Chapter map

1. 1. Target
2. 2. Variance input
3. 3. Confidence level
4. 4. Tuning
5. 5. Residual
6. 6. Scaled residual
7. 7. Central score
8. 8. Saturated score
9. 9. Balance equation
10. 10. Root solver
11. 11. Mean comparison
12. 12. Median comparison
13. 13. Deviation claim
14. 14. Dependence check
15. 15. Scale sensitivity
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. Target

Here the geometry becomes operational. Define the population mean and sampling unit.

A concrete situation

Estimate the mean of 1,2,3,4,20.

What the mathematics does not license

Dependence changes the deviation guarantee. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Audit design. 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. Variance input

This step deserves a careful pause. Provide a valid variance bound or robust estimate.

A concrete situation

Separate the variance input from the five values.

What the mathematics does not license

Understating variance makes tuning overconfident. The same reasoning helps students move from an exercise to a real dataset.

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.

---

3. Confidence level

The useful question is not only how to calculate, but what the calculation assumes. Choose δ before examining results.

A concrete situation

Use one stated tail probability.

What the mathematics does not license

Trying many δ values changes interpretation. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Predefine reporting. 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. Tuning

A small example makes the mechanism visible. Calculate α from n, variance information and δ or state an illustrative value.

A concrete situation

Use α=.2 only for the worked mechanism.

What the mathematics does not license

α has inverse data units. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Label illustration. 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. Residual

This is where modelling choices begin to affect evidence. Form r_i=x_i−θ.

A concrete situation

At θ≈3.40482 residuals range from about −2.405 to 16.595.

What the mathematics does not license

Residuals depend on the candidate root. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Keep signs. 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. Scaled residual

The central mathematical idea is concrete. Compute u_i=αr_i.

A concrete situation

The largest scaled residual exceeds one.

What the mathematics does not license

Unit conversion requires reciprocal α conversion. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Check dimensions. 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. Central score

Here the geometry becomes operational. Use the logarithmic score for |u|<1.

A concrete situation

Four observations remain inside.

What the mathematics does not license

Do not replace the formula with hard clipping. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Show values. 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. Saturated score

This step deserves a careful pause. Use sign(u)log2 beyond the stated boundary.

A concrete situation

The value 20 contributes +.69315.

What the mathematics does not license

Saturation limits score, not data legitimacy. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Keep provenance. 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. Balance equation

The useful question is not only how to calculate, but what the calculation assumes. Sum every ψ(u_i).

A concrete situation

The total is approximately zero at θ≈3.40482.

What the mathematics does not license

Rounding can move a small residual sum. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Keep precision. 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. Root solver

A small example makes the mechanism visible. Bracket and solve a monotone equation.

A concrete situation

Bisection reproduces the teaching root.

What the mathematics does not license

A failed bracket is an implementation problem. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record tolerance. 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. Mean comparison

This is where modelling choices begin to affect evidence. Calculate the empirical mean separately.

A concrete situation

The ordinary mean is 6.

What the mathematics does not license

Difference alone does not identify contamination. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Explain mechanism. 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. Median comparison

The central mathematical idea is concrete. Calculate the median separately.

A concrete situation

The median is 3.

What the mathematics does not license

Catoni is not a renamed median. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Compare losses. 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. Deviation claim

Here the geometry becomes operational. Match the theorem’s assumptions and α rule.

A concrete situation

Finite variance can support Gaussian-order concentration.

What the mathematics does not license

The guarantee is probabilistic, not casewise. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

State scope. 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. Dependence check

This step deserves a careful pause. Inspect time, cluster or repeated-measure structure.

A concrete situation

Five independent units is a different design from five readings of one sensor.

What the mathematics does not license

Independence cannot be inferred from a list. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Model design. 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. Scale sensitivity

The useful question is not only how to calculate, but what the calculation assumes. Repeat with plausible variance inputs and α values.

A concrete situation

Track movement toward mean or saturation.

What the mathematics does not license

Cherry-picked tuning can manufacture stability. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Show curve. 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

A small example makes the mechanism visible. State a tuned log-truncated mean with assumptions.

A concrete situation

The score-balanced centre is about 3.40482 in the illustration.

What the mathematics does not license

Do not promise universal superiority. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Keep limitations. 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 Catoni mean replaces the ordinary sum of residuals with a slowly growing or bounded score. A common bounded choice is odd and equals −log(1−|u|+u²/2) times sign(u) while |u|<1, then saturates at sign(u)log 2. The estimate θ solves the balance equation Σψ(α(x_i−θ))=0. The scale α decides how quickly observations enter the saturated region.

For 1,2,3,4,20 with illustrative α=.2, the bounded-score equation has a root θ≈3.40482. At that root the five scores are approximately −.45460,−.27640,−.08087,.11873 and .69315, which sum to zero. The ordinary mean is 6 and the median is 3. The Catoni value is a separate estimator whose position follows from score balance, not an average with hidden deletion.

Catoni’s 2012 result is valuable because it obtains Gaussian-order deviation control for suitably tuned estimators under finite-variance assumptions, including heavy-tailed settings where the empirical mean can have poor worst-case deviation. That guarantee depends on independent sampling, the chosen confidence level, a valid variance bound or estimate, and the specified influence function. It is not a promise that every realized estimate is closer to truth.

The tuning parameter is part of the estimand-producing procedure. A smaller α keeps more scores nearly linear and closer to the sample mean; a larger α saturates residuals sooner. Reusing α across different units is invalid because α multiplies a measured residual. Report the score function, α rule, variance input, confidence level, solver, stopping tolerance and any sample splitting.

---

Frequently asked questions

What does the Catoni estimator solve?

A nonlinear balance equation in which transformed residual scores sum to zero.

Why log-truncate?

The score grows much more slowly than a raw residual, limiting tail domination.

What does α control?

How quickly scaled residuals leave the nearly linear region.

Does it need variance information?

The theoretical tuning commonly uses a variance bound or defensible variance estimate and a confidence level.

Is it always better than the mean?

No. Performance depends on the distribution, tuning, assumptions and decision target.

What must be reported?

Score formula, α rule, variance input, confidence level, solver, tolerance and sampling design.

---

Primary source and fact-check note

Olivier Catoni, “Challenging the Empirical Mean and Empirical Variance: A Deviation Study,” Annales de l’Institut Henri Poincaré, Probabilités et Statistiques 48(4), 1148–1185 (2012), DOI 10.1214/11-AIHP454. 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

• Rousseeuw–Croux Qn robust scale
• Rousseeuw–Croux Sn robust scale
• How Robust Statistics Work
• 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: topic-specific validation and responsible use

Design audit

Freeze observations, sampling units, dependence, missingness and the target parameter.

Scale audit

State the residual scale, its units, its robustness and whether it is recomputed.

Score audit

Show every standardized residual, threshold, score, weight and contribution.

Tuning audit

Connect confidence tuning or rejection constants to a stated model and goal.

Solver audit

Record initialization, bracketing or IRLS updates, tolerance, objective and all plausible roots.

Graph audit

Plot raw data, residuals, score curves, weights and start-to-solution paths.

Sensitivity audit

Change one variance input, scale, tuning constant, starting fit or cluster at a time.

Final claim

Keep the conclusion inside the sampling design and the protection the selected score actually supplies.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Target

Begin with this mechanism: Define the population mean and sampling unit. Recreate the concrete case: Estimate the mean of 1,2,3,4,20. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Dependence changes the deviation guarantee. Complete the transfer step: Audit design. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Variance input

Begin with this mechanism: Provide a valid variance bound or robust estimate. Recreate the concrete case: Separate the variance input from the five values. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Understating variance makes tuning overconfident. Complete the transfer step: Run sensitivity. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Confidence level

Begin with this mechanism: Choose δ before examining results. Recreate the concrete case: Use one stated tail probability. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Trying many δ values changes interpretation. Complete the transfer step: Predefine reporting. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Tuning

Begin with this mechanism: Calculate α from n, variance information and δ or state an illustrative value. Recreate the concrete case: Use α=.2 only for the worked mechanism. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: α has inverse data units. Complete the transfer step: Label illustration. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 5: Residual

Begin with this mechanism: Form r_i=x_i−θ. Recreate the concrete case: At θ≈3.40482 residuals range from about −2.405 to 16.595. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Residuals depend on the candidate root. Complete the transfer step: Keep signs. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Scaled residual

Begin with this mechanism: Compute u_i=αr_i. Recreate the concrete case: The largest scaled residual exceeds one. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Unit conversion requires reciprocal α conversion. Complete the transfer step: Check dimensions. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Central score

Begin with this mechanism: Use the logarithmic score for |u|<1. Recreate the concrete case: Four observations remain inside. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not replace the formula with hard clipping. Complete the transfer step: Show values. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Saturated score

Begin with this mechanism: Use sign(u)log2 beyond the stated boundary. Recreate the concrete case: The value 20 contributes +.69315. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Saturation limits score, not data legitimacy. Complete the transfer step: Keep provenance. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 9: Balance equation

Begin with this mechanism: Sum every ψ(u_i). Recreate the concrete case: The total is approximately zero at θ≈3.40482. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Rounding can move a small residual sum. Complete the transfer step: Keep precision. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Root solver

Begin with this mechanism: Bracket and solve a monotone equation. Recreate the concrete case: Bisection reproduces the teaching root. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A failed bracket is an implementation problem. Complete the transfer step: Record tolerance. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Mean comparison

Begin with this mechanism: Calculate the empirical mean separately. Recreate the concrete case: The ordinary mean is 6. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Difference alone does not identify contamination. Complete the transfer step: Explain mechanism. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Median comparison

Begin with this mechanism: Calculate the median separately. Recreate the concrete case: The median is 3. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Catoni is not a renamed median. Complete the transfer step: Compare losses. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 13: Deviation claim

Begin with this mechanism: Match the theorem’s assumptions and α rule. Recreate the concrete case: Finite variance can support Gaussian-order concentration. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The guarantee is probabilistic, not casewise. Complete the transfer step: State scope. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Dependence check

Begin with this mechanism: Inspect time, cluster or repeated-measure structure. Recreate the concrete case: Five independent units is a different design from five readings of one sensor. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Independence cannot be inferred from a list. Complete the transfer step: Model design. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Scale sensitivity

Begin with this mechanism: Repeat with plausible variance inputs and α values. Recreate the concrete case: Track movement toward mean or saturation. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Cherry-picked tuning can manufacture stability. Complete the transfer step: Show curve. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this mechanism: State a tuned log-truncated mean with assumptions. Recreate the concrete case: The score-balanced centre is about 3.40482 in the illustration. Write observations, residuals, scaled residuals, logarithmic scores, variance inputs, confidence probabilities, tuning parameters, score sums, roots and original measurement units with units, signs, thresholds and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not promise universal superiority. Complete the transfer step: Keep limitations. Change exactly one observation, sampling unit, variance input, confidence level, residual scale, tuning parameter, starting location, score boundary, cluster, leverage value, solver tolerance, resampling rule or covariance formula. Explain which mathematical object changes, which root is selected, which conclusion survives and which claim must weaken. Keep design, standardization, score, weight, objective, solver, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the score table and once to a parent with the real decision and original units. Add a counterexample in which the same robust estimate comes from rounding, censoring, dependence, mixture structure, leverage or genuine rare events. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.