Observed arrival · 2026-10-12
A policy audit that follows the refund into the ledger
Keyturn Labs compares agent releases against reviewed policy tests, tracing tool calls to the effects recorded by a test backend.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Teams releasing agents that can issue refunds or take actions
- Worth noticing
- In one reported case, a model refunded $49 in all three runs despite a missing chargeback status; the gate blocked all three calls.
Field notes
The sample centers on a simulated refund workflow with 50 cases and four approaches to gating actions. A separate report describes 52 reviewed cases run three times each against the site's own billing backend, which keeps its own ledger. The site reports that an execution gate refused the problematic refund call in all three reruns; it also states that the company, customers, and data were fictional.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
$PaidCommerce or pricing visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue