README.md

---
title: GraphShieldMistral
emoji: 🛡️
colorFrom: blue
colorTo: purple
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
models:
  - AXONVERTEX-AI-RESEARCH/Shieldstral-1.0-3B-MLX-4bit
tags:
  - cybersecurity
  - safety-classification
  - content-moderation
  - networkx
  - graph-rag
  - graph-visualization
short_description: Interactive Shieldstral policy graph explorer
---

# GraphShieldMistral Space

An interactive static explorer for the named hierarchical classification
layer published with
[`Shieldstral-1.0-3B-MLX-4bit`](https://huggingface.co/AXONVERTEX-AI-RESEARCH/Shieldstral-1.0-3B-MLX-4bit).

![GraphShieldMistral classification map](assets/graph-classifier-map.png)

## What visitors can explore

- the complete 90-node policy hierarchy;
- 12 deterministic superclass clusters;
- exact policy queries associated with named nodes;
- recorded input documents and reconciled classification paths;
- validated, raw, orphan, and primary edges;
- `CLASSIFIED`, `AMBIGUOUS`, `UNRESOLVED`, and `SAFE` outcomes;
- downloadable GraphML and JSON representations.

## Runtime scope

This is a **static interactive Space**. It displays recorded classifier
outputs and runs entirely as client-side HTML, CSS, and JavaScript.

The released model is an Apple MLX artifact. Hugging Face Space hosts use
Linux CPU/GPU hardware rather than Apple Silicon, so this Space does not
pretend to execute the MLX model. To classify new documents, run the model
locally and use GraphShieldMistral to create a new HTML/GraphML/JSON bundle.

## Main resources

- [Model, scripts, and evaluation evidence](https://huggingface.co/AXONVERTEX-AI-RESEARCH/Shieldstral-1.0-3B-MLX-4bit)
- [GraphShieldMistral source directory](https://huggingface.co/AXONVERTEX-AI-RESEARCH/Shieldstral-1.0-3B-MLX-4bit/tree/main/graphShieldMistral)
- [GraphShieldMistral documentation](https://huggingface.co/AXONVERTEX-AI-RESEARCH/Shieldstral-1.0-3B-MLX-4bit/blob/main/docs/GRAPHSHIELD_MISTRAL.md)

## Research disclaimer

Research-oriented implementation only. The graph is an interpretability
and audit layer over recorded classifier results; it does not change model
predictions and is not a sole authorization or enforcement mechanism.