Resettable Stripe-like mock with synthetic state and 22 MCP tools
Interactive Proof Lab
Inspect the run. Check the state. Verify the result.
Explore sanitized evidence from real Shapd AI evaluation work. Choose a retained run, replay its steps and see why the verifier accepted or rejected it.
Verified passRejected runReplay evidence · No live production systems
Stateful MCP · Financial operations
Create exactly one customer after checking for an existing record
Confirm that a customer with the supplied email does not already exist. If absent, create exactly one customer with the requested identity.
A plausible agent can skip the lookup, create a duplicate or report success without proving the final state. Each mistake creates operational cleanup and unreliable downstream state.
Customer searchCreate customerState inspectionIndependent verifier
Choose retained run
RESULTPASS
REWARD1
DURATION17.6s
PROVENANCEModel trajectory
The agent searched first, found no match, created one customer and left a single valid resulting record.
Codex baseline · retained 17 Aug 2026Plan
Check whether the target email is already present before creating anything.
Want this level of evidence for your agent?
Bring us one consequential workflow. We will propose the smallest useful evaluation pilot.