Jev Decision Gates — RUN / CONFIRM / BLOCK

Jev Decision Gates — RUN / CONFIRM / BLOCK A workflow diagram generated by Archify. 01 / Inbound 02 / Jev Gate (System 1) 03 / Auto Outcome EX / Human Check EX / Refused EX / Logs & Calibration Capture Typed decision Outcome One forward pass ~420ms Rewrite + approval Prompt / Command · user or agent · Inbound › Capture Prompt / Command user or agent Context Pack · session + tool state · Inbound › Capture Context Pack session + tool state Jev Decision Call · action · radius · safe · Jev Gate (System 1) › One forward pass ~420ms › Typed decision · typed probabilities Jev Decision Call action · radius · safe typed probabilities RUN · safe ≥ 0.85, silent · Auto Outcome › Outcome RUN safe ≥ 0.85, silent Reasoning Model · System 2 on survivors · Auto Outcome › Outcome Reasoning Model System 2 on survivors Rewrite + Approval · 0.5–0.85 or low quality · Human Check › Rewrite + approval › Outcome · human gate Rewrite + Approval 0.5–0.85 or low quality human gate Rewritten Prompt · overrides original · Human Check › Rewrite + approval › Outcome Rewritten Prompt overrides original BLOCK · safe < 0.5 · Refused › Outcome BLOCK safe < 0.5 Decision Log · JSONL, every call · Logs & Calibration › Typed decision Decision Log JSONL, every call Calibration Review · after ~20 prompts · Logs & Calibration › Outcome Calibration Review after ~20 prompts state in with context log verdict high safe labeled data approved survivors only Legend Agent logic Policy Tool action Context / trace Cloud service External system

System 1 vs System 2

  • • Jev: ~420ms, ~$0.0000086/call, typed probabilities, no explanation
  • • Reasoning model: seconds, per-token cost, full chain of thought
  • • Division of labor: Jev gates volume; LLM analyzes survivors

Known failure mode

  • • Confidently wrong within schema (destructive prune scored 0.54)
  • • Band + log design keeps the miss reviewable
  • • Thresholds provisional until calibration review