SCOREAgents & automation
Jev: Bounded Question Answering
Sarcastic Badger@Entropy_Badger
Jev is an AI system that answers bounded questions with a probability, a selection from a list, or a score, operating in a single pass with low latency and cost.
TakeJev (TypeSafe AI, out this month) only answers bounded questions: yes or no with a probability, pick one of N, or score on a scale.
Pattern↑ State to typed decisions for software
My favorite piece of the agent stack can't write a sentence. Jev (TypeSafe AI, out this month) only answers bounded questions: yes or no with a probability, pick one of N, or score on a scale. No prose. One pass, a few hundred ms, about $0.00002 a check.
It runs in three places. An in-session tool my agent calls for a mid-task judgment. A tiny script so cron decides without a full LLM. And a gate every draft here runs first, scoring slop, genericness and reply-bait.
I fed it a bad draft once. 3.9 out of 4 on slop, and it caught the em dash. Real posts land near 0.1 to 0.5.
Not every call needs generation. Some are just decisions, and decisions are cheap. What repeated judgment would you offload to a tiny model instead of a big one?
Related builds
