Way Out.
Reconnecting
Connected to live data · every number on this page comes from the running system
Under the hoodEvery number,
Every number,
measured.
Honesty, speed and dispatch results come straight from the running API, the evals and the simulation. If a number isn't measured yet, this page says so.
The AI never writes ride status, ETAs or driver names. Code does.
Honesty at a glanceChecked by code
Not measured yet. Run evals/run.py.
0calls so far—median turn (no real calls yet)
Voice turn latency
—real phone p50 · 0 turns
—real phone p95
—stand-ins p50 · 0 turns
—stand-ins p95
Measured inside the API, per turn. Interpreter: —. Budget: interpreter ≤ 1,000 ms at p50.
Dispatch vs. todaySIM
Waiting for the impact simulation (data/sim-results.json).
What the agent heardSIM
Not measured yet. Run evals/run.py.
Census right now
- 0
- open ride requests
- 0
- duplicates merged
How a turn works
- 1Cartesia turns speech into text and sends it to /agent/turn.
- 2Gemini turns the words into record fields (Claude if Gemini is slow).
- 3Code picks the next question. Templates write every ride sentence.
- 4The honesty guard drops any AI words that sound like a promise.
- 5Cartesia speaks the reply. The record and the map update live.
Audit trail
Nothing logged yet.