The Parameterisation Gap

The Tempting Answer

Here is a comforting thought about Koher: it is training wheels.

AI models are improving. Prompt engineering is maturing. Soon — perhaps in a few years — people will be able to define rubric values directly in their prompts, and the models will apply them consistently. The flakiness that makes the Koher architecture necessary today will disappear. At that point, the three-layer separation becomes unnecessary overhead. The training wheels come off.

This is a reasonable position. It is also wrong.

I know because I held it for about ten minutes one morning before correcting myself.

The Self-Correction

The training-wheels framing assumes that the gap Koher addresses is a gap in AI capability — that AI is currently flaky at applying rubric values, and will eventually not be flaky. This frames the architecture as a temporary scaffold around a temporary limitation.

But the gap is not in AI's capability. The gap is between what can be parameterised and what is lived. And that gap does not close with better models.

All of the parameters of what it means to be human can never be parameterised. There will always be something evading the parameterisation of human experience. That something — irreducible, resistant to discrete encoding — is the permanent territory of work like Koher's.

The gap is not in AI's capability. It is between what can be computed and what is lived.

What Can Be Parameterised

Parameterisation works. It works well. The Koher tools are built on it.

The Coherence Diagnostic parameterises a concept statement into five dimensions — CLAIM, EVIDENCE, SCOPE, ASSUMPTIONS, GAPS — and assigns confidence scores to each. This parameterisation captures real patterns. A student who reads their coherence diagnostic and sees that their evidence is thin while their claim is bold has learned something true about their work.

The Play Shape Diagnostic parameterises experiential qualities of play and computes relationships between them. A designer who learns that their three chosen qualities create tension rather than synergy has encountered a real structural property of their design.

Parameterisation — turning lived experience into discrete, computable terms — is powerful. The question is not whether it works. The question is whether it is sufficient.

What Remains

Consider what happens when a teacher sits with a struggling student.

The teacher notices hesitation before it becomes silence. They sense whether the confusion is productive — the kind that precedes insight — or destructive, the kind that precedes shutdown. They know when to push and when to wait. They read the room: who else is watching, what the student's body is doing, what happened in last week's class that might be surfacing now.

None of this is captured by dimensions. Not because the dimensions are wrong, but because dimensions are a way of freezing what is fluid. They take the stream and measure its depth at fixed points. The measurement is real. But the stream is not its measurements.

A teacher's judgment operates on the stream itself — the movement, the temperature, the direction of flow. Parameterisation captures the depth at five points along the bank. The 60% that remains — the part that requires a human — is not a technical limitation waiting to be solved. It is the nature of what is being measured.

Evidence from Teaching

I discovered this gap concretely while designing a tool for mentor-student pairing.

The tool needed to capture how students approach learning — their pace, their tolerance for uncertainty, their relationship to feedback. I defined five dimensions. For teachers, those dimensions made sense: PACE, CERTAINTY_TOLERANCE, FEEDBACK_ORIENTATION, INDEPENDENCE, LEARNING_STANCE. These are pedagogical terms. Teachers understand them because pedagogy is their special mode of being social.

For students, these same dimensions were illegible. Not because students are less sophisticated, but because pedagogy is not their reality. Their reality is social. A student does not think of themselves as having a "certainty tolerance." They think: when I don't know what to do next, I get anxious — or — when I don't know what to do next, I wait and something usually comes.

The same underlying pattern, expressed in two different languages. The teacher's language is pedagogical. The student's language is social. Neither is complete. Neither captures everything the other does. Both are parameterisations of something that, in the living, is neither pedagogical nor social — it is simply how this particular person moves through the world.

Dimension	Teacher Sees	Student Lives
Pace	Rusher, lingerer, oscillator	"I jump in" / "I need to sit with it"
Uncertainty	Settler, dweller, context-dependent	"Not knowing stresses me out" / "Something will come"
Feedback	Direct-seeker, discovery-seeker, evidence-seeker	"Just tell me" / "Ask me things that make me think"

The two languages map to the same underlying structure. The deterministic rules — Stage 2 — do not change regardless of which language produced the input. Only the surface of Stage 1 changes depending on who is speaking.

This is a small example of a large principle. The parameterisation captures something real. But the choice of which parameterisation — pedagogical or social — is itself an act of judgment that no model can make, because the model does not know who is speaking or what language they live in.

The Horizon

The current Koher architecture works with dimensions: named, discrete, thresholded. This is not the final form.

As of today, we have binary devices that store and measure values in discrete terms. Dimensions and thresholds are what discrete computers can work with. They are the best approximation available to us of something that, in the living, is not discrete at all.

There is work happening in the world — in computational biology, in analogue computing, in models of cognition — that may eventually find ways to compute with amorphous terms. Terms that are not pinned to fixed dimensions. Terms that shift, overlap, resist boundaries. We do not know what that looks like yet. We do not know when it will arrive.

When it does, the form of the adapter will change. Dimensions may give way to something we do not yet have a name for. The three-layer separation — qualification, rules, language — may take new shapes.

What will not change is the gap itself. The gap between what can be computed and what is lived does not close with better computing. It is not a technical limitation. It is an ontological one — a fact about the relationship between the map and the territory.

The form of the adapter will change. The need for the adapter will not.

What This Means for the Architecture

The Koher architecture separates qualification, rules, and language. This position statement says: that separation is the current best encoding of a permanent need.

• The need is permanent: AI will always require something between its computation and human experience. Not because AI is deficient, but because human experience exceeds parameterisation.
• The form is current: Dimensions, thresholds, and deterministic rules are the best tools we have for encoding judgment in discrete terms. They are not the only tools that will ever exist.
• The practice is honest: Every Koher tool explicitly states what it parameterises and what it does not. The 40% that is captured and the 60% that remains. This honesty is the architecture's deepest feature — not the dimensions themselves, but the refusal to pretend the dimensions are sufficient.

The architecture is not a claim that dimensions are forever. It is a claim that the gap is forever, and that the honest response to the gap is to build tools that name what they can see and acknowledge what they cannot.

Summary

The tempting answer is that Koher is training wheels — a temporary scaffold for a temporary AI limitation. The better answer is that the gap between what can be parameterised and what is lived does not close. AI will keep improving. The gap remains.

The current architecture works with dimensions because discrete computers work with discrete terms. The form will evolve. What will not evolve is the need for something that sits between computation and experience, making the translation honest.

The architecture's deepest contribution is not separating qualification from rules from language. It is insisting that the separation be visible — that the tool shows you what it measured and admits what it did not.

The gap between the parameterisable and the lived is permanent territory. Koher sits in that gap. Not to close it, but to make it navigable.