01 / provider assessment
resolved
Agent: “Your $48 refund has been issued.”
source / conversation + provider label
independent evaluation for AI agents
An agent said it worked.Did it?
UnSure is being built to check the work. Evaluate AI agents against your requirements, trace each finding to its evidence, and make the business consequences easier to see.
a small exampleyour criteria. What the agent was expected to do, made explicit.
the evidence. What the records support, with sources you can inspect.
room for unknowns. When evidence is missing, the question stays open.
Same conversation. Different picture.
A fictional case. Both assessments ask: did the refund complete?
02 / outcome evidence
refund.failed
The transaction log records a failed refund. No completed refund is recorded.
source / transaction log
purchaser criterion
The requested refund was completed.
finding against this criterion
Evidence-supported discrepancy.
The provider’s completion assessment conflicts with the recorded outcome.
Evidence for the next decision.
We’re building toward a clearer basis for comparing agents, investigating their work, and exploring loss and exposure under stated assumptions.
Starting with customer service. Extending to marketing and PR.