# Documents that carry liability

- Page: https://scanovich.ai/research/documents-that-carry-liability/
- Language: en
- Updated: 2026-11-11

- Status: active

**Question.** Can a system turn messy operational documents into a record a named person can sign, and stop when it does not know?

## Why it is uncertain

Supplier spreadsheets, degraded scans and handwritten notes arrive incomplete and inconsistent. A language model will produce a plausible value for every empty cell. Work that carries liability needs something else: a value with a source, or no value at all.

The open problem is where the system should stop. What counts as enough evidence for a field, a classification or a code is different in every domain, and it is rarely written down.

## Method

- Procedure before generation. Classification follows the legal method step by step instead of asking a model for an answer.
- Every value keeps its provenance: which document, which line, which rule.
- "Insufficient data" is a normal outcome, recorded and shown to the reviewer.
- An adversarial recheck tries to break each proposal before a person sees it.
- Outputs are validated against the official schema, so a draft is structurally correct or it is rejected.
- A person with authority signs. The system prepares; it does not decide.

## What we measure

- Share of fields that carry a source.
- How often the system stops, against how often it is wrong when it does not stop.
- Reviewer correction time per record.
- Behaviour on bait cases built to invite a plausible wrong code.

## Results so far

A working prototype for trade documents. Incomplete supplier lines become a reviewable product registry with sources and gaps. Classification proposals cite the rules they rely on. Draft declarations validate against the official schema.

An accuracy figure is not published. It waits for a held-out set of real declarations, which is the next piece of work.

## Next

Assemble the held-out set. Compare against a strong baseline on the same inputs and the same budget.

## Projects on this line

- [Trade-document engine](https://scanovich.ai/en/customs.md) — prototype, closed code. Incomplete supplier lines become a reviewable record with sources and gaps. A person keeps the decision.
- [Document-intelligence core](https://scanovich.ai/projects/document-core.md) — prototype, closed code. One engine loop for every document domain — a new domain adds its own validators and precedents, not a new system.