jevbooks

← All projects

jev-shadow

dreyfus92/jev-shadow

Measures Jev's hazard judgments against Claude Code's classifier.

What it is

A Claude Code plugin that, in the background, asks Jev five hazard questions about every command and edit Claude proposes, logs the probabilities next to the built-in classifier's decision, and later reports agreement, over-refusals, latency, and tokens.

How it uses Jev

For each proposed Bash, Write, Edit, WebFetch, or MCP call, Jev answers five typed hazard questions (destructive, exfiltration, remote_code, weakens_security, outside_project) over redacted, clipped state. The probabilities and risk score are logged as what enforce mode would have done; in enforce mode the same path prints ask or deny before the tool runs.

Primitives:choicescore

Technique worth stealing

Shadow-mode measurement: run the model in a background hook that cannot print, log its verdicts beside the real classifier's, and join them later by tool_use_id.

Try it

claude plugin marketplace add dreyfus92/jev-shadow; claude plugin install jev-shadow@jev-shadow; jev-shadow mode shadow

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.83
  • Shows a System One patternyes0.91
  • Handles uncertaintyno0.07
  • Measuredno0.10
  • Runnableyes0.82
  • Worth recommendingunclear0.53
  • Model replicano0.04
  • Problem scopescore on a 0–2 scale1.19
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.