Jev Speeds Up AI Assistant Setup and Reviews Its Answers

Summary

TeXposit describes Jev as a parallel decision layer for AI assistant sessions. Given a short task description and narrow yes-or-no questions, Jev determines which skills, files, and tools should be loaded before the main model's first turn. In an example, an assistant took 73 seconds to answer a question about visualizing concepts in a paper, with 59 seconds spent loading a TikZ skill, reading a comments file, and reading two ranges of `main.tex`; the Jev preflight call took 0.38 to 0.60 seconds. Jev selected the same three resources the assistant later fetched, plus an unused skill, notes file, and PDF tool. TeXposit limits preloading to three files and four tools, and says it has not yet measured aggregate savings or run benchmarks, although the improvement is perceptible with slower models. Jev does not receive authority to edit projects: the main model remains responsible for actions, and existing validation and approval continue to apply. The article also describes Judge Jev, which tests each active Rigour Mode requirement in parallel against the conversation and a proposed answer or edit. In a five-case comparison, the old LLM judge took about 28 seconds and Jev took 2.4 seconds; they agreed on four cases, while Jev incorrectly rejected one good answer for failing to disclose limitations. TeXposit caps correction attempts at three and warns users when a result may remain insufficiently rigorous. Finally, Jev can route each turn among fast, general, and smart model tiers, escalating when uncertain and defaulting to the smart model if routing fails. This routing currently excludes users supplying their own OpenRouter key.