Hello there,
Artificial intelligence is advancing rapidly. But many of the systems organizations use to govern software were designed for deterministic behavior: defined inputs, repeatable execution, and predictable outputs.
AI agents introduce a different challenge. Their behavior is probabilistic. Similar inputs can produce different outputs, making runtime policy enforcement, oversight, and auditability more difficult.
SAFi grew from a framework for understanding human cognition and became an open-source runtime governance engine for agentic AI.
SAFi is designed to place deterministic governance boundaries around probabilistic systems. It enforces policies in real time, governs tool calls, and records every decision for audit.
Four design laws behind SAFi
1. The Law of Probabilistic Entropy
A probabilistic engine, including a large language model, cannot be relied upon to enforce absolute boundaries consistently on its own.
2. The Law of Structural Determinism
Critical policy 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, evaluate its own conduct, and authorize its own execution without separation of responsibilities.
SAFi separates Values, Intellect, Will, Conscience, and Spirit for practical governance purposes. Values define what matters. The other four faculties describe how the governed turn is processed. This separation creates a more inspectable decision path.
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 an organization’s charter and policies to evaluate whether an agent’s response or proposed action is aligned with its defined purpose, rather than relying only on blocklists.
Where to start
Choose the path that best matches your role. Each one begins with a different question.
If you build or run the platform
The first thing to know is that SAFi does not require you to rebuild your existing agent stack.
Evaluate an existing agent
Use the /evaluate gateway to govern the output of an agent you have already built. Your orchestration, prompts, and tool layer can remain where they are.
Run the quick start
Clone the repository and run SAFi locally with Docker and a database.
Read the developer guide
Explore the repository layout, architecture, policy authoring, tool authorization, and integration surfaces.
Find a good first issue
Review accessible contribution opportunities and open an issue when you find something worth improving.
Clone the repository, run SAFi, and tell us where it fails.
If you own technology strategy
The strategic question is portability.
Your charter, policies, and audit trail live in your database rather than with a model provider. Changing or upgrading the model that drafts a response does not require you to restart your governance architecture or abandon the evidence it has accumulated.
Run the live demo
See what a governed turn produces before evaluating the architecture.
Read the governance documentation
Review how SAFi represents policies, evaluates responses, governs tool calls, and records decisions.
Review the readiness material
Examine the documentation discussing SEC, FINRA, the EU AI Act, HIPAA, and GDPR. The material distinguishes current functionality from planned work and makes no certification claims.
Run the demo and inspect the audit trail.
If you practice AI governance, ethics, or research
The central question is whether a governance theory can be operationalized and tested through evidence.
Read a worked example
Examine an organization’s value set, the response produced by its agent, the value-by-value evaluation ledger, and the hash-chained audit entry for that turn.
Review the mathematical specification
See how the evaluation is defined and what each faculty is deliberately permitted or denied to do. The separation of responsibilities is central to the independence of the audit.
Review benchmarks and validation
Examine the methods, results, and supporting evidence behind SAFi’s reported performance.
Inspect a real audit record and open an issue with the part you find least convincing.
Ways to participate
SAFi is an open-source project. You can participate by:
- Exploring the GitHub repository
- Running the demo
- Inspecting an audit trail
- Reading the research and technical articles
- Reviewing the documentation
- Opening an issue
- Suggesting an improvement
- Challenging a design assumption
- Contributing code, tests, documentation, or research
The most useful contribution is not agreement. It is a substantive question, a reproducible problem, or evidence that helps us improve the system.
Licensing
SAFi is released under the AGPL-3.0 license so that its governance infrastructure remains available for inspection, modification, and responsible reuse.
The Self Alignment Framework is licensed separately for free use with attribution.
Get in touch
If your question does not fit a standard issue, please reach out. We welcome people interested in collaboration, research, deployment, documentation, and critical examination of the ideas behind SAFi.
If you think we have something wrong, especially in a way that can be demonstrated through the source code, a live demo, or an audit record, we would particularly like to hear from you.
Clone the repository, run SAFi, and tell us where it fails.
