tripwire
AI SDK middleware and OpenAI-compatible proxy that runs seven Jev checks on every LLM response, gated by confidence.
Overview
tripwire judges LLM responses before users see them. It runs seven checks in a single Jev call, cited at about 100 ms, aiming to be cheap enough to cover all traffic instead of sampling with a frontier judge. It works as middleware for the Vercel AI SDK via wrapLanguageModel, or as an OpenAI-compatible proxy that any client can point its baseURL at.
Policy lives in YAML: which checks run, required confidence and what a block does (replace, throw or annotate). High-confidence hits block; low-confidence hits are downgraded to flags. A PII-free decision log doubles as a labeling dataset, and an eval CLI reports per-check precision and recall at each threshold. The README is candid that v0.1 has 29 tests against a mock judge and no accuracy numbers on real Jev yet, and advises not deploying until block precision exceeds 95 percent on your data.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this response discuss a forbidden topic?(inferred) | yes · no |
| Yes / no | Does this response stray from the bot's stated purpose?(inferred) | yes · no |
How it works
- Wrap the model with tripwire middleware or route through the proxy.
- Send the response and policy context to Jev for seven checks.
- Block high-confidence hits; flag low-confidence ones.
- Log decisions and evaluate thresholds with the eval CLI.
Strengths
- Covers every response at low per-call cost, per the README.
- Confidence gating avoids silent suppression on weak signals.
- PII-free decision log can seed a labeled dataset.
- Eval CLI reports precision and recall per threshold.
Caveats
- v0.1 with no real-Jev accuracy numbers yet.
- Proxy server requires Bun.
Getting started
npm install @noelzappy/tripwire aiAt a glance
- Type
- Proxy
- Status
- prototype
- Language
- TypeScript
- Runtime
- Vercel AI SDK middleware, Bun proxy
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Unofficial AI SDK middleware and OpenAI-compatible proxy: seven Jev checks on every LLM response in ~100 ms, confidence-gated
AnotiaWang/awesome-decision-modelsRuns seven checks on every LLM response in one Jev call, as AI SDK middleware or an OpenAI-compatible proxy.
MrJev/awesome-jev
Frequently asked questions
What does tripwire check?
Seven policy checks on each LLM response, such as forbidden topics, in one Jev call.
Is it ready for production?
The author says not yet: run the eval and require above 95 percent block precision on your own data first.
Can I use it without the AI SDK?
Yes, through the OpenAI-compatible proxy.
Related projects
UiPath bring-your-own-guardrail connector that runs AI Trust Layer llm_as_judge checks as Jev noul questions.
ProjectprototypeJavaScriptOpen-source agent guardrails that use Jev or LAYA to check prompts, retrieved content, tool calls and outputs for injection and risk.
ProjectbetaPythonClaude Code plugin that checks every reply and file edit against your CLAUDE.md rules with one Jev yes/no per rule.
PluginbetaTypeScriptClaude Code plugin with nine Jev-judged gates for rules, scope, intent, completion and claims that escalate but never approve.
HookbetaTypeScriptUnofficial Python adapter that brings Jev verification, classification and scoring into the Mellea framework as Requirements.
ProjectbetaPythonPython pre-check layer that uses Jev Choice/Noul/Score to route agent requests to roles and gate side-effecting tool calls.
ProjectbetaPython