Know what you are approving
FusionWall is designed around a simple idea: an important action should not disappear between a button click and an AI provider. The approval checkpoint shows you the exact protected request before it leaves. The private receipt records what happened afterward. Let me show you how to read both without needing to be a security engineer.
Pause at the checkpoint
At the checkpoint, start with the basics. Is this the question you meant to ask? Is the selected context appropriate? Is it going to the provider and model you intended to use? Then read the policy result and any warnings. The screen is there to support a real decision, not to become a form you click through.
Read what may leave
Now read the protected request itself. Confirm that names, account details, and custom terms have been replaced where expected. Make sure the remaining text still contains enough meaning for a useful answer. Detection is not perfect, so stop if anything private remains. Canceling is a normal and safe outcome.
Send, revise, or cancel
You have three reasonable choices. Send it when the protected view is correct. Revise it when the question or context needs work. Cancel it when the request should not leave at all. The approval applies to this visible action and route. It is not blanket permission for future questions.
Open the linked receipt
After the answer, open the linked receipt. You can see the request fingerprint, rules that applied, chosen model, approval, status, and connections to related records. The ledger is tamper-evident, which means unexpected changes can be detected. It is evidence of the workflow, not a certification that the AI answer is correct.
Follow the chain
If a result surprises you later, start with the receipt. Compare it with the conversation, selected source boundary, active policy, and recorded result. That gives a founder, contractor, operator, or auditor a common set of facts. You can correct the work without relying on somebody's memory of what they clicked.
See it, decide it, keep it
A useful approval is specific enough to understand and limited to one action. A useful receipt connects that decision to the outcome. Read the request, choose deliberately, and use the proof when you need to explain what happened. That is how human authority stays real instead of ceremonial.
The idea to keep
Approval applies to the visible action. A receipt records the governed path; it does not certify that an AI answer is correct.