Hello there,

Artificial intelligence is advancing rapidly. But the systems organizations have traditionally used to govern software were designed primarily for deterministic behavior: defined inputs, repeatable execution, and predictable outputs.

AI systems, especially large language models, introduce a different challenge. Their behavior is probabilistic. They can produce different outputs from similar inputs, which makes policy enforcement, oversight, and auditability more difficult.

I have been thinking about these challenges for a long time. That work began as a framework for understanding human cognition and eventually became the foundation for SAFi, an open-source governance engine for AI agents.

SAFi is designed to govern probabilistic machines with strict, deterministic boundaries.

SAFi’s Four Governing Laws

1. The Law of Probabilistic Entropy. A probabilistic engine, such as a large language model, cannot be relied upon to enforce absolute boundaries consistently on its own.

2. The Law of Structural Determinism. Critical safety gates and execution permissions should be enforced through deterministic logic, not semantic inference alone.

3. The Law of Faculty Isolation. A single model should not generate a response, judge its own ethics, and authorize its own execution. Separating those responsibilities creates a more defensible audit.

4. The Law of Defined Teleology. To govern a system rather than merely restrict it, you must define the outcome it is intended to pursue. SAFi uses that defined purpose to guide behavior toward an organization’s standards, rather than relying only on blocklists.

Where to Start

Pick the role that fits — each path starts somewhere different.

If you run the platform

The most useful thing to know first is that SAFi does not require you to rebuild your stack.

• The /evaluate gateway — govern the output of an agent you have already built. Your orchestration, prompts and tool layer stay where they are.
• Quick Start — run SAFi locally with Docker and a database.
• Developer Guide — repository layout, architecture, policy authoring, tool authorization and integration surfaces.
• Good first issues — accessible ways to contribute.

If you own the technology strategy

The strategic value is portability. Your charter, policies and audit trail live in your database. Changing or upgrading your model provider changes which model drafts the response, but it does not restart your governance or orphan your evidence.

• Live demo — see what a governed turn actually produces.
• Regulatory readiness — documentation for SEC/FINRA, the EU AI Act, HIPAA and GDPR. Each distinguishes what ships today from what is roadmap, and makes no certification claims.

If you practice AI governance, ethics or research

The central question is whether the theory can be operationalized, and whether its results can be verified through evidence.

• A worked example — a governed agent for a real organization: its value set, the answer produced, the value-by-value ledger with a confidence on each score, and the hash-chained audit entry for that turn.
• Math specification — every formula, and what each faculty is deliberately denied. That separation is where the independence of the audit originates.
• Benchmarks and validation — the methods, results and supporting evidence behind SAFi’s reported performance.

Ways to participate

SAFi is an open-source project, and there are several ways to participate:

• Explore the GitHub repository
• Run the live demo
• Read the articles
• Review the documentation
• Open an issue
• Suggest an improvement

Licensing

SAFi is released under the AGPL-3.0 license, because governance infrastructure should remain available for inspection, modification and responsible reuse.

Get in touch

If your question does not fit a standard issue, please reach out. I would be glad to hear from people interested in collaboration, research, deploying SAFi, or challenging the ideas behind it.

If you think I have got something wrong, I would especially like to hear from you.