Chapter map

1. 1. Candidate range
2. 2. Observed dispersion
3. 3. Reference model
4. 4. Reference means
5. 5. Gap at one
6. 6. Gap at two
7. 7. Gap at three
8. 8. Reference SD
9. 9. Adjusted error
10. 10. One-SE boundary
11. 11. Selection
12. 12. Scaling
13. 13. Reference rotation
14. 14. Simulation count
15. 15. Algorithm
16. 16. Claim

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Candidate range

A small example makes the mechanism visible. Evaluate k=1,2,3.

A concrete situation

Use the same observations and preprocessing.

What the mathematics does not license

A narrow range can force the answer. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Declare range. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Observed dispersion

This is where modelling choices begin to affect evidence. Record log W as 4.0,2.5,2.2.

A concrete situation

Compactness improves as k rises.

What the mathematics does not license

Raw W normally decreases with k. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Keep logs. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. Reference model

The central mathematical idea is concrete. Generate four null datasets per k.

A concrete situation

Preserve the declared sampling geometry.

What the mathematics does not license

The null distribution is a modelling choice. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Name geometry. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Reference means

Here the geometry becomes operational. Use 4.2,3.4,3.0.

A concrete situation

These are means of reference log W.

What the mathematics does not license

Average logs, not unlogged dispersions silently. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Show simulations. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Gap at one

This step deserves a careful pause. Compute 4.2-4.0.

A concrete situation

Gap1=.2.

What the mathematics does not license

Small positive difference is weak evidence. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record value. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Gap at two

The useful question is not only how to calculate, but what the calculation assumes. Compute 3.4-2.5.

A concrete situation

Gap2=.9.

What the mathematics does not license

This is the largest example gap. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record value. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Gap at three

A small example makes the mechanism visible. Compute 3.0-2.2.

A concrete situation

Gap3=.8.

What the mathematics does not license

A smaller gap can still be within simulation error. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record value. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. Reference SD

This is where modelling choices begin to affect evidence. Use SD=.10 at k=3.

A concrete situation

This measures Monte Carlo spread of log W.

What the mathematics does not license

Four replicates give a noisy SD. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Increase B. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Adjusted error

The central mathematical idea is concrete. Compute .10sqrt(1+1/4).

A concrete situation

s3=.1118.

What the mathematics does not license

Include the finite-reference correction. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Show factor. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. One-SE boundary

Here the geometry becomes operational. Compute Gap3-s3.

A concrete situation

.8-.1118=.6882.

What the mathematics does not license

The rule favours parsimony. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Keep threshold. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Selection

This step deserves a careful pause. Compare Gap2=.9 with .6882.

A concrete situation

Select k=2.

What the mathematics does not license

Use the smallest satisfying k. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

State rule. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Scaling

The useful question is not only how to calculate, but what the calculation assumes. Rescale one feature.

A concrete situation

Observed and reference dispersions change.

What the mathematics does not license

Train-consistent scaling is essential. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Document transform. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Reference rotation

A small example makes the mechanism visible. Rotate correlated features before bounding-box simulation.

A concrete situation

The null geometry may change.

What the mathematics does not license

Axis-aligned boxes can waste volume. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Run sensitivity. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Simulation count

This is where modelling choices begin to affect evidence. Increase B.

A concrete situation

Means and standard errors stabilise.

What the mathematics does not license

Small B can change selection. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report B. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. Algorithm

The central mathematical idea is concrete. Replace k-means with hierarchical clustering.

A concrete situation

Wk and gaps can change.

What the mathematics does not license

Gap is not algorithm free. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Name method. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Claim

Here the geometry becomes operational. Name null, B, W, algorithm and k range.

A concrete situation

Teaching selection k=2.

What the mathematics does not license

Do not claim natural truth. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Bound conclusion. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

The gap statistic compares the compactness of an observed clustering with compactness expected from data generated under a declared reference distribution with no comparable cluster structure. For each candidate number of clusters k, it subtracts the observed logarithm of within-cluster dispersion from the average reference logarithm. A large gap means the observed groups are more compact than the reference expects.

Consider candidate k values 1, 2 and 3. Suppose the observed log dispersions are 4.0, 2.5 and 2.2, while four reference datasets give mean log dispersions 4.2, 3.4 and 3.0. The gaps are therefore .2, .9 and .8. The largest raw gap occurs at k=2, but the published selection rule also carries Monte Carlo uncertainty forward.

If the standard deviation of the four reference log dispersions at k=3 is .10, the adjusted simulation error is s3=.10sqrt(1+1/4)=.1118. Choose the smallest k satisfying Gap(k) at least Gap(k+1)-s(k+1). Since .9 is at least .8-.1118, the rule selects two clusters. This is a one-standard-error style choice, not simply an instruction to take the largest number.

The answer depends on the clustering algorithm, dispersion definition, feature scaling, candidate range, number of reference simulations and reference geometry. A bounding box in correlated high-dimensional data may be a poor null; rotation or principal-component alignment can matter. The statistic estimates a useful cluster count under stated choices. It does not prove that clusters are stable, semantic or valuable.

---

Frequently asked questions

What is the gap statistic?

A comparison between observed log within-cluster dispersion and its mean under a declared reference distribution.

Why use logarithms?

The method compares dispersion on a multiplicative scale and follows the published formulation.

How is k selected?

Choose the smallest k whose gap is within the simulation-adjusted error of the next candidate under the stated rule.

What is the reference distribution?

A null data-generating geometry, often uniform over a transformed or axis-aligned bounding region, which must be justified.

Does the method prove a true k?

No. It estimates a useful count under chosen data, algorithm, dispersion, null and candidate range.

What must be reported?

Features, scaling, algorithm, Wk definition, k range, reference geometry, simulation count, gaps, errors, rule and stability.

---

Primary source and fact-check note

Robert Tibshirani, Guenther Walther and Trevor Hastie, “Estimating the Number of Clusters in a Data Set via the Gap Statistic,” JRSS Series B 63(2), 411–423 (2001), DOI 10.1111/1467-9868.00293. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• Calinski–Harabasz variance ratio
• Silhouette coefficient and cluster count
• Davies–Bouldin index
• Feature scaling and distance geometry
• How Mathematics Works: Statistics

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: clustering evidence and responsible use

Data audit

Freeze objects, features, units, exclusions, missing values, duplicates, transformations and random seeds.

Geometry audit

Document metric, scaling, linkage, sampling boundary, null reference and every pairwise or nearest-neighbour construction.

Formula audit

Show dispersion logs, Monte Carlo errors, powered distances, cophenetic pairs, co-membership counts and denominators.

Convention audit

Name gap rule, Hopkins orientation and exponent, correlation definition, classifier, threshold and weakest-cluster rule.

Uncertainty audit

Repeat reference simulations, random points, splits, seeds, bootstrap samples and plausible cluster counts.

Failure audit

Test outliers, duplicates, tiny clusters, high dimensions, correlated features, ties, leakage and boundary choices.

Meaning audit

Compare numerical tendency or stability with domain evidence without assuming groups are natural or fair.

Final claim

State exact tendency, fidelity, compactness or reproducibility evidence and stop before causal or guaranteed-usefulness claims.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Candidate range

Begin with this clustering mechanism: Evaluate k=1,2,3. Recreate the concrete case: Use the same observations and preprocessing. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A narrow range can force the answer. Complete the transfer step: Declare range. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Observed dispersion

Begin with this clustering mechanism: Record log W as 4.0,2.5,2.2. Recreate the concrete case: Compactness improves as k rises. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Raw W normally decreases with k. Complete the transfer step: Keep logs. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Reference model

Begin with this clustering mechanism: Generate four null datasets per k. Recreate the concrete case: Preserve the declared sampling geometry. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The null distribution is a modelling choice. Complete the transfer step: Name geometry. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Reference means

Begin with this clustering mechanism: Use 4.2,3.4,3.0. Recreate the concrete case: These are means of reference log W. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Average logs, not unlogged dispersions silently. Complete the transfer step: Show simulations. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 5: Gap at one

Begin with this clustering mechanism: Compute 4.2-4.0. Recreate the concrete case: Gap1=.2. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Small positive difference is weak evidence. Complete the transfer step: Record value. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Gap at two

Begin with this clustering mechanism: Compute 3.4-2.5. Recreate the concrete case: Gap2=.9. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: This is the largest example gap. Complete the transfer step: Record value. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Gap at three

Begin with this clustering mechanism: Compute 3.0-2.2. Recreate the concrete case: Gap3=.8. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A smaller gap can still be within simulation error. Complete the transfer step: Record value. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Reference SD

Begin with this clustering mechanism: Use SD=.10 at k=3. Recreate the concrete case: This measures Monte Carlo spread of log W. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Four replicates give a noisy SD. Complete the transfer step: Increase B. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 9: Adjusted error

Begin with this clustering mechanism: Compute .10sqrt(1+1/4). Recreate the concrete case: s3=.1118. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Include the finite-reference correction. Complete the transfer step: Show factor. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: One-SE boundary

Begin with this clustering mechanism: Compute Gap3-s3. Recreate the concrete case: .8-.1118=.6882. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The rule favours parsimony. Complete the transfer step: Keep threshold. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Selection

Begin with this clustering mechanism: Compare Gap2=.9 with .6882. Recreate the concrete case: Select k=2. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Use the smallest satisfying k. Complete the transfer step: State rule. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Scaling

Begin with this clustering mechanism: Rescale one feature. Recreate the concrete case: Observed and reference dispersions change. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Train-consistent scaling is essential. Complete the transfer step: Document transform. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 13: Reference rotation

Begin with this clustering mechanism: Rotate correlated features before bounding-box simulation. Recreate the concrete case: The null geometry may change. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Axis-aligned boxes can waste volume. Complete the transfer step: Run sensitivity. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Simulation count

Begin with this clustering mechanism: Increase B. Recreate the concrete case: Means and standard errors stabilise. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Small B can change selection. Complete the transfer step: Report B. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Algorithm

Begin with this clustering mechanism: Replace k-means with hierarchical clustering. Recreate the concrete case: Wk and gaps can change. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Gap is not algorithm free. Complete the transfer step: Name method. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this clustering mechanism: Name null, B, W, algorithm and k range. Recreate the concrete case: Teaching selection k=2. Write objects, features, candidate k, observed log dispersions, reference replicate logs, means, standard deviations, adjusted errors, gaps and selected k with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not claim natural truth. Complete the transfer step: Bound conclusion. Change exactly one feature scale, cluster count, null boundary, simulation replicate, exponent, random location, nearest neighbour, distance, linkage, merge height, test split, classifier label, cluster size, threshold or random seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep tendency, compactness, tree fidelity, reproducibility, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a uniform cloud, a correlated box, one outlier, tied distances, a tiny cluster or leaked preprocessing. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.