Question 1

Why add a decision API when chat review already exists?

Accepted Answer

The System One path turns review into one POST to baseURL systemone with parallel decision and risk choice questions, deterministically mapped to the same verdict JSON as chat. It suits teams already serving Jev, SiliconFlow, or Laya.

Question 2

Will defer mode flood me with manual approvals?

Accepted Answer

Partly. With defer, low-confidence verdicts fall back to human approval and are audited as reviewer-low-confidence. To reduce prompts, broaden allowRules for narrow repeated tool plus operation prefixes or adjust confidenceThreshold.

Question 3

What data leaves my machine for a review?

Accepted Answer

The pending tool call, sandbox state, reviewer-requested read-only facts, and a bounded context of 10 messages or 6000 chars. File contents stay off by default, and keys are redacted from requests, audit rows, and traces.

Question 4

What happens if the reviewer times out or crashes?

Accepted Answer

It fails closed: denyOnReviewerError defaults true so crashes or timeouts deny the request, the circuit breaker blocks repeated probing, and critical-risk findings always deny regardless of decision answers.

Question 5

How does it turn a probability into a confidence?

Accepted Answer

It computes the confidence as the absolute distance of the allow probability from one half, doubled, so a p(allow) of 0.500 yields a confidence of 0.000. That is deliberate: a coin-flip decision is treated as cannot judge rather than quietly allowed, and the value then feeds the onLowConfidence policy.

Question 6

Why does Jev so often return low-confidence verdicts here?

Accepted Answer

The README reports a calibration finding from real use: on escalation-type requests Jev often deliberately abstains, giving confidence near 0.000. That is not an error, it means the model declines to back either answer, so onLowConfidence decides whether it becomes a deny or a defer to a human, and it is why broad allowRules coverage matters.

Question 7

What is the fact-finding step the reviewer can use?

Accepted Answer

Before deciding, the reviewer can call bounded read-only fact tools: by default up to two rounds and three facts. Reading file contents is a separate switch that is off by default and capped at 4096 bytes when enabled. That is the seam that lets the reviewer see sandbox state without handing it the whole transcript.

Dsh Auto Review Jev

Highlights

Questions about this entry

More Jev resources like this

From the Jev community