Reproduce
Plant a failed booking result and capture the unsupported confirmation.
Synthetic booking-truth test · BT-001
A booking is not real because the agent said it nicely.
Run the planted failure, switch on the truth gate, and rerun both the failure and success paths. The verifier compares the agent’s claim with the booking result every time.
- Runtime
- Your browser
- External calls
- None
- Test data
- Synthetic only
- Expected time
- Under 30 sec
00 / Ready
Booking commitment verifier
- 00Ready
- 01Reproduce
- 02Gate
- 03Retest
- 04Control
Start with the planted failure. No API or provider will be called.
Input ABooking backend
Tool result
- Status
- WAITING
- Appointment ID
- none
Rule BTruth boundary
Confirmation gate
A confirmation needs two things:
status=success
appointment_id≠empty
NOT CHECKED
Output CAgent response
What the caller hears
Waiting for the synthetic booking attempt.
Check DInvariant
Evidence verdict
READYThe verifier will compare the agent's words with committed backend evidence.
Machine-readable idea, human-readable trail
Event evidence
- fixtureSynthetic caller asks for Tuesday at 3:00 PM
- boundaryNo network, credential, phone call, or real customer data
Why this matters
A false confirmation becomes tomorrow’s support ticket.
Constrain
Allow success language only when the write succeeded and returned its identifier.
Retest
Run the same failure again, then use a valid success as the anti-cheat control.
Apply it to one real workflow
Start with the failure your client cannot afford to fake.
The first $199 pilot covers one authorized workflow, 10–12 selected deterministic scenarios, one agreed remediation pass, and one retest. Provider usage is separately capped or client-funded.