The trust layer has to do some work

A well-written answer can still be wrong. A consistent answer can be wrong too.

In legal AI, I do not think trust can come from how confident the output sounds. It has to come from the way the work is controlled.

Take a financial calculation. Let the model help interpret the question, but use a defined calculation method for the numbers. Keep the inputs visible. Check that changing the presentation has not changed the calculation. Repeatability is useful, but it does not replace checking that the method is right.

The same principle applies to legal workflows. What information was used? Which rule was applied? What could the system do on its own? What required approval? And who accepted the result?

These questions should shape the workflow before a tool is put into use. An audit trail written after the event cannot undo an action that should never have happened.

This is what I mean by a trust layer: checks, limits and approvals between an AI output and the next step. Some work can proceed automatically once those checks pass. Some should stop for a person.

Put simply: govern the action, not just the answer. That is where the policy has to become working software.