Nothing from another family or business. Nothing from the audit log. Thaw does not have tools that fetch more; it sees the one document and the context above.

What comes back

A structured record: category, form, type (or a proposed type in plain words), title, summary, document date, period, facts with roles and page numbers, the people and things mentioned with roles, obligations with problem and solution, a supersession note, a confidence score, and a reason if the document needs a look.

Every reading is stored: model, prompt version, tokens, estimated cost, duration, whether it succeeded. You can see them on the document page.

What happens to identifiers

Facts with the identifier role are encrypted with a blind index and redacted from page text before anything is indexed. Ask sends the model only the excerpts that match your question, from that redacted text.

Your corrections win

Correct a title, a type, a date or a fact and it is marked as yours. A later re-read updates everything else and never touches those fields. There is a test for it.

Retention at the provider

API inputs are not used to train models. We use models eligible for zero data retention and will move to a ZDR agreement when volume justifies it; until then the provider keeps inputs for 30 days, which the security page says.