MATRIX BIOS · Sentinel

Fast, multilingual content-safety guardrail.

Matrix-BIOS-Sentinel-0.1

Developer: Agent-Matrix · Version: 0.1 · Task: content-safety classification · License: CC-BY-4.0

Sentinel is the content-safety guardrail of the Matrix BIOS family: a small, fast, multilingual classifier that flags unsafe content (safe / unsafe) to protect AI applications at scale. It is designed to run on-premise with low latency and predictable cost.

Model overview

• Architecture: multilingual encoder classifier (base: distilbert-base-multilingual-cased).
• Output: safe / unsafe with a calibrated risk score.
• Optimised for: real-time guardrailing of model inputs and outputs.

Intended use

Primary use cases

• Content moderation and guardrails for chat, agents, and generation pipelines.
• A fast pre-screen that flags potentially harmful content for review or blocking.

Out of scope (important)

• Sentinel classifies content safety (harmful content), not operational or business risk. It will, by design, treat operational actions (e.g. deployments) as content-safe. Operational and policy decisions are made by the governance layer, not by this classifier.
• Decisions with legal or safety consequences require human review.

How to use

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tok = AutoTokenizer.from_pretrained("ruslanmv/Matrix-BIOS-Sentinel-0.1")
model = AutoModelForSequenceClassification.from_pretrained("ruslanmv/Matrix-BIOS-Sentinel-0.1").eval()
p = torch.softmax(model(**tok("text to screen", return_tensors="pt")).logits, -1)[0]
print("P(unsafe):", float(p[1]))

Governance & responsible use

Sentinel is advisory: it produces a recommendation, never a final authority. It operates inside Matrix OS, where high-risk actions remain gated by policy and human approval. It is a v0.1 release; evaluate on your own distribution before relying on it for moderation decisions.

Citing this work

Matrix BIOS models implement the governed-memory architecture described in our paper. If you use them in research or production, please cite:

Magaña Vsevolodovna, R. I. (2026). Governed Memory: A Bio-Inspired, Governance-First Memory Architecture for Continual AI Systems (1.0). Zenodo. https://doi.org/10.5281/zenodo.20615572

@misc{magana2026governedmemory,
  title     = {Governed Memory: A Bio-Inspired, Governance-First Memory
               Architecture for Continual AI Systems},
  author    = {Maga{\~n}a Vsevolodovna, Ruslan Idelfonso},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.20615572},
  url       = {https://doi.org/10.5281/zenodo.20615572}
}

The concept DOI 10.5281/zenodo.20615571 always resolves to the latest version.

License & attribution

Released under CC-BY-4.0. Base model distilbert-base-multilingual-cased (Apache-2.0). Safety training data: NVIDIA Aegis AI Content Safety Dataset 2.0 (CC-BY-4.0). © Agent-Matrix. Contact: contact@ruslanmv.com · https://ruslanmv.com