We just swapped a chunk of freeform LLM reasoning in our lead classification pipeline for Jev, and it's a clean pattern worth sharing.

We just swapped a chunk of freeform LLM reasoning in our lead classification pipeline for Jev, and it's a clean pattern worth sharing. The old setup: a lead comes in with a raw signal (a hiring post, a funding raise, a shipping stall), and a subagent read a rubric and reasoned in prose to land on a fit band: vertical match, signal strength, geography fit, stage/budget fit, offer fit, then High/Medium/Low. It worked, but every lead cost a full agentic pass, and the reasoning was a paragraph you had to re-read to trust. Now the five dimensions go out as one Jev call: a Choice for vertical match, a Noul for "is this signal actually fresh and specific," a Choice for geography fit, a Score for stage/budget fit, a Choice for offer fit. They run in parallel over the same state and come back typed, with probabilities: "vertical_match": {"choice": "in_vertical", "confidence": 0.94} "signal_strength": {"noul": 0.93} "offer_fit": {"choice": "ai-demo-to-prod", "confidence": 1.0} Code still owns the banding logic (weights, dealbreakers), and low-confidence answers fall back to a human read. T