Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

paved-gate

System 1 ingestion gate that asks Jev intent, risk and sensitivity in one call, then blocks, answers deterministically or passes to an LLM.

Projectprototype
View on GitHub ↗

Overview

Paved Gate puts a typed decision model in front of AI agents to judge every inbound request quickly instead of using a frontier LLM. It detects and masks PII, then asks three questions in one parallel round trip targeting about 100 ms: intent (Choice over three routes), risk (Score on a 1-5 rubric) and sensitivity (Noul for PII or PHI). Based on policy, it blocks the request with a policy error, answers it with a deterministic handler with no LLM call, or passes it to a frontier agent such as Claude, OpenAI or your own orchestrator.

Every decision writes a structured JSONL audit record including raw scores and the hash of the policy that produced it. Policies are code in a config file, there is a FastEvaluator interface, FastAPI/Starlette middleware, and an NVIDIA OpenShell integration. The README states it is alpha; the Jev adapter ships with a contract-accurate mock and was run against the live API for 144 requests across four benchmark runs without errors.

What the decision model decides

TypeQuestionAnswers
ChoiceWhat is the intent of this request (which of three routes)?—
ScoreHow risky is this request?1 · 2 · 3 · 4 · 5
Yes / noDoes it contain sensitive data such as PII or PHI?yes · no

How it works

  1. Detect and mask PII in the request.
  2. Ask intent, risk and sensitivity in one Jev call.
  3. Apply policy: block, answer deterministically or pass to an LLM.
  4. Write an audit record with scores and policy hash.

Strengths

  • Avoids LLM calls for blocked or simple requests.
  • Full audit trail tied to policy versions.
  • Mock adapter for testing.

Caveats

  • Alpha status.
  • Masking before judgment can hide information the evaluator needs (discussed in README).

At a glance

Type
Project
Status
prototype
Language
Python
Runtime
FastAPI/Starlette middleware
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What three questions are asked?

Intent (choice), risk (1-5 score) and sensitivity (yes/no).

Can it avoid calling an LLM?

Yes, blocked requests and those answerable by deterministic handlers never reach an LLM.

Is it tested against real Jev?

Yes, 144 live requests across four benchmark runs, per the README.

Related projects

  • claude-code-templates

    Large catalog of Claude Code agents, commands, hooks and MCPs, listed with community Jev mods for guardrails and subagent models.

    CLIproductionJavaScript
  • guardrail-chatbot-jev

    Python/TypeScript/Go chatbot safety library checking user input and bot replies with Jev and returning one actionable verdict per message.

    LibraryproductionPython
  • jev-decisionops

    Offline-by-default demo of six Jev decision workflows: model tiering, prompt screening, tool-call risk, reranking and more.

    ProjectprototypePython
  • Pi Jev Guard

    Pi guard plugin with timing-based rules, Jev risk checks on tool calls, output redaction and reminders for repeated failures.

    PluginprototypeTypeScript
  • jev-route

    Router that uses Jev's cloud decisions today, logs every decision with probabilities, and distils a local routing model from your traffic.

    ProjectbetaPython
  • Jev demos

    Six side-by-side demos where Jev makes the decision, Claude or Kimi writes the words, and code owns control flow.

    AppprototypeTypeScript