Start with the guided demo
See a Franchise Lead Safety Pack run without sign-in, customer data or an API request. It shows the intended workflow and a deliberately blocked release.
Practical guide
Turn an AI update into a release decision.
ReleaseProof compares the same customer scenarios across a baseline and candidate assistant. It helps an agency collect evidence before an update reaches a client - it does not replace a human decision.
Prepare one comparable set
Use 10-12 representative, redacted customer scenarios. For each, collect the same prompt and the answer from Version A and Version B.
Define the rules that matter
Add objective checks, semantic criteria and your client-specific risky promises - for example, no guaranteed profit, territory or callback commitment.
Run a private review
Sign in to evaluate your own dataset. Results, evidence and run history stay in your private workspace.
Share only a deliberate report
After checking the evidence, publish a saved review only when it is safe to share. Anyone with that report link can read it.
Private review
What to prepare before a private review
- Representative, redacted conversations - not raw production logs.
- The same customer input and output from both assistant versions.
- A short description of what changed in the update.
- The commercial or safety promises that need a human decision.
Release decision
What to do with the result
Confirm and release
Review the evidence, document the human confirmation, then release the candidate version.
Resolve the flagged cases
Assign an owner, clarify the expected behavior and rerun the same cases after the change.
Do not expose this version yet
Remove the material regression, test it against the same scenario, then create a new decision record.
Important boundary
ReleaseProof is a release-quality workflow, not legal, financial or compliance certification. Critical findings still require independent human confirmation.