Jev 决策流水线：状态打包、问题堆叠、门控路由

Archimedes @ArchimedesLore原推发布于 2026.10.28

在 X 上打开原推

WHOEVER LEAKED THIS JEV DECISION PIPELINE SHOULD BE WRITING THE RFC I thought routing an agent through five steps meant five prompts, five parsers, and five places something could silently drift. Apparently not. This pipeline's setup is much more interesting: → STATE packs the inbound case into typed cards indexed across 12 use cases (email, meetings, contracts, leads, trading), no raw text passes the first column → TYPED QUESTIONS stacks 7 asks against the same state (Choice folder, Noul worth_reply, Score urgency, Choice action, Score viral_fit, Score lead_quality, Choice next_worker), one shot, no retries → JEV SYSTEM ONE runs choice, score and noul heads in parallel with no generation at all, 256ms median, tokens out literally zero → GATE routes each answer at 0.85, below drops to quarantine or review, above forwards to the next worker (draft_short, book_call, type_destination), no silent escalation → CODE acts on the typed label deterministically (mail moved, calendar link sent, slack ping, review folder), flags 6 routes drifting from the baseline this run so you get: States packed. Questions stacked. Gates routed. Acts deterministic. the interesting part isn't adding more steps. it's giving every column one job the next column can't do for it. Instead of one model that writes, decides and acts in one freeform turn, you split the pipeline: LLM writes, JEV decides, code acts, with 10,000 decisions at $0.42 and gate 0.85 holding every route. Pack the state. Stack the questions. Act in code.

这条是引用着另一条写的：「长文：如何搭好 Jev 驱动的 agent 技术栈」