Research.

The open question: how to make a system for high-stakes documents show its sources, abstain when evidence is missing, and leave the decision to a person.

A model’s intelligence does not create trust. Trust is the architecture around it: evidence binding, a deterministic check, an explicit abstention, and a hand-off to a person.

  1. 01What counts as evidence when sources disagree?
  2. 02When must the system stay silent instead of guessing?
  3. 03How do you tell a code candidate from a decision someone could file?

What exists today.

  • Working prototype — trade documents, on the Customs page.
  • Demonstration rows 8457 and 8458: a class of goods, not a ten-digit code, not a client file.
  • Research layer — a gap, a question for the plant, a “needs review” state.

Customs

Limitations.

No customers, no revenue, no completed deployments. We do not publish accuracy percentages on live declarations. The prototype does not file and does not replace a specialist.

Discuss an evaluation