WHY MATHEMATICS?

Why Mathematics? | CaSTLe for Transfer Learning: Mutual Information, Ordinal Binning, XGBoost Trees and Cross-Dataset Cell Labels

How CaSTLe engineers a shared source–target feature space, converts platform-sensitive expression values into ordinal bins and transfers cell labels through a regularised boosted-tree ensemble.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Intersect source and target genes without using target labels
Select informative, measurable and nonredundant features
Map expression into four common ordinal bins
Train XGBoost on the source and validate transfer on a separate target

Chapter map

1. Why transfer is difficult
2. Shared-gene intersection
3. Rare-expression filter
4. High-mean features
5. Mutual information
6. Union of feature views
7. Correlation pruning
8. Four ordinal bins
9. Why bins can transfer
10. Boosted decision trees
11. Regularised objective
12. Multiclass transfer
13. Binary threshold route
14. Label-blind target access
15. A student boosting project
16. Governance and careers

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

1. Why transfer is difficult

Here the geometry becomes operational. Protocols alter depth, dropout and feature distributions even when biology is shared. A reusable classifier needs a common representation.

A concrete situation

One source has twice the reads per cell.

What the mathematics does not license

Simple scaling does not repair detection differences. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Compare source and target distributions before modelling. 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. Shared-gene intersection

This step deserves a careful pause. Features present in only one dataset cannot support a transferred rule, so the workflow starts from common genes.

A concrete situation

Two studies share 12,000 identifiers.

What the mathematics does not license

Identifier mismatches can masquerade as missing biology. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Version symbol and accession mapping. 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. Rare-expression filter

The useful question is not only how to calculate, but what the calculation assumes. Genes detected in fewer than a declared number of cells are removed because their empirical bins and information estimates are unstable.

A concrete situation

A gene appears in only nine cells.

What the mathematics does not license

Rare markers may be biologically valuable. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Audit lost known markers separately. 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. High-mean features

A small example makes the mechanism visible. Genes with large overall mean across source and target are measurable enough to support robust ordinal comparisons.

A concrete situation

A housekeeping-like gene passes the mean screen.

What the mathematics does not license

High expression need not be discriminative. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Keep mean and label information conceptually separate. 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.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Mutual information

This is where modelling choices begin to affect evidence. Mutual information measures how much knowing a binned gene reduces uncertainty about source labels.

A concrete situation

A gene’s bins strongly separate two types.

What the mathematics does not license

Biased estimates arise with sparse counts and small classes. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Calculate a two-class contingency table. 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. Union of feature views

The central mathematical idea is concrete. Combining high-mean and high-information genes balances measurability with discrimination.

A concrete situation

Some genes qualify through both lists.

What the mathematics does not license

Fixed top counts are modelling choices. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Plot performance versus feature count. 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. Correlation pruning

Here the geometry becomes operational. For highly correlated gene pairs, one feature is removed to reduce duplicated evidence and computation.

A concrete situation

Two ribosomal genes exceed 0.9 Pearson correlation.

What the mathematics does not license

Arbitrary removal can discard a more portable gene. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat pruning under resampling. 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. Four ordinal bins

This step deserves a careful pause. Continuous expression is mapped to [0], (0,1], (1,6] and (6,infinity) in the published workflow.

A concrete situation

A value of four enters the third bin.

What the mathematics does not license

Thresholds reflect the original scale and may age poorly. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Publish units and bin occupancy. 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. Why bins can transfer

The useful question is not only how to calculate, but what the calculation assumes. Broad ranges reduce sensitivity to modest scale shifts and turn each gene into ordered categories.

A concrete situation

A source value 3 and target value 5 remain in one bin.

What the mathematics does not license

Large shifts still cross boundaries. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Measure per-feature bin drift. 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. Boosted decision trees

A small example makes the mechanism visible. XGBoost builds trees sequentially, each responding to gradients from the current classification loss.

A concrete situation

A later tree repairs errors among related types.

What the mathematics does not license

Boosting can learn source-specific artefacts. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Track training and held-out loss. 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. Regularised objective

This is where modelling choices begin to affect evidence. Loss is combined with penalties on tree complexity, leaf weights and minimum child evidence.

A concrete situation

A split is rejected when gain is too small.

What the mathematics does not license

Pre-tuned settings are not universally optimal. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Tune only inside source validation. 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. Multiclass transfer

The central mathematical idea is concrete. One model predicts among source types when the target is assumed to contain those same classes.

A concrete situation

Pancreas source labels a second pancreas study.

What the mathematics does not license

The assumption fails if a target type is absent from source. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Audit reference coverage first. 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. Binary threshold route

Here the geometry becomes operational. Separate one-versus-rest models can produce scores for each type; a cell is assigned only if the highest score clears a threshold.

A concrete situation

A low-score cell remains uncertain.

What the mathematics does not license

Several models can pass simultaneously. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Define tie and coexpression rules. 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. Label-blind target access

This step deserves a careful pause. Feature engineering may inspect target expression but must not inspect target labels. This is transductive unsupervised access, not supervised testing.

A concrete situation

Target means help define measurable genes.

What the mathematics does not license

Accidental label use invalidates transfer estimates. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Log every target-dependent step. 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. A student boosting project

The useful question is not only how to calculate, but what the calculation assumes. Bin two datasets, estimate mutual information, prune correlation and fit a small boosted tree ensemble.

A concrete situation

Known labels reveal scale robustness.

What the mathematics does not license

Real data add thousands of genes and dropout. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Withhold one class from the source. 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. Governance and careers

A small example makes the mechanism visible. Version datasets, genes, bins, information scores, correlations, split, XGBoost settings and thresholds. This joins information theory and ensemble learning.

A concrete situation

Students can begin with contingency tables and trees.

What the mathematics does not license

A transferred label proves no mechanism or diagnosis. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Write a transfer model card. 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

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

A CaSTLe audit begins with a labelled source dataset and an unlabelled target dataset. It intersects their genes, removes rarely expressed features, combines high-mean genes with genes carrying mutual information about source labels, removes highly correlated features, bins expression into four ordinal ranges and trains an XGBoost classifier on the transformed source.

Record source and target accessions, shared genes, rarity filter, mean-expression ranking, mutual-information calculation, correlation filter, bin edges, source split, XGBoost parameters, class scores and decision threshold. The workflow deliberately uses target feature distributions during engineering, so evaluation must keep target labels invisible and distinguish unsupervised access from label leakage.

Validate with entire studies, platforms and tissues held out, per-class sensitivity and specificity, accuracy, AUC, calibration, threshold coverage and unknown-type challenges. Transfer learning works only when source and target share enough biology and measurement structure; it cannot manufacture a missing cell type.

Frequently asked questions

What is CaSTLe?

CaSTLe is a transfer-learning workflow that engineers a shared source–target feature representation and trains an XGBoost cell-type classifier on labelled source cells.

How are features selected?

The published workflow combines genes with high overall mean expression and genes with high mutual information with source labels, then removes strongly correlated genes.

Why bin expression?

Four ordinal bins—zero, low, medium and high in the published workflow—reduce sensitivity to cross-dataset scale differences.

What does XGBoost learn?

It adds shallow decision trees sequentially so later trees correct residual classification errors while regularisation controls complexity.

Can CaSTLe find a new cell type?

It cannot name a type absent from the source; thresholded binary models can flag cells that fit none of the known types for further study.

How should transfer be validated?

Use whole target studies with labels hidden during engineering, per-class metrics, threshold curves, calibration and absent-type challenges.

Primary source and fact-check note

Yuval Lieberman, Lior Rokach and Tal Shay, “CaSTLe—Classification of single cells by transfer learning: Harnessing the power of publicly available single cell RNA sequencing experiments to annotate new experiments,” PLOS ONE 13, e0205499 (2018), DOI 10.1371/journal.pone.0205499. 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

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

Feature-ledger audit

Store shared genes, detection counts, mean ranks, mutual-information estimates, selected lists and correlation removals. Every transferred rule should be reproducible.

Target-label firewall

Document which target quantities were inspected and confirm that no target labels entered feature engineering, bin selection or model tuning. Label leakage invalidates transfer.

Bin-drift audit

For every selected gene, compare source and target occupancy in the four ordinal bins. Large drift identifies fragile transfer features.

Correlation sensitivity

Repeat pruning under biological resampling and keep the more portable feature from each correlated pair. Arbitrary deletion can change boosted splits.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

XGBoost stability

Repeat source splits and seeds. Report tree depth, selected features, source validation and untouched target metrics.

Threshold challenge

For binary classifiers, plot sensitivity, specificity, coverage and conflicting passes across thresholds. Lock tie rules before final testing.

Absent-type test

Remove one known source type and quantify forced labels versus uncertain results. A strong multiclass score can still be wrong out of domain.

Final claim

Name datasets, shared genes, feature rules, bins, model settings and target metrics. Transfer is conditional on source–target comparability.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Why transfer is difficult

Start with the mechanism: Protocols alter depth, dropout and feature distributions even when biology is shared. A reusable classifier needs a common representation. The worked case is concrete: One source has twice the reads per cell. Expose the boundary: Simple scaling does not repair detection differences. Complete this evidence step: Compare source and target distributions before modelling. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 2: Shared-gene intersection

Start with the mechanism: Features present in only one dataset cannot support a transferred rule, so the workflow starts from common genes. The worked case is concrete: Two studies share 12,000 identifiers. Expose the boundary: Identifier mismatches can masquerade as missing biology. Complete this evidence step: Version symbol and accession mapping. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 3: Rare-expression filter

Start with the mechanism: Genes detected in fewer than a declared number of cells are removed because their empirical bins and information estimates are unstable. The worked case is concrete: A gene appears in only nine cells. Expose the boundary: Rare markers may be biologically valuable. Complete this evidence step: Audit lost known markers separately. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 4: High-mean features

Start with the mechanism: Genes with large overall mean across source and target are measurable enough to support robust ordinal comparisons. The worked case is concrete: A housekeeping-like gene passes the mean screen. Expose the boundary: High expression need not be discriminative. Complete this evidence step: Keep mean and label information conceptually separate. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Mutual information

Start with the mechanism: Mutual information measures how much knowing a binned gene reduces uncertainty about source labels. The worked case is concrete: A gene’s bins strongly separate two types. Expose the boundary: Biased estimates arise with sparse counts and small classes. Complete this evidence step: Calculate a two-class contingency table. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 6: Union of feature views

Start with the mechanism: Combining high-mean and high-information genes balances measurability with discrimination. The worked case is concrete: Some genes qualify through both lists. Expose the boundary: Fixed top counts are modelling choices. Complete this evidence step: Plot performance versus feature count. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 7: Correlation pruning

Start with the mechanism: For highly correlated gene pairs, one feature is removed to reduce duplicated evidence and computation. The worked case is concrete: Two ribosomal genes exceed 0.9 Pearson correlation. Expose the boundary: Arbitrary removal can discard a more portable gene. Complete this evidence step: Repeat pruning under resampling. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 8: Four ordinal bins

Start with the mechanism: Continuous expression is mapped to [0], (0,1], (1,6] and (6,infinity) in the published workflow. The worked case is concrete: A value of four enters the third bin. Expose the boundary: Thresholds reflect the original scale and may age poorly. Complete this evidence step: Publish units and bin occupancy. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Why bins can transfer

Start with the mechanism: Broad ranges reduce sensitivity to modest scale shifts and turn each gene into ordered categories. The worked case is concrete: A source value 3 and target value 5 remain in one bin. Expose the boundary: Large shifts still cross boundaries. Complete this evidence step: Measure per-feature bin drift. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 10: Boosted decision trees

Start with the mechanism: XGBoost builds trees sequentially, each responding to gradients from the current classification loss. The worked case is concrete: A later tree repairs errors among related types. Expose the boundary: Boosting can learn source-specific artefacts. Complete this evidence step: Track training and held-out loss. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 11: Regularised objective

Start with the mechanism: Loss is combined with penalties on tree complexity, leaf weights and minimum child evidence. The worked case is concrete: A split is rejected when gain is too small. Expose the boundary: Pre-tuned settings are not universally optimal. Complete this evidence step: Tune only inside source validation. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 12: Multiclass transfer

Start with the mechanism: One model predicts among source types when the target is assumed to contain those same classes. The worked case is concrete: Pancreas source labels a second pancreas study. Expose the boundary: The assumption fails if a target type is absent from source. Complete this evidence step: Audit reference coverage first. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: Binary threshold route

Start with the mechanism: Separate one-versus-rest models can produce scores for each type; a cell is assigned only if the highest score clears a threshold. The worked case is concrete: A low-score cell remains uncertain. Expose the boundary: Several models can pass simultaneously. Complete this evidence step: Define tie and coexpression rules. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 14: Label-blind target access

Start with the mechanism: Feature engineering may inspect target expression but must not inspect target labels. This is transductive unsupervised access, not supervised testing. The worked case is concrete: Target means help define measurable genes. Expose the boundary: Accidental label use invalidates transfer estimates. Complete this evidence step: Log every target-dependent step. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 15: A student boosting project

Start with the mechanism: Bin two datasets, estimate mutual information, prune correlation and fit a small boosted tree ensemble. The worked case is concrete: Known labels reveal scale robustness. Expose the boundary: Real data add thousands of genes and dropout. Complete this evidence step: Withhold one class from the source. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 16: Governance and careers

Start with the mechanism: Version datasets, genes, bins, information scores, correlations, split, XGBoost settings and thresholds. This joins information theory and ensemble learning. The worked case is concrete: Students can begin with contingency tables and trees. Expose the boundary: A transferred label proves no mechanism or diagnosis. Complete this evidence step: Write a transfer model card. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare features, bins, trees, activations, probabilities, hierarchy edits or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.