# AI SAFE² Implementation Examples
### Governed runtime patterns across agent stacks

[![AI SAFE²](https://img.shields.io/badge/AI_SAFE%C2%B2-v3.1-F6921E?style=flat-square)](../README.md)
[![Surface](https://img.shields.io/badge/Surface-Examples-820F1A?style=flat-square)](./README.md)
[![Status](https://img.shields.io/badge/Status-Reference_Implementations-808080?style=flat-square)](../docs/REPOSITORY-UX-STANDARD.md)

[Framework Home](../README.md) | [Cross-Pillar Governance](../00-cross-pillar/README.md) | [AISM](../AISM/) | [NEXUS](../NEXUS/) | [Dashboard](https://cyberstrategyinstitute.github.io/ai-safe2-framework/dashboard/)

---

## Reviewing technology contributions

Use the [Technology Card](../AISM/templates/TECHNOLOGY-CARD.md) to
document each example's actual implemented, enforced, validated, and evidenced
controls. The [nine-example backlog](../research/technology-contribution-examples.md)
illustrates source-attributed candidates; it adds no runnable integration to this
index. Reference code, passing fixture tests, and installed examples do not
establish an adopting organization's maturity. Apply the
[profile method](../AISM/TECHNOLOGY-CONTRIBUTION-PROFILE.md) before extending claims.

## Purpose

The examples directory shows how AI SAFE² controls can be operationalized around real agent frameworks, automation systems, gateways, and security tooling.

Examples demonstrate implementation patterns. They do not create framework conformance merely by being copied or installed. A deployment must satisfy applicable control outcomes and evidence requirements in its own operating context.

---

## Example Index

<!-- EXAMPLES:START -->
<!-- This table is auto-generated by scripts/generate_examples_table.py. Do not hand-edit between these markers; edit each example's README.md instead. -->

| Example | Stack | What It Governs |
|---|---|---|
| [`aism-decision-card/`](./aism-decision-card/) | AISM Decision Support | Executable assessment demonstrating evidence-aware AISM scoring and a human Decision Card. |
| [`aism-remediation/`](./aism-remediation/) | AISM Remediation Decision Support | Executable evidence-bound AISM remediation plan with ownership, alternatives, exit criteria, history, and human decision gates. |
| [`anti-gravity-sovereign-runtime/`](./anti-gravity-sovereign-runtime/) | Anti-Gravity | Sovereign runtime defense package for Anti-Gravity agent deployments. |
| [`autogen-sovereign-runtime/`](./autogen-sovereign-runtime/) | Microsoft AutoGen | Governance wrapper for AutoGen multi-agent orchestration. |
| [`challenge-controlled-executor/`](./challenge-controlled-executor/) | Challenge 001 Controlled Executor | Reproducible process-boundary evaluator with pre-registration, independent grading, incomplete states, and cross-bound evidence receipts. |
| [`claude-code-sovereign-runtime/`](./claude-code-sovereign-runtime/) | Claude Code | Hardens agentic coding sessions against unsafe tool execution, scope creep, and drift. |
| [`codex-sovereign-runtime/`](./codex-sovereign-runtime/) | OpenAI Codex | Runtime enforcement wrapper for Codex-driven coding agents. |
| [`crewai-sovereign-runtime/`](./crewai-sovereign-runtime/) | CrewAI | Governance wrapper for CrewAI crew-based orchestration. |
| [`cursor-sovereign-runtime/`](./cursor-sovereign-runtime/) | Cursor | Runtime controls for Cursor IDE chains, .cursor/rules injection surfaces, and AI-generated code execution. |
| [`environment-decision-card/`](./environment-decision-card/) | Environment Discovery, Drift, Evidence, and Policy | Privacy-safe baseline-to-decision workflow with human cards and a unified evidence manifest. |
| [`hermes-sovereign-runtime/`](./hermes-sovereign-runtime/) | NousResearch Hermes | Closes four critical audit findings identified in the Hermes agent. |
| [`ishi/`](./ishi/) | Supervisory agent | Reviews agent plans and enforces governance decisions before high-risk actions execute. |
| [`langchain-sovereign-runtime/`](./langchain-sovereign-runtime/) | LangChain | Runtime controls for LangChain chains and tool-calling agents. |
| [`langflow-sovereign-runtime/`](./langflow-sovereign-runtime/) | Langflow | Sovereign runtime defense package for the Langflow visual builder. |
| [`langgraph-sovereign-runtime/`](./langgraph-sovereign-runtime/) | LangGraph | Runtime controls for LangGraph stateful agent graphs. |
| [`lovable-sovereign-runtime/`](./lovable-sovereign-runtime/) | Lovable | Runtime enforcement package for Lovable Agent mode, database actions, deployments, and MCP integrations |
| [`love_equation/`](./love_equation/) | Alignment research | Drift-scoring model adapted into AI SAFE2 behavioral monitoring. |
| [`make-sovereign-runtime/`](./make-sovereign-runtime/) | Make.com | Sovereign runtime defense package for no-code automation scenarios. |
| [`mcp-security-toolkit/`](./mcp-security-toolkit/) | MCP (protocol security) | Moved. Use safe2 scan mcp, safe2 score mcp, safe2 gate mcp and safe2 mcp wrap-stdio / wrap-proxy (CP.5.MCP v3.1). |
| [`openclaw/`](./openclaw/) | OpenClaw | Full two-layer governance stack, internal identity and memory files plus external scanner and gateway. |
| [`slowmist-overlay/`](./slowmist-overlay/) | SlowMist / OpenClaw | Threat-intelligence overlay mapping SlowMist OpenClaw security practices to AI SAFE² controls. |
| [`xai-grok-sovereign-runtime/`](./xai-grok-sovereign-runtime/) | xAI Grok | Sovereign runtime defense package for Grok-based agents. |

<!-- EXAMPLES:END -->

---

## v3.1 Notes

- The overall framework remains **161 core controls**.
- MCP implementations should use the [CP.5.MCP v3.1 profile](../00-cross-pillar/cp5_mcp_server_security.md).
- UAS is a 27-requirement regulatory profile extension; it does not create CP.11 as a core control.
- NEXUS examples are reference implementations, not mandatory AI SAFE² dependencies.
- Legacy session-oriented state terminology may appear in older examples during the migration window; new governance-bearing examples should use `request`, `handle_scoped`, and `durable` semantics where applicable.

---

## Adding an Example

Every new `examples/<name>/README.md` included in the generated index should contain:

```text
<!-- stack: Stack Name -->
<!-- description: One concise description. -->
```

The generator validates those tags and updates the table above. The Lovable example has an explicit compatibility metadata entry because it predates this convention.

---

## Navigation

[Framework Home](../README.md) | [Cross-Pillar Governance](../00-cross-pillar/README.md) | [NEXUS](../NEXUS/) | [Scanner](../scanner/README.md) | [Gateway](../gateway/) | [Dashboard](https://cyberstrategyinstitute.github.io/ai-safe2-framework/dashboard/)

---

*AI SAFE² v3.1 · [Cyber Strategy Institute](https://cyberstrategyinstitute.com/ai-safe2/)*
