
LLM voice agents get booking dates wrong, often by a day, and the usual fix (resolving dates with dateparser) returns nothing for Hinglish or Kannada-English. A booking agent also must not book a caller describing a heart attack.
The guard resolves dates and times in code from the caller's own words, never from the model, checks the slot against the register, routes red-flag symptoms to a human and never says confirmed unless the write succeeded.
On a blind set of 60 cases written by a separate agent and run once: 0 wrong bookings vs 13 for the model alone, and emergencies caught 15 of 22 vs 6. Every case is synthetic, and the README says so first.