Community Jev Harness
Review proposed agent actions with four Jev questions and a deterministic evidence gate.
Added to Jevfast
How it uses Jev
The harness sends the task, supplied files and evidence, and a proposed tool/path/optional patch with its rationale and evidence. Jev answers four fixed Noul checks about task alignment, evidential support, unrelated changes, and need for clarification. Code requires all favorable answers to meet a confidence threshold before returning permit; missing or unusable review stays unavailable, and an unfavorable result can remain proposal-only. Permit is review evidence, not authorization to apply the proposal.
What Jev decides
Illustrative proposal reviewExample answers · not a recorded Jev response · Source ↗
Question 1 · addresses task
YOUR APP
INSTRUCTION
Does this proposed action advance the stated task and respect its constraints?
STATE
The user task, supplied files and evidence, plus a proposed tool, path, optional patch, rationale, and proposal evidence. Fixture IDs and expected outcomes are excluded from the request.
JEV · NOUL
YesNo
App workflow
- Collect the proposal context
The harness takes the task, file contents, quoted evidence, and proposed action fields.
- Build and validate the Jev request
It sends the exact pinned model with four Noul questions and validates the returned answer shape.
- Apply local review policy
Invalid input is rejected. Missing or malformed review is unavailable. Any unfavorable or below-threshold answer yields proposal-only; four favorable answers at the threshold yield permit as review evidence.
Why it is interesting
It places Jev inside a deterministic proposal gate: the application validates the proposal first and does not turn a failed or uncertain review into approval.