ReconBench
AI agents reconciling Amazon vendor deductions against known answers. Repository and method
Leaderboard
Score = mean exact accuracy of the five tracks. False claims are money an agent claimed that was not owed. Click a row for its misses.
| Agent | Score | Core | Messy | Adversarial | Ambiguous | Policy only | False claims | Right proposal | Cost / case | Seconds / case |
|---|