Question 1

Are the token savings real, or was the baseline stacked?

Accepted Answer

The author argues the control was not weak, scoring 93.4% to 98.2% across repeated runs, and states the honest claim is parity rather than superiority — through jev-mode scored 96.1% against 93.7% on department routing. The corpus was synthetic, and the same write-up reports an earlier question design that lost 5.3 points, which is the more useful warning: measure your own accuracy before trusting the ratio.

Question 2

Does the gate actually catch every step an agent takes?

Accepted Answer

No, and the project is built around measuring that gap rather than claiming coverage. Subagent sessions receive no pre-action hooks — 339 subagent sessions, 29,146 tool calls, zero hook invocations — so the gate protects the parent only. Run jev-mode coverage to see per-session and per-subagent numbers, and note the harness refuses a spawn unless the delegation message commits the child to per-step Jev use, since the spawn is the only enforcement point available.

Question 3

Why isn't this on PyPI, and what do I need installed?

Accepted Answer

Install from the git URL with pip, or clone and run PYTHONPATH=src python3 -m jev_mode.cli check. There are no dependencies either way — Python 3.8 and the standard library. The TypeSafe key comes from TYPESAFE_API_KEY or ~/.config/jev-mode/typesafe.env, and you can point TYPESAFE_ENV_FILE at another file instead. jev-mode check verifies the config and reaches the API.

Question 4

Can the gate block something important by mistake?

Accepted Answer

It is deliberately narrow. It only blocks repeats and off-goal steps above 0.8 confidence; the unsafe_or_irreversible answer is recorded but never enforced, because a typed model is not a policy engine. It has a denial budget of three consecutive refusals after which it fails open so an agent cannot be trapped, and any allowed step clears the counter. It is off until you enable it and only gates state-changing tools, not every read.

Question 5

What did the measurement actually show?

Accepted Answer

Over 1000 triage judgments against a Jev-free control: total tokens fell from 7,987,983 to 1,770,917, work-attributable input from 3,539,466 to 220,879, and model round trips from 40 to 14. Department accuracy went from 93.70% to 96.10% and the boolean field stayed at 100%. Jev's own share of that run was about four cents.

Question 6

When do I batch and when do I fan out?

Accepted Answer

The two shapes have opposite answers. Many items is one call per item driven from code: jev-mode batch --items items.jsonl --questions questions.json --out answers.jsonl --pool 6. One item with many questions is a single call: jev-mode ask --state-file one.txt --questions-file questions.json. Answers stream out as they complete, so a failed item is isolated rather than fatal and retries are counted per item.

Question 7

What does one gate call cost against the step it guards?

Accepted Answer

One Jev call per gated step, roughly two hundredths of a cent, measured against a per-step agent context of about 109k tokens. A refusal comes back as a PreToolUse deny carrying the reason, so the agent is told why a step was refused rather than simply losing it. The rules table records what each question type returns: choice, score and noul answers with their confidence.

Jev Mode

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community