What happens if I deploy with no API keys set?

It falls back to the local heuristic engine automatically and reports heuristic fallback status, so check engine and isHeuristicFallback fields.

Will angry but legitimate prompts get blocked?

Risk bands map 0-9 Safe to ALLOW through 80-100 Critical to BLOCK, with MONITOR and INSPECT in between; test legitimate queries with one-click presets and history compare.

How do intent and pattern scores combine?

The response returns both intentScore with summary and patternScore with summary plus an overall score, level, categories, and action for audit.

Can I trust Gemini fallback to match Jev calibration?

Treat them as tiers, not equivalents: Jev is primary via @typesafe-ai/sdk with probabilities and confidence, Gemini is secondary, heuristics last; /api/health shows the active engine.

What risk levels and actions does it define?

Critical at 80-100 percent triggers BLOCK, High at 60-79 triggers QUARANTINE, Medium at 30-59 triggers INSPECT, Low at 10-29 triggers MONITOR, and Safe at 0-9 triggers ALLOW.

Which primitives does it use for what?

Choice for the five-level risk classification, Score over 0-9 ordered levels for weighted risk, intent, and pattern scoring, and Noul for per-vector booleans such as instruction override and misinformation.

How many engines are in the fallback chain?

Three: TypeSafe JEV as primary, Google Gemini as secondary, and a local heuristic engine as the tertiary option that runs fully offline when neither key is set.

CommunityJev Tools & CLIs

JEV-Prompt-Injection-Guardian

TypeScript dashboard that scores prompts with Jev Choice, Score, and Noul, maps scores to block-through-allow actions, and falls back to Gemini or offline heuristics.

See highlights

Open github.com

Stars: 0
Language: TypeScript
Published by: Community
Author: YuyaForest
Last verified: Oct 2026

YuyaForest/JEV-Prompt-Injection-Guardian

User repository on GitHub

View on GitHub

JEV Prompt Injection Guardian is a prompt injection quarantine and risk-scoring system for LLMs powered by JEV (jev-1.13.0), TypeSafe AI's innovative "System One" model, as its primary engine. By eliminating ambiguous, long-winded natural language explanations, it instantly computes rigorously calibrated probability risk-scores.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity github

Works forJev for Scoring

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Choice gives Critical to Safe, Score averages 0-9 risk levels, Noul rates leakage, bypass, override, misinformation.
Dual-spectrum view separates semantic intent from jailbreak syntax, tag spoofing, and obfuscation.
Three engines in order: Jev with TypeSafe key, Gemini with its key, then local heuristic offline.
Express /api/evaluate returns score, level, threat categories, action, and per-engine detail.

Watch out

Needs API keys for hosted engines or it uses local heuristics; presets and history are for testing, not proof of safety. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

tictacjev

Tic-tac-toe demo where Jev plays one side through a Choice question over open cells, returning a chosen cell, confidence, and per-cell probabilities shown in an expandable move log.