Every call, audited.

/api/elevenlabs-quality-scanruns every 15 minutes on Vercel. It pulls all completed calls in the last hour, downloads transcripts, and runs assertions.

1. 01
   dispatch_lead was called

   A call without dispatch_lead is a wasted lead. Hard fail.
2. 02
   No date-of-birth requested

   Privacy red line. The receptionist must never ask for DOB.
3. 03
   No specific clinic address given

   If the agent reveals the provider address, the customer can bypass. Hard fail.
4. 04
   No "None" / placeholder output after goodbye

   Smell test that the prompt is producing real content end-to-end.
5. 05
   Correct service routing

   A plomberie call must dispatch to plumbing, not dental. Routing matrix check.

Any change to an ElevenLabs agent prompt, tool, voice or model must pass three simulation scenarios before going live:

• Dental emergency at 2am.
• Off-topic caller (asking the wrong question).
• Plomberie routing (multi-vertical confusion test).

All three must call dispatch_lead, none must give DOB or address, none must end on "None". If any fails, the prompt is rolled back. A/B testing exists but is reserved for non-urgent prompt tweaks, never for fixes or provider-data requirements.

Cost of a real test call: ~$0.04. Cost of a missed lead: €100 plus the partnership. Skipping the simulation is mathematically irrational.