Jev 决策流水线：置信度门控加代码执行

Archimedes @ArchimedesLore原推发布于 2026.10.27

在 X 上打开原推

WHOEVER LEAKED THIS JEV LANDSCAPE SHOULD BE OWNING THE EXECUTOR I thought the "code acts" half of an agent was just a switch statement everyone leaves to the intern. Apparently not. This landscape's downstream is much more interesting: → ACT runs the deterministic action for every answer whose confidence beat the 0.85 gate (click_next 0.95, review 0.91, phishing 0.86), no model wake, pure code → ROUTE_TO_HUMAN fires for anything under the floor and attaches the raw answer with name and value, no silent fallback, no "retry with more thinking" → LOG commits every action to disk in 3ms per entry, every row carrying the state hash it fired against, no backfill, no eventual-consistency gap → SAVE_PROGRESS persists the full world between frames with 1,008 strands indexed, so the next tick starts where this one ended, no re-read from source → GUARDRAILS cap irreversible actions at 7 per run on the Agent Drone source alone, flags 5 drafts that tried to fire without their save_progress commit so you get: Actions determined. Humans routed. Logs committed. Progress saved. the interesting part isn't making jev faster. it's giving every answer a code path that commits before the next tick fires. Instead of a model that writes JSON and prays the executor reads it right, you route 119,575 decisions through typed answers at 270ms median, let code act in 3ms per entry, and save every frame's progress against its own state hash. Act on the answer. Route the human. Save the state.

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」