A timeout is not a failed refund.
Separate the intent, the attempt and the effect. Then reconcile what you actually know.
Research & reproducibility
What makes an automated action dependable? Our first experiment examines a refund retry. Inspect its inputs, method, result and limits; the example is synthetic.
Separate the intent, the attempt and the effect. Then reconcile what you actually know.
An approval needs an exact scope, a current authority and a meaningful expiry.
Measure adherence to active constraints, rather than similarity to a preferred path.
Three questions to ask of a certificate: what was checked, is the record intact, and does it still apply?
For a customer's security or risk reviewer: three gauges to read separately, who issued it, what a fingerprint lets you check and what a certificate cannot tell you.
A starting checklist for buyers, risk reviewers and builders assessing an agent that moves money, changes records or makes commitments.
A refund agent can pass every code scan and still refund twice. What each kind of check covers, and where they overlap.
A reproducible synthetic experiment shows why losing a reply and losing an effect are different events.
We plan to publish real audits as case studies with their own certificates. Each will identify an exact revision, a permitted reproduction, problems and limits. The aim is evidence another team can use when deciding what to delegate.
Our longer-term questions include meaningful oversight, faithful explanations and response to changing conditions. See the future research direction.
Certificates & case studies →