How to test
The tour as a checklist: 7 stops
- 1. Every reply, traced and checked
DoPress a measured question under the chat.
CheckToday's model left, the open model on Dvar right. Each reply shows its calls, ₹, latency and 7 checks.
- 2. The scorecard
DoRead the Gates column.
CheckToday 89.4%, bar 88.4%, DeepSeek V4 Flash on the harness 92.5%. 2 of 4 gates; not yet: Delayed, on hold or failed 74%, p99 23.3 s a chat.
- 3. A reply the checks blocked
DoOpen a case that failed a critical check.
CheckCaught in code before the merchant saw it, re-run on the frontier model, kept as an eval case.
- 4. Traffic as traces
DoRead the Topics bar, then open a chat.
Check7 intents, each chat in one. Names and account numbers masked before storage.
- 5. The rubric
DoRead the 7 checks.
Check3 critical: facts exact, privacy, hand-off. Then language, right action, facts complete, and one judged.
- 6. Compare and promote
DoRead why each candidate fails, then press Promote to gateway.
CheckFour gates decide. Promote routes Ray's next answer to that model, in this browser.
- 7. Two ways to connect
DoRead options A and B.
CheckBeside Ray's calls in Razorpay's network, or through the gateway. Nothing leaves by default.