MindOS Learning Manual: Example-Generation State | Can You Invent a New Example That Really Fits?

Wait, What?

You can know the definition and still invent the wrong example.

A student can recite a concept perfectly, feel confident, and then produce an example that quietly violates the definition.

That failure is useful.

It tells us that remembering the words and applying the concept are not the same learning operation.

Quick Answer

Example generation asks the learner to take an abstract idea and construct a new concrete instance that genuinely satisfies it. Done well, this can strengthen concept learning and transfer because the learner must decide which features are essential. Done badly, it can reinforce a misconception—especially when the learner cannot judge whether the example is valid.

The operation is not:

Think of anything vaguely related.

It is:

Construct a new case, test it against the concept, and repair it if it does not fit.

Owned Learning Operation

EXAMPLE-GENERATION STATE = understand the concept → construct a novel instance → justify why it fits → test the boundary → revise.

This is distinct from Generation State, where the learner attempts an answer before seeing instruction. Here, the learner has already encountered the concept and must now instantiate it correctly.

It is also different from studying worked examples. A worked example is provided. Example generation requires the learner to build one.

Why This Is Harder Than It Looks

Abstract concepts usually contain a relationship, condition, rule or pattern that can appear in many surface forms. Generating an example requires the learner to keep the deep structure while changing the surface.

Suppose a student learns:

A variable is a factor that can change in an investigation.

The student writes:

The beaker is a variable because we use it in the experiment.

The sentence sounds scientific. It contains the target word. It may even feel right.

But the beaker is not a variable merely because it is present. The learner has attached the label to a related object without preserving the defining relationship.

That is exactly why example generation can be diagnostic as well as educational.

Four States of Example Generation

1. Echo Example

The learner changes a name or number but keeps almost all of the original example.

This can be a useful early scaffold, but it does not yet show that the concept can travel far.

2. Surface Example

The learner produces something that looks related but does not satisfy the defining structure.

This often exposes a misconception.

3. Valid Example

The learner changes the context while preserving the concept.

4. Boundary Example

The learner deliberately generates a case near the edge and can explain why it still belongs—or why it does not.

This is stronger because the learner is no longer merely matching a prototype. They are reasoning about the concept’s conditions.

The MindOS Example-Generation Protocol

Step 1 — State the Rule in a Testable Form

Before inventing an example, identify what must be true.

For Mathematics, that may be a structural condition. For English, it may be a language relationship. For Science, it may be a mechanism or variable relation.

Step 2 — Change the Surface

Use a different context, object, number, sentence, scenario or representation from the one used in teaching.

Step 3 — Justify the Fit

Finish the sentence:

This is an example because…

The explanation matters. It exposes whether the learner is using the defining property or a superficial cue.

Step 4 — Generate a Non-Example

Now construct a case that looks similar but fails one important condition.

This forces the boundary into view.

Step 5 — Check Against an External Standard

Use the definition, a trusted worked example, teacher feedback or an authoritative source. Do not assume that a plausible-looking self-generated example is correct simply because it was generated independently.

Step 6 — Return Later

Generate another example after a delay and in a different context. The operation should survive beyond the original prompt.

Worked Examples Across Subjects

English: Inference

Definition-level knowledge: an inference combines textual evidence with reasoning to reach a conclusion not stated directly.

Generated example:

The floor is wet, umbrellas are open by the door, and students are shaking water from their sleeves. I infer that it has been raining.

Why it fits: the conclusion is not explicitly stated; it is supported by several clues.

Non-example: “The passage says it is raining, so I know it is raining.” That is retrieval of stated information, not inference.

Mathematics: Direct Proportion

A learner should not merely swap apples for oranges in a textbook question. They should preserve the multiplicative relationship and be able to explain why a constant ratio is maintained.

Science: Fair Test

Ask the learner to invent a fair-test scenario different from the classroom experiment. Then ask which variable changes, which outcome is measured, and which relevant conditions must be controlled.

If the learner can invent the setting but cannot preserve variable control, the story changed while the concept disappeared.

Competing Explanations When Example Generation Fails

The learner may not understand the definition.
The learner may understand but lack relevant background knowledge to invent a sensible context.
The concept may be too abstract for unguided generation.
The learner may over-focus on surface similarity.
The learner may produce a valid example but explain it poorly.
The learner may be unable to judge the quality of their own example.

These are not the same problem. A student who lacks prior knowledge may need a provided example first. A student who generates plausible but invalid examples needs comparison against standards. A student who can generate only near-copies may need more varied cases.

How Do We Know?

The evidence is promising but not simple.

Rawson and Dunlosky examined example generation for declarative concepts across three experiments involving 487 undergraduates. Their combined results showed moderate learning benefits over restudy alone, while also showing that the size of the benefit depended on conditions such as how practice was scheduled.

A 2025 experimental study with student teachers found stronger retention and transfer when participants generated their own examples than when they studied provided examples or reread instruction only. But the same paper carefully reviews earlier studies with mixed findings. That history matters: example generation is not universally superior to provided examples.

Another important line of research shows a practical danger: learners are not always good at evaluating the quality of their own generated examples. Independent generation therefore should not automatically mean independent validation.

Useful sources:

Evidence Boundary

Most controlled research on example generation has used particular concepts, age groups and learning materials. Do not assume the same effect for every school subject or every learner. The operation is most defensible when the concept is understood well enough to instantiate, the learner can compare the example with an external standard, and the generated case is used to deepen—not replace—accurate instruction.

When Example Generation Is the Wrong Tool

When the learner does not yet understand the core concept.
When a novice needs a clear provided example before producing one.
When the concept has high-stakes safety implications and unsupervised invention could reinforce dangerous misconceptions.
When the actual weak link is retrieval, not application.
When the learner is repeatedly producing invalid examples and cannot yet evaluate them.

The Independence Test

Do not stop after one successful example.

Generate one example now.
Generate a non-example that differs by one defining feature.
Return tomorrow and generate another example in a different context.
Explain why both examples belong to the same concept.
Classify a new case someone else provides.

If the learner can only reproduce yesterday’s example, the operation has not yet become flexible.

Teaching Guide for Parents, Tutors and Teachers

A useful prompt sequence is:

“Tell me the rule.”
“Give me a new example.”
“Why does it fit?”
“Give me something that almost fits but does not.”
“What exact condition fails?”
“Can you make another example that looks completely different?”

At first, keep the definition visible. Later, remove it. Early support protects against practising misconceptions; later removal tests whether the concept can be reconstructed independently.

MindOS Direction

If the learner cannot state why the example fits: move to Explanation State.

If the learner confuses neighbouring concepts: move to Comparison State or a concept-boundary task.

If the learner’s example is valid but knowledge disappears later: use Retrieval State and spacing.

If examples are accurate but do not travel to new problems: continue to Transfer State.

MindOS rule: a self-generated example is valuable not because the learner invented it, but because the learner preserved the concept while changing the world around it.