The Experience Fidelity Model

A few years ago I sat through a vendor demo for a compliance course that had clearly cost a great deal of money. It recreated the company’s actual software down to the button placement. The screens responded, the avatars moved, and the whole thing was almost indistinguishable from the system the learners used every day. Everyone in the room was impressed. As far as I could tell, it taught almost nothing, because every decision it asked of the learner came down to clicking the highlighted box to continue.

It took me a long time to work out why that bothered me as much as it did. This model is what I eventually arrived at.

Think about how we judge a speaker. A good one is judged on one thing: how faithfully it reproduces the original recording. Not how it looks, not what it cost, not how many drivers it has. A beautiful speaker playing the wrong song is no use to anyone. A learning experience is a reproduction too. It stands in for some part of reality — a sales call, a clinical decision, a difficult conversation — and tries to prepare someone for the real version. So the question I want to ask of any learning experience is the one I would ask of a speaker: how faithful is it to its source?

That is what the Experience Fidelity Model is for. And the thing it keeps surfacing is uncomfortable. A lot of what our profession builds looks exactly like the workplace while reproducing almost none of the judgement the work actually requires. It is faithful to the wrong thing.

The problem with “experiential”

The word did not start out vague. When Kolb set out experiential learning in 1984, building on Dewey and Lewin before him, he meant something specific: a cycle in which a concrete experience is followed by reflection, then by forming a concept, then by testing that concept in action. The experience was the start of the loop, not the whole of it. What mattered was what the learner did with it.

Four decades on, the term has stretched until it covers a multiple-choice quiz with consequences and a full-motion flight simulator equally well. Ask ten learning designers what experiential learning means and you will get ten answers: branching scenarios, gamification, simulations, learning by doing. That vagueness is convenient. It lets us present passive content as practice and promise capability we do not deliver, without ever answering the question that would expose the gap: how much of the real thing is actually in here? Dewey’s warning that not all experience is educative, and that some of it is actively miseducative, has mostly been forgotten.

I built the model to make that question answerable. The first thing it does is separate the part of fidelity that builds capability from the part that only looks as though it does.

What the model measures

Most frameworks in this area treat every dimension as the same kind of thing, stacked on one scale. This one does not, because the dimensions are not the same kind of thing.

Two of them build capability. Decision Fidelity asks whether the judgements the learner makes are the same as the real ones. Learning Responsiveness asks whether anything responds when the learner acts, and whether they can learn from it, including from their mistakes. These two are where learning is or is not made. I think of them together as the capability plane.

The third dimension does something else. Surface Fidelity, meaning whether it looks, feels and behaves like the real environment, does not build capability at all. It governs two other things. It governs cost, because it is the most expensive dimension to raise, and it governs transfer, meaning how well the capability you built holds up when the learner moves into the real environment. It is not a peer of the other two. It sits over the capability plane and modifies it. It changes what your learning costs to produce and how far it travels, but it is not itself a source of learning.

This is the part that took me longest to see clearly, so I will put it plainly. Two things make the learning: real decisions, and a real response to them. A third, surface fidelity, decides how far that learning travels and what it costs to build. Most of the failed e-learning I have seen pays for the third and never invests in the first two. The problem is not that it scored low on one axis of three. The problem is that it paid for the transfer of a capability that was never built.

The capability plane

Decision Fidelity: are the judgements real?

This dimension builds capability, and it does not care about production value. It asks whether the cognitive work the learner is doing matches the work the real task demands. A learner can sit inside a flawless copy of their own environment and make no real decisions at all. Another can sit in front of plain text and wrestle with the same judgement an expert faces. The scale runs:

• None — no decisions; pure reception.
• Recall — decisions about facts, not the task. “Which of these is correct?” Memory, not judgement.
• Recognition — judging work done by others. “What is wrong here?” Building the evaluative frame without yet performing.
• Constrained — real task decisions, but chosen from a small authored set. Genuine judgement, bounded answer space. Where most branching scenarios live.
• Open — the learner constructs their own response rather than choosing from a list. The answer space is no longer limited to what the designer thought of.
• Real — the actual judgement of the real task, in full ambiguity, with competing factors and incomplete information.
• Reality — the real decision with real stakes. No longer a design problem.

Learning Responsiveness: can the learner act, err and learn?

This dimension turns activity into capability, and it is more than the quality of the feedback. It is about whether errors are dead ends or openings, whether the environment adapts, and whether the learner can try things, get them wrong, and find a way back. Being able to make a mistake, live with the consequence, and work out a way through is not a nice feature of good feedback. It is how skill forms.

• None — actions produce no response.
• Binary — right or wrong. Information, not learning.
• Explanatory — tells you why, then stops. Something to think with, but a dead end.
• Consequential — choices produce outcomes that play out. Cause and effect become visible.
• Recoverable — errors are not dead ends; the learner can navigate out of a mistake and experience the recovery. Real work is full of error and recovery.
• Adaptive — the environment responds to the learner’s pattern over time and adjusts.
• Emergent — feedback comes from the environment itself, unscripted. The learner can do what the designer never anticipated and still be answered meaningfully. The ceiling of digital responsiveness.
• Reality — feedback from the real world, in its full complexity.

Independent in theory, clustered in practice

The two axes are independent in principle. You can build a real example at any combination: high decision and no responsiveness (an essay marked with a single comment), no decision and high responsiveness (spaced-repetition flashcards), low on both (a recorded lecture), high on both (a case that adapts as you reason). Every combination is buildable.

In practice they travel together. You cannot respond in an open, unscripted way to a multiple-choice answer, so real builds cluster along a diagonal. That clustering is worth noticing rather than ignoring. Most e-learning ends up at Constrained decisions with Consequential responsiveness, because that band is the cheap one to build. The question is whether the cheap band is where your particular task actually belongs.

Surface fidelity: a modifier, not a third axis

Surface fidelity is the dimension that flatters everyone. It photographs well, it impresses stakeholders, it feels like progress. And most of the time it has very little to do with whether anyone learns anything. The scale runs from Abstract (text, numbers, diagrams) through Illustrated, Representational and Replicated to Immersive (VR, physical simulators), and then to Reality, the actual environment.

Surface earns its cost in one situation: when transfer depends on it. If the capability is perceptual or physical, like reading a real instrument, operating a real interface, or performing a technique with your hands, then the look and feel of the environment is part of the skill, and spending on surface is right. If the capability is judgement, like what to decide, what to say, or how to reason, then surface has almost nothing to do with whether that capability forms, and spending on it is waste. That is why I treat surface as a modifier rather than a third axis. It does not move you towards capability. It changes what your capability costs, and how well it survives the move into the real world.

The fidelity signature

Every learning experience has a signature: its position on the two capability axes, with its surface level as a multiplier. The capability core sits in brackets, and surface multiplies it.

[ Decision · Responsiveness ] × Surface

A few real ones. A compliance video with a tick-box is [Recall · Binary] × Illustrated. An anti-bribery branching scenario is [Constrained · Consequential] × Representational. A text-only AI roleplay coach is [Open · Adaptive] × Abstract. A scripted surgical VR module is [Constrained · Explanatory] × Immersive. A full-motion flight simulator is [Real · Emergent] × Immersive.

Put the surgical VR next to the AI roleplay. The VR has a much higher surface multiplier and a weaker capability core. The roleplay has almost no surface and a strong core. If the task is judgement, the cheap text tool builds more capability than the expensive simulator. Writing both as signatures makes that visible in a way a feature list never does.

The reality gap

On every dimension the top level is Reality, and it cannot be reached by design. Not because the technology is not good enough yet, but in principle. Once the decision carries real stakes, the environment is the real one, and the feedback comes from the world itself, the experience has stopped being a learning design and become the job.

I call the distance between the best experience you can build and the real thing the reality gap. No signature closes it. Even the flight simulator, which is about as far as digital fidelity goes, stops short. I think naming the gap matters, because it tells you where your job as a designer ends. It stops us promising capability we cannot actually deliver, and it changes the goal from reaching reality to building the most faithful reproduction the task warrants, and then handing off honestly to real practice.

Reading a signature

The first use of the model is to look backwards at something you have already built. Write a struggling experience as a signature. If it comes out as [Constrained · Consequential] × Immersive, the diagnosis is right there: a lot of money on surface, a middling capability core. If behaviour is not changing, more polish on the surface will not fix it, because the surface was never the problem. The learning would have to come from the capability core, and the capability core was left thin.

The signature also gives you something to say to a stakeholder, which I have found almost as useful as the diagnosis. You can say: we can build to this signature for this budget, and reaching this other signature costs this much more, and most of the difference is surface fidelity that this task does not need. That is a better conversation than arguing about whether something counts as experiential.

Prescribing a signature

The more useful version is to decide the target signature before you spend anything. This is where the model becomes a method, and the method has one discipline that most design conversations lack: deciding which dimension you are going to refuse to spend on.

For declarative tasks, like knowing a policy, a product, or a regulation, the goal is knowing, not doing. Aim for something like [Recall · Adaptive] × Abstract. Spend on responsiveness, because a good spaced-repetition loop keeps knowledge alive at scale. Refuse to spend on decision fidelity and surface. Building a realistic simulation to teach a policy fact is waste, and Abstract is the right answer, not a compromise.

For recognition tasks, like reading a scan, inspecting for defects, or spotting fraud, the skill is perceptual, and this is the rare case where the surface is the task. Aim for [Recognition · Adaptive] × Replicated. The stimulus has to look like the real thing, so spend on surface and responsiveness, and do not push decision fidelity past recognition. The learner is not building a plan; they are learning to see.

For judgement under ambiguity, like clinical reasoning, difficult conversations, leadership calls, or negotiation, aim for [Open · Adaptive-to-Emergent] × Abstract-to-Representational. Spend on decision fidelity, because the judgement has to be real and the learner’s own, and on responsiveness, because it has to answer what the learner actually did. Refuse to spend on surface. A lightly illustrated environment with a strong capability core will beat a beautiful branching video with a fixed menu, at a fraction of the cost. This is the category we most often over-build on surface and under-build on judgement.

For procedural and physical tasks, like operating a machine, a clinical technique, flying, or pulling a shot of espresso, aim for [Constrained-to-Real · Recoverable+] × Replicated-to-Immersive, and spend on all three. This is the one task type where high surface fidelity is genuinely correct, because transfer depends on the feel of the real interface, and the learner has to be able to fail and recover safely. There is nothing to refuse here, except the assumption that maxing all three closes the gap. It does not, because physical skill only finishes with real repetitions.

The discipline is in the refusal. A designer who can say we are staying at Abstract surface and spending everything on decision fidelity, because this is a judgement task, is using the model the way I intended.

The coffee test

Take making espresso. Not knowing about coffee, but making it: setting the grind, tamping by feel, reading the extraction, adjusting as you go. It is a procedural, physical task, so by the method above the surface matters here. Watch the signatures climb. A video of a barista is [None · None] × Representational. A match-the-grind-to-the-drink quiz is [Recall · Binary] × Illustrated. A branching “the shot ran too fast, what now?” is [Constrained · Consequential] × Representational. A drag-the-steps sim is [Recall · Explanatory] × Replicated. A VR espresso machine that responds realistically is [Constrained · Emergent] × Immersive.

For coffee, the immersive surface on that last one is justified, because the feel is part of the skill. This is not the expensive mistake. And yet [Constrained · Emergent] × Immersive, which is about as high as digital can reach, still does not teach anyone to make good coffee. The felt judgement of the tamp and the pour sits at [Real · Reality] × Reality, on the far side of the reality gap, and you only get there by pulling real shots. So the coffee test shows both halves of the model at once: surface fidelity spent correctly, and a gap that no amount of spending closes.

What AI changes

Large language model tools change what is reachable in the capability core, and barely touch the surface. An AI roleplay coach can sit at [Open · Adaptive], or even [Open · Emergent], while staying × Abstract. The learner constructs a real response and gets answered in the moment, in an environment that does not need to look like anything in particular. That matters, because rich responsiveness at real decision fidelity used to mean expensive custom development, which is a large part of why the industry settled on the cheap diagonal in the first place. AI lowers the cost of the capability core without paying the surface premium. For judgement tasks, that is the right place to spend.

The model also points at a limit, though. If the intelligence runs at design time and ships a fixed artefact that behaves the same for every learner, responsiveness is capped at about Recoverable: the learner can make a mistake and recover, but only down paths the designer anticipated. Emergent means responding to what the designer did not anticipate, which a fixed artefact cannot do. To reach Emergent, or to move decisions from Constrained to Open, the intelligence has to be present at runtime, responding live. That is a different kind of system with different costs, and writing it as a signature makes the trade-off explicit: two of the most valuable upgrades you can make both require putting the intelligence into the runtime.

Where this leaves me

To gather it up. The Experience Fidelity Model asks how faithfully a learning experience reproduces the reality it is meant to prepare someone for. Decision fidelity asks whether the judgements are real, and it builds capability. Learning responsiveness asks whether the learner can act, get things wrong, and learn, and it turns activity into capability. Surface fidelity asks whether it looks and feels real, and it does neither. It is a modifier that governs cost and transfer.

Every design has a signature: [ Decision · Responsiveness ] × Surface. Read backwards, it shows you where the budget went, which is usually surface bought instead of capability. Read forwards, it helps you pick a target for the task in front of you, and decide which axis to leave deliberately low. Above all of it sits the reality gap, the ceiling no experience reaches, where digital learning has to hand off to real practice.

I keep coming back to the speaker. You can build a beautiful one and still have it playing the wrong song. Most of the time, the song is the part we forget to check.

---

This is the fourth in a series of articles on the structural challenges facing the L&D profession. The previous pieces drew on other people’s frameworks: training evaluation, needs analysis, and manager support for transfer. This one puts forward a model of my own. If you use it, break it, or disagree with it, I would like to hear about it.