There is a gap between understanding something and doing something about it.
This seems obvious. But it is easy to underestimate how much is lost when a system — human or artificial — moves too quickly from the words that express a thought to the actions that are supposed to follow from it. A document is written. It looks clear. Someone implements it. What gets built is not what was meant.
The problem is not usually bad writing or bad reading. It is that human-readable text carries meaning through a context that is partly shared and partly assumed — and when that context is not preserved, the text loses its grip on the original intent. The words remain. The understanding has thinned.
This post proposes that responsible systems — systems that want to convert human intention into design, task, policy, or action without distorting it — need to operate on something beneath immediate interpretation: a layer where the states that give text its meaning can be held, examined, and carried forward.
We are calling this layer the primitive states of meaning.
The problem with fluent output
Language models and AI systems can produce text that sounds complete. It has the shape of a conclusion. It uses the right vocabulary. It answers the question that was asked.
But fluency is not the same as understanding. A system can generate a requirements document, a policy statement, or a task breakdown without having preserved the epistemic state that the original human intent occupied — without tracking what was certain and what was approximate, what was a constraint and what was a preference, what carried authority and what was speculative.
When that state is lost, the output becomes detached from the intent. It may still be useful. But it is no longer governed. Corrections become harder to reason about. Amendments cannot trace back to what they are amending. The record of how a claim became a commitment disappears.
Primitive states of meaning
What follows is a working list — not a final taxonomy — of cognitive states that a governed system should be able to preserve as meaning moves from perception to expression to action. These are proposed as a starting vocabulary. Some are more developed than others. All are offered as candidates for scrutiny.
Before interpretation
Presence — something is being attended to. Before any analysis begins, there is a subject of attention. A system that cannot mark what it is attending to cannot be held accountable for whether it attended to the right thing.
Distinction — two things are not the same. Responsible reasoning depends on the ability to hold boundaries between concepts, categories, and entities without collapsing them prematurely. Distinction is the precondition for precision.
Continuity — this is the same thing as before. Identity across time and context. Without continuity, a system cannot track whether a concept, entity, or commitment has persisted or changed.
Difference — something has changed. Where continuity tracks sameness, difference marks departure. Both are necessary. A system needs to know when it is dealing with a revision, not merely a restatement.
Context — this meaning belongs to this situation. Every claim is made in a setting that shapes what it means. Stripping context from a claim and treating it as universal is one of the most common sources of misapplication.
As meaning takes shape
Salience — this stands out. Attention is not evenly distributed, and salience records what has been foregrounded. But salience is not authority: something can draw attention because it is unusual, repeated, or emotionally charged without that making it more important or more reliable. A governed system should preserve salience without letting it determine weight.
Relevance — this applies here. Not everything present is pertinent. A governed system should be able to distinguish what it is drawing on from what it is setting aside, and why.
Association — these things go together. Meaning is relational. Concepts acquire their sense partly from what they are connected to. Preserving associations preserves the web of meaning that holds a concept in place.
Tension — these things are in conflict. When two things that are both present cannot be fully reconciled, that tension is itself information. Resolving it prematurely is a form of distortion. A system should be able to hold tension without collapsing it.
Uncertainty — this is not settled. The epistemic status of a claim is part of the claim. Uncertainty is not a failure state. It is a condition that, when preserved, allows downstream reasoning to remain calibrated.
Intention — this is the purpose the statement seems to serve. Not just what was said, but the direction it was pointed. A system that cannot distinguish intent from expression will implement the letter and miss the spirit. This is an inference, not a direct read — the system proposes what the statement appears to be trying to do, and that proposal can be corrected.
Modality — this is what could, should, or must be. The difference between a possibility, a recommendation, and a requirement is not always visible in the text alone. Modal state must be preserved explicitly.
Claim type — this is an assertion, a question, a proposal, or a commitment. The kind of speech act a statement represents changes what can be done with it. Conflating claim types is a source of downstream confusion.
Before action
Authority — what gives this claim binding force, if anything. Authority is not only a matter of who says something. It can derive from document status, process, role, prior approval, or governing rule. Tracking authority is part of tracking what a claim can authorize.
Adequacy — is this sufficient for the purpose it is meant to serve? Adequacy is a relation between a claim and a use. A claim that is adequate for exploration may not be adequate for implementation.
Implementability — how, if at all, may this become action? A claim may be possible to act on in several ways — as code, as a test, as a review criterion, as a prohibition, as a deferred question, or as a documentation note. Implementability is not only a question of feasibility. It is a question of what downstream disposition is appropriate.
Commitment — this has been agreed to. A commitment is a claim that has crossed a threshold — it is no longer under consideration but under obligation. The transition from proposal to commitment is consequential and should be marked.
Amendment — the record or framework must change. In the ordinary case, an amendment points at something that was said before and changes it. But amendment can also mean something deeper: that the current categories cannot hold the meaning adequately, and the representational framework itself needs revision. A governed record must be able to support both kinds of change and distinguish between them.
A note on the child machine
In 1950, Alan Turing proposed an alternative to the question of whether machines could think. Rather than simulating an adult human mind, he suggested, it might be better to simulate a child's mind — and then educate it. The child machine, as he called it, would not begin with full capability but would develop it through a process that resembles learning.
The proposal was speculative. Turing did not develop it into a technical program. But the intuition it points toward is still worth taking seriously: that intelligence might not be best approached by installing a finished system, but by building one that can develop from simpler states toward more complex ones.
The primitive states proposed above are not a theory of mind. But they are shaped by a similar intuition: that the responsible path from perception to action passes through a layer of structured intermediate states — states that can be examined, preserved, and reasoned about — rather than moving directly from language in to language out.
Substrate as an emerging system for this
Always Becoming is developing a layer called Substrate. The intention is to build a governed system that can preserve these kinds of primitive states as meaning moves from human intent through interpretation, memory, design, and action.
This is not complete. Substrate is a direction, not a finished product. What exists is a set of commitments: that human authority remains visible at every step, that AI assistance is disclosed, that corrections are preserved alongside the original claims, and that no output is treated as final without human review.
The list of primitives above represents current thinking, not an agreed specification. It will be revised. Some concepts on the list will prove redundant. Others that belong will be missing. That is expected.
What matters is that the question is being held: what must a system preserve in order to handle meaning responsibly?
This post is a draft. It has not been through adversarial review. The primitives listed are offered as a working vocabulary for further examination, not as settled doctrine.