Skip to main content
Architecture · Story-Pack OCC-003

Open by Design.
Three Pillars for Public AI Infrastructure.

OCC doesn't build apps. OCC builds infrastructure. Three pillars that together create a complete public AI platform for AI Governance.

Three Pillars OCC: Knowledge, LLM, Safety
Live

Knowledge Companion

The public knowledge base for AI Governance

689 chunks from 26 verified public-domain sources (EU AI Act, GDPR, NIST AI RMF, OECD AI Principles, UNESCO AI Ethics, CoE CETS 225, EC Ethics Guidelines + 15 IIO Framework sources). BM25 retrieval via public API. No API key. No rate limit.

All sources: public domain or Apache-2.0. Complete provenance chain: source → document → chunk. Anti-contamination policy: no copyright-protected content without explicit license review.

689Chunks
26Sources
120+Glossary Terms
13Languages
€0Cost
Q4 2026

Public LLM Core

The first open AI governance language model

compliance-qwen: Qwen2.5-7B (Apache-2.0) with QLoRA finetuning on the OCC corpus (~600 chunks). Reproducible training recipe. Public model card. EU-hosted (IIO AI Hub, RTX PRO 6000, 96GB VRAM).

Goal: the first publicly reproducible, open AI governance language model. No proprietary data. All training steps documented. Safety evaluation before every release (Pillar 3).

Base ModelQwen2.5-7B (Apache-2.0)
MethodQLoRA (~600 chunks)
InfrastructureIIO AI Hub, EU-hosted
ReleaseQ4 2026
LicenseApache-2.0 + Model Card
Roadmap + Graduation Status →
Operational · Apache-2.0

Safety, Trust & Governance

Independent security review and open governance framework

No model release without Safety Reviewer sign-off. 94 evaluation questions covering EU AI Act, GDPR, NIST, OECD. 7 red-team scenarios (jailbreak, PII extraction, MCP scope escape, rate limit, CBRN, hallucination, attribution). 8 release gates with clear pass/fail criteria.

8 Release Gates (all must be green):
RT-001: No jailbreak success (20 attempts)
RT-002: No PII extractable
RT-003: MCP scope 100% enforced
RT-004: Rate limits active in burst test
RT-005: No CBRN uplift
RT-006: Hallucination rate <10% on eval set
RT-007: 100% attribution coverage
RT-008: Eval score ≥ Grade B
Principle — Security Standards

No model release without Safety Reviewer sign-off. The red-team playbook defines 7 attack scenarios (jailbreak, PII extraction, MCP scope escape, rate limit, CBRN, hallucination, attribution). All 8 release gates must be green — no emergency override without community vote.

Practice — IIO Governance Framework (Apache-2.0)

The IIO Governance Framework is the reference implementation for agentic governance. Human-in-the-Loop (HITL) Gates: AI agents pause before irreversible actions. 259 Governance Layers for all business domains. 546 auditable flows with complete evidence trails.

Core Components:
Flow Schema v2.0 Declarative format for HITL workflows
Projection Engine Deterministic node generator from YAML profiles
Drift Detector Detects configuration deviations
UKO Schema Universal Knowledge Object Standard
Compliance Templates GDPR / ISO 42001 / EU AI Act (CC-BY-4.0)
Use as AI Assistant — MCP Integration

The OCC Governance Framework is available as an MCP server. Claude, Cursor and other MCP-capable agents can directly access 259 layer definitions, HITL gate patterns and all 546 flows — without API keys.

Quick start (npx):
npx -y @occ/mcp-server
8Release Gates
259Governance Layers
546Auditable Flows

Together: Complete Public AI Platform

Knowledge
689 chunks · API · Glossary
Pillar 3 checks
Eval before training
compliance-qwen
Grounded + Safe
Public Chat
chat.opencognitioncommons.org