The engine, not a prompt

A VARTA call moves through seven explicit layers -- L1 through L7 -- each with a narrow, inspectable job, instead of one growing prompt handed to a model every turn. Call flows are authored as deterministic step graphs: the same input takes the same path, every time, which is what makes a call flow testable at all.

Because most of what a caller says has been said before, VARTA plays cached audio for repeated phrasing instead of re-synthesising it, and only escalates to an LLM call for the turns that genuinely require reasoning. That is the architecture the cost calculator models.

This isn't a lab demo of the idea: humanise.ai is a live, multi-tenant SaaS product running on this exact core today, in production, for real calls.