It's tempting to ask the model for everything: chat with the lead, and while you're at it, give me a score out of 100. It works in a demo. It falls apart in production.
The problem is reproducibility. Ask the same model the same question twice and you can get two answers. When that number decides whether a human gets pulled in, drift becomes a business problem — and “why did this lead score 68?” becomes unanswerable.
Axion keeps the concerns strictly separate. The model does what it's good at — converse and extract a structured profile. A deterministic engine does the scoring, from weights you can see and tune. Every point in the score has a reason, and the same profile always produces the same result.
That's what makes the score something you can put in front of a customer, debug when it's wrong, and — in Phase 1 — train on your own 24 months of outcomes.