jev
Delegate a repeated decision to Jev, TypeSafe AI's System One model, instead of deciding every item yourself or spending a Claude agent on it. Use when the same yes/no (or choice, or score) judgment i
SKILL.md.claude/skills/jev/SKILL.md
GitHub/jev: delegating decisions to Jev
Jev does not write text. It takes a state (a JSON object) and a typed question, and returns a
calibrated probability that code can branch on: a smart if for conditions code cannot compute.
It answers in well under a second, for a fraction of a cent (vendor figures).
Does the job fit? All five must hold
- Repeated. The same question is asked of many items: dozens now, or every session.
- Closed. The answers are known up front (yes/no), and the action for each is known.
- It fits in a state. An item is a small JSON object, such as an issue's title, body and labels.
- Cheap to be wrong at the ends. A wrong answer at a confident end costs less than you deciding every item, and a person can still overturn it.
- Cleared to leave. The content may be sent to a third-party API. Ask the user before sending anything from a private repository for the first time.
Not Jev's:
- code (OpenCode);
- reviews and anything that needs a reason (Claude);
- anything a script can compute (file lists, the fix-lane test);
- a one-off decision.
The procedure
Define. Add the decision to
harness.jsonunderjev.decisions:{"type": "noul", "instructionsFile": "docs/jev/<name>.md"}. The instructions say:- what the state is;
- what yes means, and what no means;
- the confusing cases, and which side they fall on.
docs/jev/breaks-play.mdis an example.Label. Collect at least 60 items with the right answer, as JSONL lines of
{"id", "state", "label": true|false}.- The best labels are decisions people already made: issue labels, triage outcomes, merged or closed.
- Labels you make yourself measure Jev against you, not against the truth. Use them only if the user agrees, and say so in the record.
Trial. Run
node tools/harness/jev.mjs trial --decision <name> --in labels.jsonl. It picks each end's cutoff on half the labels and reports it on the other half.- Adopt the cutoffs only if the held-out precision holds and there are no warnings.
- Record them in
harness.json:autoYesAt,autoNoBelow, and atrialobject with the date, the model, the held-out counts, precision and coverage. - Commit with the numbers in the message.
- An end without a cutoff stays yours.
Route. Build
items.jsonl(for issues:gh issue list ... --json number,title,body,labels --jq '.[] | {id: .number, state: {title, body, labels: [.labels[].name]}}'), then runnode tools/harness/jev.mjs route --decision <name> --in items.jsonl --out routed.jsonl.- Act on the
yesandnorows, and say on each item that Jev decided it and with whatp. - Decide the
middlerows yourself.
- Act on the
Keep it honest.
- A person's decision always outranks Jev's.
- When the user overturns a Jev answer, add the item to the labels.
- Re-run the trial when the model pin or the instructions change. The cache is keyed on both, so a rerun asks again.
Exits
- 0: done.
- 1: refused (an untrialled decision), or stopped by a provider failure. The answers so far were written and cached; decide the rest yourself, or rerun later.
- 3: Jev is off (no key, or no
jevsection). Decide every item yourself, as before.
ask returns raw answers for any question type. trial and route handle noul only, until the
response formats of Choice and Score are confirmed.