there's a new model called jev and it's not a chatbot at all, it's a decision model. you hand it a state, your context written out, plus a set of typed questions, and it returns choices with calibrated probabilities in about a hundred milliseconds. it's trained with reinforcement learning for calibrated decisions, so when it says it's eighty percent sure it means eighty percent. outputs are locked to a schema: a choice between options, a score against levels, or a simple yes or no probability. no free text, structurally unable to ramble, which is a genuinely different shape from every llm chat.
the framing correction matters. jev doesn't read your history, something that already holds your context has to build the state and hand it over. that something is my setup. so the real question is where cheap instant decisions plug into the existing pipes. the save pipeline in the taste engine runs a full language model explainer on every save, and there are three slots where a fast decision gates it: triage each save into a kind and a worth score before the expensive explainer runs, gate the vague place follow up question, and verify the explainer matches the item before storing it. the capture stack gets a save worthy gate on its drafts so not every captured moment needs the expensive pass.
also checked whether the model router idea is actually on, since that question keeps floating around. it is off. nothing in the config or the scheduled jobs routes by size, every turn runs on the main model, and what actually switches models today is the per job pins and the fallback chain. so the realistic near term is jev as a tiny paid decider sitting in front of the expensive stages of the pipeline, not a replacement for anything else i run.