The Gap: Facts Need a Deterministic Judge

Every assertion in the system today is an LLM reading a transcript and returning a probabilistic verdict. That is the right tool for a genuinely fuzzy question (was the explanation clear, was the tone right, did it handle the objection well) and the wrong tool for a question with a definite answer. Two classes fall outside what it can do: outcomes that are not in the transcript at all (did it transfer to the correct destination, did the appointment actually get created, did the CRM write land), and transcript facts that are exactly checkable (did it say the required disclosure verbatim, did it read back the callback number, did it use a forbidden phrase), where an LLM answers probabilistically and at inference cost what a plain match settles exactly. The failure that makes this concrete: an agent says "you're all booked for Tuesday at 2" while the CRM write silently failed. Hamming hears a confident, correct-sounding confirmation and passes the call; the customer never gets a technician. Only a check against the actual record catches it, and a transcript judge by construction cannot be that check.