Define success, choose the covered action, and retain every approval decision.
For AI support teams handling refunds and order cancellations
Two requests. One refund. →Replay the recorded demo · no new refundsCheck authority before the action, then read the target system to verify the result. Our first partners are AI support and ecommerce automation companies with an engineering team and one action path to connect.
Become a design partner · founder-guided setup →Assurance for AI Agent actions
Help keep AI mistakes from becoming real losses.
Check permissions and control AI actions before they happen. Verify the results afterward and keep a receipt. Protection covers connected, configured action paths.
Recorded test-mode replay · no new refundsSee the method →
Stop unauthorized actions before they happen.
An AI request is not permission to act. Witnora blocks requests outside the rules and holds requests that need your approval. Authorized actions stay within the configured scope; results are checked afterward.
- Held for approval
A refund that needs your approval is held before execution.
No refund yet - Authorized
Your approval covers only this refund, once.
Customer decision - Executed
The approved request uses a single-use execution grant.
Controlled action - Verified
A separate read checks the actual refund record.
Observed outcome - Preserved
The receipt links approval, execution, the observed result, and limits.
Signed Receipt
Verify the result. Preserve the proof.
Witnora keeps what the Agent said separate from what a customer-operated read path actually observed, then binds authority, execution, outcome, and limitations into one signed Receipt.
The statement is retained, but it is not treated as proof of the business result.
ReportedA separate customer-operated read path observed the declared result.
Outcome verifiedOne action. One authority chain. One inspectable result.
- 01Requested$50 test-mode refund
- 02AuthorizedApproved once by the account owner
- 03ControlledExecuted through the registered boundary
- 04ObservedProvider returned status: succeeded
ReportedRecordedEnforcedOutcome verifiedIndependently reviewed
Current sample: outcome verified. Not independently reviewed. A lower evidence level is never presented as a stronger claim.
Evidence makes improvement believable.Improvement makes evidence useful.
What happened, and who can prove it?
Independent evidence connects customer authority, controlled execution, an observed outcome, and a signed Receipt.
- The covered action crossed the customer boundary
- The business result was checked separately
- The proof keeps its scope and limitations
Did the failure leave, and did it stay gone?
The same contract drives baseline reproduction, candidate comparison, Runtime Watch, and regression detection.
- Baseline and candidate remain comparable
- Missing evidence stays inconclusive
- A regression withdraws stale confidence
A fix passed. Runtime Watch found the failure returning.
Prevent a duplicate customer-feedback update after retry.
1 duplicate outcome in 2 approved attempts
0 failures in 2 matching attempts
The duplicate outcome returned after verification
Watch healthRegression detected
Covered observations2 observations: 1 healthy, 1 regression
Last checkAug 26 at 2:08 PM
Next checkScheduled automatically
Keep the lesson. Check it again before trusting the next result.
Retained failures can inform regression checks. Blocking a future write requires a configured Gateway guard and customer-authorized rule for that covered path.
Keep the exact incident as evidence for a task-specific regression check.
A configured guard enforces the customer-authorized rule on its covered path.
A running Harness and configured triggers rerun the relevant checks.
When a watched check finds the failure again, the affected result needs review.