Skip to content
Menu

Design principles / PRAXENOR DESIGN NOTES

Keep important AI actions behind approval.

For actions that need human approval, the system should wait for that decision. A label on a screen does not enforce permission.

A Praxenor design note. For what the products currently do and their limits, see Veristade’s current status.

Approval should control the next action.

For example, an AI tool may draft an email while sending it requires approval. The system must hold the send action until the right person approves it.

Define what is paused, which records the reviewer can see, what action needs approval and what that approval will allow.

Show the evidence behind the proposed action.

Show the requested result, source material, assumptions, checks, unresolved questions and the exact action awaiting approval.

Give the reviewer a way to ask whether the result meets the original criteria. The interface should not make approval feel inevitable.

Make it easy to ask for changes.

Record the changes the reviewer asks for, send the work back to the right step and clear any earlier approval.

The revised work should pass through review again. It should not inherit approval from the version that was rejected.

Test the approval boundary.

In a fictional email task, try to send before approval. Revise the recipient or message after approval and try again. Check that approval for the earlier version cannot authorize the changed request.

Record the decision and the execution result separately. A working button does not prove that the receiving system enforces the approval. For implementation questions, read technical details.

← Back to security insights