refund-agent-v1
Refund of £42.00 approved for order #10428. You'll see it within 3–5 days.
The answer stayed correct. The route did not.
The same answer, twice
refund-agent-v2
Refund of £42.00 approved for order #10428. You'll see it within 3–5 days.
The customer-facing answer is materially identical. The difference is in what the two releases did: one checked the refund policy and the customer’s fraud signal before issuing a single payment, and one did not.
Demo state
- Demo mode
- ENABLED
- Project
- 019f9fff-f7f8-736e-aeac-dff9358c3e2a
- Agent
- 019f9fff-f7f9-7147-a7a1-7f5d13b3688f
- Baselines captured
- 1
- Contracts
- 1
- Active contract
- 019fa000-1484-718f-a748-2f040a8afab5
- Checked
- 2026-07-27 01:12:27
Controls
- Run approved v1
- DEMO_RUNS=25 make demo-v1
- Run unsafe v2
- DEMO_RUNS=8 make demo-v2
- Capture baseline
- make demo-seed
- Evaluate v2
- flightrules release evaluate --release refund-agent-v2
- Reset demo
- make demo-reset
Each control runs real product behaviour: real OTLP telemetry, a real baseline mined from it, a real contract, and a real release decision. Nothing here fabricates a trace or a result.
Then look at what happened
Once both releases have run and been evaluated, the release decision and its evidence are on the releases page for the agent.