How SAF Was Born: The Philosophical Foundation of SAFi

The Self-Alignment Framework, or SAF, did not begin as an artificial intelligence project. It began with a personal question:

How does a person move from understanding something to acting on it, judging the result, and remaining true to what they believe?

That question emerged from a lifelong curiosity about nature, human thought, philosophy, and the structure beneath apparently complex things.

From curiosity to philosophy

I grew up in a small village in El Salvador with limited access to formal education, but I was surrounded by questions. I wanted to understand how things fit together, how natural systems worked, and what gave life structure and meaning.

When I moved to the United States at sixteen, I encountered philosophy more deeply. The works of Plato, Aristotle, Kant, Nietzsche, and Thomas Aquinas gave language to questions I had already begun asking. Their ideas did not provide a single answer. Instead, they offered different ways to examine knowledge, action, responsibility, purpose, and the human search for truth.

Even while building a career in technology, philosophy remained part of my thinking. I continued asking:

• How do we distinguish understanding from assumption?
• What causes us to choose one action over another?
• How should we evaluate the consequences of a decision?
• How can a person remain coherent when circumstances change?
• What role do values play in practical action?

Over time, these questions began to form a pattern.

The alignment loop

I came to understand human thought and action as a continuous loop with five functional stages:

• Values ground
• Intellect discerns
• Will decides
• Conscience judges
• Spirit integrates

These terms are drawn from human experience, but SAF does not claim that they represent separate physical components or that every person experiences them in exactly the same way. They describe functions within a model of principled action.

Values: the ground of the loop

Values are the foundation. Before anything can be understood, chosen, or judged, there must be something to be faithful to. Values are what a person or an institution has decided to stand for, held explicitly enough that action can be measured against them.

Values are supplied to the loop rather than produced by it. That distinction matters. A system that invents its own standard as it proceeds is not aligned to anything beyond itself; it is only consistent.

In an AI system, this role can be reflected in questions such as:

• Which principles govern this agent?
• Who declared them, and when?
• Which of them apply to the request at hand?
• How are competing values weighed against one another?
• Can the governing set be changed without anyone noticing?

Intellect: discerning what is understood

Intellect is the function of understanding. It gathers information, distinguishes relevant facts from distractions, identifies uncertainty, and considers possible interpretations. Without discernment, a decision may be confident but poorly grounded.

In an AI system, this role can be reflected in questions such as:

• What does the system know?
• What information is missing?
• Which sources support the proposed conclusion?
• How confident should the system be?
• Is the request ambiguous or outside its authority?

Will: choosing an action

Will is the function of decision. After considering what is known, a system or person must choose what to do. That choice may involve acting, waiting, asking for clarification, refusing a request, or referring the matter to someone with appropriate authority.

A decision is not automatically aligned simply because it is efficient. It must also remain within the principles, permissions, and responsibilities that govern the situation.

Conscience: judging the proposed action

Conscience is the function of judgment. It evaluates whether a proposed action is consistent with applicable values, rules, duties, and constraints. In human life, this judgment can be deeply personal. In SAFi, the concept is operationalized through explicit policies and governance controls.

This can include asking:

• Is the action permitted?
• Does it violate a safety or privacy requirement?
• Is the system exceeding its authority?
• Does the action conflict with the stated purpose of the agent?
• Should the action be denied, redirected, or flagged for review?

In this sense, conscience is not presented as proof that an AI system has human moral agency. It is a functional part of a governance model.

Spirit: integrating the whole system

Spirit is the integrating function. It preserves coherence among understanding, decision, judgment, identity, and purpose. It asks whether the action fits the larger character of the system rather than merely satisfying an isolated instruction.

In SAF, Spirit is not a claim about supernatural consciousness. It represents the function that maintains continuity and coherence across the alignment loop.

For an AI system, this might involve checking whether a response:

• Remains consistent with the agent’s defined role
• Reflects its governing principles
• Preserves continuity across decisions
• Makes uncertainty and limitations visible
• Supports the system’s intended purpose

Spirit therefore helps prevent alignment from being reduced to a single rule or one-time approval. It represents the integrity of the whole process.

The influence of Thomas Aquinas

Thomas Aquinas became an especially important influence because his treatment of Intellect and Will closely corresponded to the structure I was developing.

At the same time, SAF is not a restatement of Aquinas’s philosophy. I understood the sequence differently in two important ways.

1. First, I placed Conscience after Will, as a function that evaluates the proposed action.
2. Second, I introduced Spirit as the integrator and guardian of the entire loop.

Not everything was a departure. Aquinas describes synderesis, the habit that holds first principles and does not deliberate about them. That is very close to the role Values plays here: the standard is held, not negotiated in the moment. When SAF was later implemented, this was the name the component kept.

The concept of Spirit grew from biblical passages and personal reflection. I had not encountered it framed in precisely this way in the sources I had studied. That does not mean the idea exists without precedent. It means this particular arrangement emerged through my own attempt to understand how knowledge, choice, judgment, and coherence relate to one another.

SAF came before AI

SAF was conceived before the age of AI. It was not created as an artificial intelligence framework.

It began as a model of the human cognitive process and of how principled action might be understood. AI later became a powerful environment in which to test whether the framework could be translated into practical systems.

That distinction matters.

SAF does not begin with the assumption that machines are human. It does not claim that AI systems possess souls, human consciousness, or moral status. Nor does it claim that software can carry human responsibility in the same way a person can.

Instead, SAF provides a way to describe and structure the functions required for behavior that is:

• Principled
• Transparent
• Coherent
• Reviewable
• Accountable to human-defined governance

From SAF to SAFi

SAF is the theory. SAFi is the implementation.

SAF describes a model of aligned action through discernment, decision, judgment, and integration. SAFi, the Self-Alignment Framework Interface, applies that model to agentic AI.

SAFi is an open-source runtime governance engine for agentic AI. It turns principles into operational controls by helping systems:

• Evaluate policies at runtime
• Constrain actions according to defined rules
• Flag a turn for supervisory review
• Record decisions and relevant context
• Support post-execution review and auditability
• Preserve a clear relationship between an agent’s purpose and its behavior

This makes SAFi different from a conventional agent framework.

An agent framework helps developers build or coordinate agents. SAFi addresses a different question:

How can an agent be governed while it is operating?

SAFi is not intended to replace the frameworks used to build agents. It is the governance layer that helps make those agents deployable with greater confidence.

From philosophical functions to runtime governance

The connection between SAF and SAFi can be represented simply:

One note on sequence. The loop above is SAF as I first understood it: strictly sequential. SAFi’s runtime is not so tidy — it consults the Will at more than one point, because in a working system authorization has to happen before an action, not only after a result. That is an implementation decision, not a revision of the framework.

This translation does not mean that an AI system literally possesses intellect, will, conscience, or spirit. These terms name governance functions. They provide a conceptual structure for designing and evaluating behavior.

The technical implementation remains concrete. Policies can be evaluated. Actions can be allowed or denied. Turns can be queued for human review. Events can be logged. Decisions can be inspected after the fact.

The philosophical model gives these controls an organizing principle. Runtime governance gives the philosophy a practical form.

Why this matters

As AI systems become more capable, the central challenge is no longer only whether they can produce useful outputs. Organizations also need to know:

• What was the system trying to do?
• Which information influenced its decision?
• Was the action within its authority?
• Which policies applied?
• Why was the action allowed or denied?
• Can the decision be reconstructed later?
• What happens when the system encounters uncertainty or conflict?

These questions are especially important for platform engineering teams, IT directors, security leaders, and AI governance practitioners. They need more than a model that can generate a response. They need systems whose behavior can be governed, observed, and reviewed.

SAFi is built around that need.

The purpose of faithful agents

The goal is not to create machines that imitate human beings or to assign them human moral responsibility.

The goal is to create agents whose behavior remains faithful to defined principles, transparent in operation, coherent across decisions, and auditable over time.

Faithfulness in this context means that an agent’s behavior reflects its declared purpose and governing policies. It does not mean obedience without judgment. A faithful system may need to refuse an unsafe request, ask for clarification, identify a conflict, or leave a decision for a human to review.

This is why governance must exist at runtime. A policy document that is never checked during execution cannot reliably govern behavior. A principle that cannot be traced to an observable decision remains difficult to evaluate.

The continuing work

SAF began as an attempt to understand human alignment. SAFi extends that inquiry into the design and governance of agentic AI.

The framework remains grounded in a simple idea:

A system should not merely act. It should act in a way that can be understood, evaluated, and held to its governing principles.

SAF provides the conceptual model. SAFi provides the operational architecture.

Together, they connect philosophy with implementation:

• Discernment becomes structured understanding.
• Decision becomes explicit intent.
• Judgment becomes enforceable policy.
• Integration becomes system-wide coherence.
• Reflection becomes an auditable record of behavior.

SAFi will be complete when governing an agent becomes as clear and repeatable as versioning code. Developers should not have to ask where governance belongs or how to reconstruct an agent’s decisions. The answer should be obvious: governance must be part of the system’s operating layer.

That is the journey from SAF to SAFi.

SAF is the theory. SAFi is the implementation. SAFi makes principled agent behavior enforceable, observable, and auditable at runtime.