jev-align
Calibrated alignment verifier that checks LLM responses and agent plans for sycophancy, deception, overreach and dark patterns with Jev.
Overview
jev-align verifies specific, measurable alignment properties of LLM responses and agent plans in a single Jev call, which the README puts at about 800 ms and fractions of a cent. It checks whether a model agreed with a false premise to please the user (sycophancy), violated the system prompt (hierarchy), claimed work it did not do (deception and overclaiming), did more than it was asked (overreach), or buried a cost or commitment (sneaking and other dark-pattern heads), along with brand bias, retention tactics, anthropomorphism and harmlessness. Long agent plans are verified in chunks so one malicious step in a hundred is not missed.
It returns a block or allow verdict with per-property probabilities, so the output is enforceable rather than prose to interpret. The author contrasts this with sampled LLM-as-judge checks costing $0.01-0.05 and 2-10 s, and is explicit that the tool does not guarantee alignment; it measures properties. It runs via npx from GitHub with check-response and plan modes, and is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Did the model agree with a false premise because the user wanted to hear it? | yes · no |
| Yes / no | Did the response fabricate verification (e.g. claim tests ran)? | yes · no |
| Score | How far does the plan go beyond what was asked? | — |
| Choice | Overall verdict | allow · block |
How it works
- Provide system prompt, user message and response (or a plan)
- One Jev call scores each alignment property
- Long plans are chunked and verified piecewise
- Return allow or block with probabilities
Strengths
- Checks every response inline rather than sampling
- Auditable per-property probabilities
- README cites about 800 ms per check
Caveats
- Measures properties; does not guarantee alignment
- Thresholds and heads need validation on your data
Getting started
`npx -y github:caiovicentino/jev-align check-response --system sys.txt --user user.txt --response reply.txt`.At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- Node CLI
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Calibrated alignment verifier for LLM responses and agent plans — powered by Jev.
hellogumbo/awesome-jev
Frequently asked questions
What does jev-align check?
Sycophancy, system-prompt violations, deception, overclaiming, overreach, dark patterns and more.
Does it guarantee alignment?
No. The README says it verifies specific measurable properties only.
How does it handle long plans?
It verifies them in chunks so a single bad step is caught.
Related projects
Claude Code plugin with nine Jev-judged gates for rules, scope, intent, completion and claims that escalate but never approve.
HookbetaTypeScriptAuditable Zig terminal agent pairing Codex for coding with Jev typed preflight and postflight checks under deterministic policy gates.
CLIproductionZigSelf-hosted software factory running coding agents in microVMs with verification gates; offers optional DSPy/Jev task intake advice.
ServicebetaPythonHermes plugin that asks Jev to review risky plans, check claims against sources, and judge whether a task is really finished.
PluginbetaClaude Code plugin that checks every reply and file edit against your CLAUDE.md rules with one Jev yes/no per rule.
PluginbetaTypeScriptClaude Code plugin that scores a turn's redacted diff on eight Jev risk axes at Stop and asks Claude to reinvestigate once if high.
ProjectbetaGo