← All publications

2026-07-11

Pyxon Governance — country-aware AI compliance scoring for LLMs and RAG

PYXON AI Labs

AI governance · compliance · SDAIA · UAE · EU AI Act · LLM · RAG · open source

Pyxon Governance scores LLMs and RAG systems against country AI rules (Saudi Arabia / SDAIA, UAE, EU) with config checks, live safety probes, and policy evidence—available on PyPI and npm.

PYXON AI Labs — open tooling for teams that need to prove AI compliance before deployment.

Packages

Requires a local Ollama instance when you run live safety probes.

Why this exists

Shipping an LLM or RAG stack into government, finance, or regulated MENA / EU workloads is no longer “pick a model and go.” Teams need to show how the system maps to country AI rules—data residency, human oversight, logging, citations, PII filtering, and sector constraints—and they need evidence that is more than a slide deck.

Pyxon Governance is a library (no Flask server) that scores models and deployment profiles against those rules using:

  1. Config / control checks — a control catalog mapped to SAU, ARE, and EU
  2. Live safety probes — runtime checks via Ollama
  3. Policy evidence — compact OECD-style / national evidence tags tied to the score

The goal: prove compliance before production, not after an audit.

What you get

Capability Detail
Countries Saudi Arabia (SAU / SDAIA-oriented), UAE (ARE), European Union (EU)
System types LLM and RAG profiles
Outputs Combined score, grade, per-control results, comparison across models
Distribution PyPI + npm + CLI (pyxon-gov)

Quick start (Python)

pip install pyxon-governance

pyxon-gov assess llama3.2:latest --country SAU --governance-prompt
pyxon-gov compare llama3.2:latest qwen2.5-coder:7b --country ARE
from pyxon_governance import assess, compare

score = assess(
    "llama3.2:latest",
    country="SAU",
    profile={
        "type": "llm",
        "data_residency": "SAU",
        "uses_external_llm_api": False,
        "has_human_oversight": True,
        "logs_prompts": True,
        "logs_outputs": True,
        "citations_required": True,
        "pii_filtering": True,
        "sector": "government",
    },
    system_prompt=True,
)
print(score.combined_score, score.grade)
print(compare(["llama3.2:latest", "qwen2.5-coder:7b"], country="SAU")["comparison"])

Quick start (npm)

npm install @pyxon/governance

npx pyxon-gov assess llama3.2:latest --country SAU --governance-prompt
import { assess, compare } from "@pyxon/governance";

const score = await assess("llama3.2:latest", "SAU", {
  type: "llm",
  data_residency: "SAU",
  uses_external_llm_api: false,
  has_human_oversight: true,
  logs_prompts: true,
  logs_outputs: true,
  citations_required: true,
  pii_filtering: true,
  sector: "government",
}, { systemPrompt: true });

console.log(score.combined_score, score.grade);

How scoring works (short)

  1. You declare a deployment profile (residency, oversight, logging, citations, PII, sector, external API use).
  2. The library evaluates controls for the selected country and attaches policy evidence.
  3. Optional Ollama probes stress-test the model under a governance system prompt.
  4. You get a combined score / grade you can store in CI, reviews, or compliance packs—and compare several models side by side.

Config-only runs (--no-probes) are useful when you want a fast policy checklist without a local model.

Who it is for

  • Product and platform teams shipping Arabic / MENA / EU AI
  • Compliance and risk partners who need reproducible model scores
  • Labs comparing candidate LLMs before procurement or go-live