What Resolve decides
- Outcome:
refund_full,refund_partialorno_refund - Refund amount in cents, and whether the item must be returned first
- Reasons in plain English, citing the evidence (
S1,S2…) and the policy sections applied - Confidence (
low,medium,high), and whether a person should review it
Who is in charge
Your policy sets the limits
The AI’s decision is final only when the amount is at or under
auto_decide_max_cents, its confidence is not low, it has not flagged the case, and every server-side check passes. Everything else waits for your reviewer.Your users can always reach a person
After any decision, either side can ask a person at your platform to review it once. If you enable arbitration, either side can ask for a licensed arbitrator, and the case goes ahead when the other side agrees.
How the AI decides
Resolve uses Claude Opus with structured output. The model sees your policy text, the transaction record and every submission, including photos and PDFs. Party names and emails are replaced with “Buyer” and “Seller.” It is told to:- apply only your policy and the evidence,
- cite a submission for every fact,
- give both sides the same care,
- flag cases the policy does not cover, where the evidence conflicts in a way only a person could judge, or where fraud or safety is involved.
- the refund is between 0 and the amount;
- the outcome and the amount agree;
- every citation exists.