the assistant that never invents a fact

A business wants a model to answer its customers. The model will, fluently, and sooner or later it will state a price, an opening hour or an appointment time that does not exist. For a business that is not a bug; it is a refund and a lost customer.

We never own the schedule and never invent a fact. The business's calendar is the only source of truth; the model may only state prices, hours, services and times that exist in the business's data or came back from a tool call. Every failure ends with a human. No path ends in silence, a wrong booking or a made-up answer; the worst case is "let me check with the owner and text you back", and the owner is told.

• Prompts are assembled at request time from the business's data, never hand-edited.
• Every number in a drafted message that looks like a price or a time is compared with the tool results and the business's data; a mismatch blocks the message and substitutes a safe one, and the owner gets an alert.
• Appointment times are never composed by the model; the scheduling code offers real slots and words them itself.
• A short model timeout with one retry, then the safe message; a stalled model never blocks a customer.
• Kill switches per business and globally: voice, booking, texting, the model itself.

Some fluency. The assistant says "let me check" more often than a free-running model would. That is the correct trade for a business, and the owner can see every held message and write the answer once, which is how the assistant learns.