Every coach has seen it. A student who looks flawless on the pads, clean lines, textbook mechanics, and then falls apart the first time a partner refuses to cooperate. The standard explanation is that they weren’t ready, or they froze, or they need more reps. I think the explanation is simpler and more uncomfortable. The training was built on a model of what a human in motion is, and that model is wrong.

This article walks the chain from that model down to the failure on the mat. It runs through five links: how we mistake correlation for cause, what happens when we train form instead of function, why skill built one way doesn’t travel, why breaking movement into parts backfires, and what all of that rests on. The last link is the one that matters. Everything above it is a symptom.

Correlation isn’t causation, and instruction runs on correlation

Two things moving together can happen for four reasons. Only one of them is that one causes the other.

The cause can run backwards. Fitter athletes do more mobility work, so it looks like mobility work makes you fit, when fitness is what makes people willing and able to do the work. A hidden third factor can drive both. Ice cream sales track drownings because summer drives both. The sample can be filtered in a way that manufactures the pattern. A gym that only keeps injury-free members can say its members rarely get injured without that telling you anything about the training. And with enough variables, some pairs will line up by chance.

What earns a causal claim is intervention. You change one thing, hold the rest steady, and watch what happens. Randomized experiments exist because randomization breaks the link to every hidden factor at once.

Most instruction lore lives in the correlation zone. “Elite performers do X” is a selection effect until proven otherwise. “My students improved after I introduced Y” has no control group, and anyone who was underperforming would probably have improved anyway. Observational data can point at a hypothesis. It can’t confirm one.

The useful reflex isn’t “correlation is worthless.” It’s “correlation is a reason to ask which of the four explanations is doing the work.”

When function improves and form gets worse

Here’s a pattern I see whenever a group switches from classical repetition to a constraints-led approach. Actual performance improves. Students solve problems under pressure they couldn’t solve before. And the movement looks worse. Rougher, less symmetrical, further from the ideal.

Whether that’s a problem depends entirely on what you’re measuring. If form is the outcome, classical training wins. If function is the outcome, CLA wins. So which one did you actually want?

Classical repetition optimizes toward a template. The student learns to reproduce an ideal movement under stable conditions, and it looks beautiful because it’s practiced in exactly those conditions. CLA optimizes toward a task under varying constraints. The movement that comes out is a solution that works for that body in that context, and by definition it deviates from the template. What gets called “worse form” is usually just “not the form I had in my head.”

Bernstein said it in the sixties: repetition without repetition. Newell and others later showed that degeneracy, many movement solutions for the same outcome, is a marker of robust skill, not a lack of it. The performer who looks identical every time is probably fragile.

And here’s where the correlation trap bites. Coaches have watched for decades that the best performers often have beautiful technique. The conclusion becomes “beautiful technique causes performance.” But performance and aesthetics could both be driven by thousands of hours of exposure, or the ones who look clean could be the ones who got selected forward in the system. Confounding and selection, both at once. Training the aesthetic directly and expecting function is treating the symptom.

The research is consistent. Blocked repetition gives better results during the session and worse retention and transfer. Instructions about body parts and form degrade outcomes compared to instructions about effects in the environment. Representative practice produces skill that holds in competition, even when it looks messier in the room. What looks efficient in the moment correlates poorly with what lasts.

If the goal is function under pressure, measure function under pressure. Did they solve the task? Under variation? Against resistance that doesn’t cooperate? Anyone still grading form is measuring a proxy that was never validated against the real outcome. Form isn’t irrelevant. Some movement patterns are more injurious than others. But that’s a safety constraint you build into the task, not a goal in itself.

Near-transfer and far-transfer

Near-transfer is when skill carries over between situations that resemble the training context. Far-transfer is when it carries over to situations that differ markedly in surface, tempo, resistance, or setting. It’s a gradient, not two boxes, but it decides which training is worth anything.

Blocked repetition under stable conditions builds a specific coupling between a perceived situation and a movement. It works until the situation changes. From an ecological standpoint, the student has calibrated against a narrow band of information. The affordance is bound to the context it was practiced in. Remove the context and the affordance disappears, because the student never learned to detect it in any other form.

That’s why someone who has done a hundred perfect blocks in the air can go completely blank when someone actually swings. The movement is there. The information that should trigger it isn’t there in a format they’ve ever practiced perceiving.

What transfers far isn’t the movement. It’s the coupling between perception and action, the ability to pick up relevant information and organize a solution from it. That gets trained through variability in the task, different partners, tempos, distances, surfaces, fatigue levels, so the student is forced to find what’s invariant instead of memorizing a surface. Through representative design, so training contains the same information sources as the real situation: resistance that actually resists, timing that isn’t predictable, consequences that resemble the real ones. And through exploration under constraint, so the student finds their own solutions inside the boundaries instead of copying one. What you found yourself is coupled to why it works. What you copied is coupled to what it looks like.

For civilian use, far-transfer is everything. The person who gets attacked never gets the situation they trained for. For professional application it’s sharper still, with fatigue, equipment, poor light, and legal constraints stacked on top. Training that only produces near-transfer produces people who are good at training.

Near-transfer isn’t worthless. Early on, a narrower task can help a student find a first working solution. But it should be a temporary constraint that gets loosened the moment the solution appears, not a permanent state. The problem with classical training isn’t that it starts narrow. It’s that it never widens.

Reductionism: the part is not the whole

Reductionism in motor learning is the assumption that a complex skill can be broken into parts, that the parts can be trained separately, and that the whole appears when the parts are reassembled. It’s a model borrowed from machines and classical physics. Movement in sport isn’t a machine. It’s a self-organizing system, and there the summation doesn’t hold.

Bernstein pointed at this in the thirties: the same muscle command gives a different movement depending on body position, velocity, and external force. The part has no fixed meaning outside its context. A step in isolation isn’t the same step as a step under pressure against an opponent, because the forces shaping it are completely different. Train the part and you get the part.

Reductionism shows up in three forms. Technique reductionism splits the movement into phases practiced separately, so the student learns five pieces and then has to link them under time pressure, when the linking was the hard part all along. Component reductionism trains physical qualities in isolation, strength here, mobility there, technique somewhere else, on the assumption they add up. Strength without perceptual coupling is strength you can’t find at the right moment. Cognitive reductionism treats perception and action as separate steps, first perceive, then decide, then act. Gibson showed it’s a circular system where movement creates the information that guides the next movement. Split it and you’re training a process that doesn’t exist.

Reductionism survives because it looks like it works. Students who are good at the parts are often good at the whole. Same confounding as before: exposure hours and selection drive both. And when part-training does improve the whole, it’s usually because the part still carried enough of the task’s information. Hitting pads with a partner who moves isn’t an isolated part. It’s a simplified whole. It works, but not for the reasons the reductionist thinks.

The alternative is simplification, not decomposition. The distinction is decisive. Simplification keeps the task’s structure and scales the difficulty: less space, slower resistance, fewer options. The perception-action coupling stays intact throughout. Decomposition removes the structure and hopes to reinstall it later.

Reductionism does have a place. Analysis. Understanding which constraints shape a movement requires thinking in parts, biomechanics, force direction, timing. That’s a tool for the coach to design tasks with. The error is letting the analytical model become the training model. That you can describe a movement in parts doesn’t mean it’s learned in parts.

The same mistake from two directions

Reductionism and the correlation trap are one error seen from two sides. Reductionism assumes parts can be isolated and summed. The correlation mistake assumes co-variation reveals cause. Both rest on the system being linear and separable, on being able to hold everything else constant and study one variable at a time. In a self-organizing system, that’s exactly what you can’t do.

When you split a movement into phases and train them separately, you also manufacture a story about cause: correct hip rotation produces a correct strike. That story is almost always a correlation promoted to causation. The elite have a certain hip rotation, therefore it causes the performance. But the rotation could be a consequence of the strike landing hard, not the cause. Or both could be driven by thousands of hours. Reductionism needs causal chains to function as a model, and it finds them by reading cause into patterns that are only co-variation.

What’s missing from both is intervention. Reductionism pretends to intervene, train the part, see the whole improve, but without control. The student who gets part-training also gets time, attention, exposure, and expectation. The improvement could come from any of them.

In a dynamic system, cause is rarely an arrow from A to B. It’s circular. Movement creates information, information shapes movement. Constraints shape the outcome without any single constraint causing it. Asking which part caused the performance is asking the wrong question. The right question is which constraints make a working solution likely. That’s still a causal question, but at the system level, not the part level.

Three checks before trusting a pattern: Have I seen co-variation, or have I actually changed something and measured the effect? If I changed something, what else changed at the same time? Am I measuring at the level where the cause operates, task and constraint, or at the part level where the pattern looks clean but explains nothing?

Reductionism makes instructors good at seeing patterns and bad at testing them. Clean correlations feel like knowledge.

The link at the bottom: interactionism versus transactionism

Everything above rests on one question. What is a human in motion? There are two answers, and they don’t mix.

Interactionism says the performer is one thing and the environment is another. They exist independently and interact through cause and effect. The environment sends stimuli, the performer processes, decides, and sends out movement. It’s Descartes with reflex arcs. The model needs an interface between inside and outside, and that interface is where representations, mental models, and motor programs get placed.

The consequences follow directly. If the performer is separate from the environment, skill can be stored inside the performer, so it can be trained in isolation and carried around. That’s reductionism. If cause runs from environment to performer and back in discrete steps, each step can be studied alone, so you can hunt for linear causal chains between parts. That’s the correlation trap. If perception and action are two processes, they can be trained separately. That’s cognitive reductionism.

Interactionism isn’t stupid. It’s intuitive, it matches how we talk about ourselves, and it yields manageable models. The problem is the models are manageable because they’ve cut away what actually happens.

Transactionism says performer and environment aren’t two things that meet. They’re aspects of a single event. Dewey and Bentley named it in 1949, Gibson built the perception, Kelso and Kugler built the dynamics. The affordance exists neither in the environment nor in the performer but in the relation. A door isn’t openable in itself. It’s openable for a body with a certain reach and strength. Remove the body and the property disappears.

That means skill can’t be stored inside the performer, because it isn’t there. It lives in the coupling. What you carry away from training isn’t a movement but a calibration, an ability to find the same relation in a new form. That’s where far-transfer comes from, and it’s why isolated part-training fails: it severs the relation the skill consists of.

Cause looks different in the two worlds. In interactionism it’s linear and directed, A causes B, which is why it hunts for links between parts and reads co-variation as cause. It has no other concept of cause to offer. In transactionism it’s circular and constraining. No part causes the outcome. Constraints make some outcomes likely and others unlikely, and the system finds stability inside that loop. Still causation, but at the system level, and tested by changing constraints and watching what emerges, not by isolating parts and measuring them.

That’s why the correlation question gets a different answer. The interactionist asks which part causes the performance and finds correlations. The transactionist asks which constraints make the performance possible and tests them by intervention. Only the second question has an answer that holds.

Where dualism leaks into practice

Most coaches who say they work with CLA still think interactionistically without noticing. It shows in the language.

“She has good technique” places the skill in the person. “Build the fundamentals first” assumes parts can be stacked in layers. “Read the situation and pick the right technique” describes perception, decision, and action as three steps. “Under pressure the technique disappears” assumes the technique was somewhere and left, instead of never having been coupled to that information in the first place.

Each of those is a dualist assumption. Each produces training that looks good in form and transfers badly.

The whole chain

Interactionism gives separable entities. Separable entities give reductionism. Reductionism gives linear causal chains between parts. Linear chains get read into correlations. The correlations confirm form training. Form training gives near-transfer. Near-transfer looks like skill in the room and vanishes outside it.

Change the ontology and the whole chain changes. Transactionism gives relational skill. Relational skill demands representative tasks. Representative tasks are trained through constraints. Constraints are tested through intervention. Intervention gives causal knowledge at the right level. The right level gives far-transfer.

This isn’t a question of method. It’s a question of what you believe a human in motion is. Answer that first, and the training design mostly writes itself. Skip it, and you’ll keep polishing form and wondering why it doesn’t hold.