Your policy.
Your agent. Actual evidence.
Bring an existing policy. Review the rules we find, run your agent against local test tools, and inspect what changed.
Open your workspace.
After we agree the audit scope and receive payment, we send your setup instructions. Use your invite code or saved access file to return here.
After checkout, your workspace opens here. Returning buyers can use a saved access link or file. Invited pilots can enter their code.
Buy pilot · $499 ↗This workspace supports one English refund workflow, up to 10 rules. Your engineer runs a compatible Python adapter against local test tools. Try the free policy check, or request a $499 audit so we can confirm fit.
Choose audit-access.json from your connector ZIP, or a saved purchase access file.
Save a way back.
Download your access file or save this link before closing the tab. Anyone with the link or file can use your workspace; keep them private.
- 01 Import policy
- 02 Review rules
- 03 Connect & capture
- 04 Inspect results
Start with the source.
Paste your English refund policy or load a plain text file. We support refund windows, amount limits, confirmation, ownership, duplicate refunds and currency restrictions.
Confirm what the policy means.
Each rule points to the text it came from. Check the interpretation; clarify the source above if anything is wrong.
Make the tests yours.
These labels were generated from the approved rules. Review each state, expected outcome and tool arguments before using them as test evidence.
Unzip the starter, connect your Python agent in customer_agent.py, then run these commands from that directory. Replace the version labels with your actual release versions.
The starter includes simulated refund tools and capture code. Your agent code and model credentials stay local. Policy text is sent during extraction and review; original capture evidence is sent when you submit an audit below.
Compare what the agent did.
Upload the connector's original files. Final assistant answers alone cannot establish that a refund happened.