This page says what this system does, what it does not do, and which of its guarantees are enforced today. The claim this system makes is narrow on purpose. It is not a claim about accuracy at all. It is that every value it reports can be traced to the sentence it came from — and that it tells you which of those traces it has actually proved.

In place today

• Quotes are grounded server-side. The model returns the sentence it read a value from; the server then searches the document’s own extracted text for that sentence and records what it found. The context shown to a reviewer is always the document’s text, never the model’s quote.
• Statuses use the data’s own words. Verified, found in several places, not found in the document text, read from the page image. The dashboard, the field chips and the exports all use the same four.
• Around twenty exception rules run in ordinary code — coverage, arithmetic, sequence, option-notice logic — with no second model call, so the findings are reproducible and free to re-run.
• Critical dates are computed, not reasoned.
• A second model can re-read a document and is deliberately not shown the first model’s quote, because being handed the quote turns the question into an easier one than the one being asked.
• Every field correction is appended to a record that keeps the value it replaced. Corrections rewrite the stored value, so anything downstream reads the correction rather than the original.
• Each deployment is isolated: its own Worker, its own database, its own document store. Neither can reach the other’s documents.
• No model credentials sit in the application for the chat path — it is authenticated by a platform binding, not a key.

On the roadmap

These are not built. They are listed because a regulated buyer will ask, and finding out later is worse than being told now.

• □ Roadmap Authenticated identity behind every review action. Today a reviewer types their name, so the trail records what changed and when — not a verified who.
• □ Roadmap Roles, and separation between the person who edits a field and the person who signs it off.
• □ Roadmap Data residency. Deployment inside your own account or region, with your own model keys.
• □ Roadmap Classification and masking of personal data in the pipeline and on export.
• □ Roadmap Cross-model verification as part of the record. It runs on demand today; its result is not yet stored, exported, or used to order the reviewer’s queue.

Retention

An uploaded file is kept so a citation can be opened against the page it came from. The storage bucket carries a rule that removes uploads after 30 days, and clearing the history removes them immediately.

That rule covers the stored file, and not every place document content comes to rest. Extracted text and fields are held in the database, and requests to the model provider may be retained by that provider on its own terms. A single documented retention position covering all of them is being written; until it exists, treat the sentence above as describing one mechanism rather than a guarantee about the whole system.

What this page does not claim

• No accuracy figure, and no benchmark. Any number you see here is a count measured on the documents this deployment has read — not a claim about how it will perform on yours.
• No blanket deletion promise. See Retention above.
• No claim that the audit trail identifies who made a change.

What this deployment deliberately does not do

• No integrations. Documents arrive by upload here, not from your mail or document systems.
• No user accounts or single sign-on — access is one shared passphrase.
• No service guarantees. This runs on demonstration infrastructure.
• Nothing is approved automatically. Approval is a person’s act, recorded against the name that person typed.

Please do not upload confidential or production documents. Use samples or redacted copies.

How we work

We have built a platform of composable blocks. Each one handles a distinct part of how documents move through a business. We assemble the blocks that fit, then implement the workflow inside your enterprise — connected to the systems your teams already use, with your rules and your reviewers.