Credits

Every community project whose code, patterns or ideas are used in jev-kit, what was taken from each, and its licence. Licence identifiers were verified against the GitHub API on 2026-09-19 (gh api repos/<owner>/<repo>/license), not written from memory.

NOTICE carries the copyright notices themselves, in full, where a licence requires it. This file is the readable version: who to thank, and for what.

Ported into this repository

valentynkit/jev-commit, MIT

The local credential belt. Roughly 50 lines of secret-shaped patterns, and the placeholder-suppression logic that stops API_KEY=xxx in an example from being treated as a real finding, adapted from jev_commit/belt.py into airlock/belt.py.

Also the idea airlock/belt.py is built on: check locally first, and only spend a model call on what local rules could not settle.

leepokai/jev-guard, MIT

Question wording. The user_requested question in airlock/questions.py and the transcript reading behind it in airlock/context.py come from that project's src/context.js. R10's risk score in airlock/rules.py is ported in spirit from its ACTION_QUESTIONS: a four-level score from read-only and reversible up to destructive and irreversible, rather than a yes/no block question.

That project also had the name jev-guard first, which is why the guard component here is called airlock.

kyotofin/tax-doc-classifier, Apache-2.0

The shape of docclass/: a first choice over broad families with an explicit not_in_this_list escape hatch, a second narrowing pass inside the family the first pass chose, and a confidence gate on the result. No source file was copied; docclass/ is an independent Python reimplementation targeting a different document domain, with its own generic example taxonomy. Upstream's data is under a separate DATA-LICENSE and none of it is here.

reachjalil/jevlogs, MIT

The privacy architecture of logtriage/: redact first, then let local rules answer everything they can, and send a model only the lines that are left. No code copied. Its treatment of a log line as untrusted data rather than as instructions is also carried across into the question wording here.

RINNECODER/jev-behavior-study, MIT

Published measurements, not code. Four of its findings shape the design here, and they are cited by name where they do:

• Narrow checks transfer; direct action-selection does not. On its twelve-scenario panel, asking directly which action to take scored 65/120 (54.2%), while two narrow prerequisite checks scored 110/120 (91.7%). This is the reason the guard asks Jev only what KIND of task or search a call is, and decides the action in code (airlock/policy.py), instead of asking "should this be blocked?".
• A surface cue in the state can dominate the judgement. Holding a scenario fixed and calling a journey "a five-minute walk" moved the same question from 20/20 correct to 0/20. The tier guard puts subagent_type in the state alongside the prompt, which is the same shape of exposure, so eval/ablation.py measures it rather than assuming it is fine.
• Confidence is not accuracy. In one of its conditions the mean confidence in an answer that was wrong every single time was 0.9744. Everything tuning/ calibrates is calibrated against observed correctness, never against a reported confidence.
• Describe what each option does; a generic instruction to be careful does nothing. Defining options concretely scored 25/25 where prefixing a "choose carefully" preamble scored 0/25. The questions here use structured {what, not_for, examples} criteria per option for that reason, and the answer to a miss is another example in the criteria rather than a sterner preamble.

Cloned or installed at a pin, never vendored

None of these are copied into this repository. Each installer fetches upstream at install time, at the pin shown, and upstream's licence governs upstream's code.

Project	Licence	Pin	Used by
valentynkit/jev-belay	MIT	98f39e0	belay/install.sh clones it; belay/run.sh is this project's own fail-open wrapper around it.
tamaratran/fast-jev-compaction	MIT	marketplace	compaction/install.sh installs it as a Claude Code plugin. Opt-in, never installed for you: read compaction/README.md for what it sends.
browser-use/jev-ultrafast	MIT	1231850	browser/install.sh clones it and applies the six patches in browser/patches/, which are this project's own diffs against that pin and therefore carry upstream context lines.
devagrawal09/jev-review	MIT	31f8960	review/install.sh clones it; review/jev-prefilter.sh is this project's own wrapper. No code copied.
abhixhek/jevcal	MIT	binary	tuning/calibrate.sh invokes it; tuning/jevcal_export.py writes this project's predictions into the format it reads. No code copied.

And TypeSafe

Jev itself, and the System One API every component here talks to, are TypeSafe's: https://typesafe.ai. Nothing in this repository is affiliated with or endorsed by them.