Machine Learning, from the Ground Up

0. How to read this chapter

You are not being hired to train models. You are being hired to decide whether somebody else's model is sound enough to fine a firm ten percent of its worldwide turnover. So every idea below comes with two things: what it is, and what goes wrong with it in a way a lawyer needs to know about.

One running example throughout: 40,000 past public tenders, each with a handful of recorded facts, and the question of which ones involved collusion between bidders.

1. The objects, then the paradigms

Everything later is built from eight things. They are worth naming precisely, because most confusion about machine learning is confusion about which of these somebody means.

Term	Definition	In the tender dataset
Dataset	A collection of examples you learn from	40,000 past tenders
Instance, example, row	One single thing you want to say something about	One tender: the Valencia road contract of March 2019
Feature, predictor, variable	One measured property of an instance. The input	Number of bidders — 3. Winning margin below estimate — 2 percent. Sector — construction. Country — Spain
Label, target, outcome	The thing you want to predict. The output. Only present in supervised learning	Was this tender rigged — yes or no
Model	A function that takes features and returns a prediction	A rule that takes those four features and returns a risk score between 0 and 1
Parameters, weights	The numbers inside the model that get adjusted during learning. You do not choose these; the algorithm finds them	The weight on number of bidders comes out at minus 0.4, meaning more bidders lowers the score
Hyperparameters	Settings you choose before learning, which control how learning happens	How deep a decision tree may grow. How much regularisation to apply. Where the decision threshold sits
Training and inference	Training is finding the parameters from data. Inference, or prediction, is applying the finished model to something new	Train on 2010 to 2020 tenders; infer on a tender published this morning

Two distinctions you will need constantly:

Pair	Difference	Example
Classification versus regression	Both are supervised. Classification predicts which category something falls into; regression predicts a quantity	Classification: is this tender rigged, yes or no. Regression: what will the winning bid be, in euro
Binary, multiclass, multilabel	Two categories; more than two but only one can apply; or several that can apply at once	Binary: rigged or not. Multiclass: which of cover bidding, rotation or suppression. Multilabel: a tender showing rotation and suppression together

Now the five paradigms. What separates them is entirely what you are given to start with.

And this is not one agency experimenting. On Schrepel's own account, all 27 EU national competition agencies, DG Competition, and most agencies worldwide now rely on computational tools. The field moved from experiment to routine in under a decade. Six deployments are publicly described, and they are doing four quite different jobs.

Agency and tool	What it does	Which technique from this chapter
Spain — BRAVA	Screens public procurement for bid rigging, mapping relationships between firms, bids and individuals	Supervised classification, with network features built from the relationship graph. Labels come from past enforcement
Brazil — Cerebro	Runs procurement documents through analysis to surface signs of collusion	Document processing and classification — closer to the NLP chapter than to tabular modelling
France	Queries its own case database in natural language, through a retrieval-augmented generation system built on large language models	Retrieval plus generation. Not a screen at all — it finds nothing new, it helps case handlers find what the agency already decided
Greece	Analyses email metadata seized in dawn raids to establish which companies communicated with each other	Network analysis over communications metadata. Who talked to whom, how often, and when
Chile	Monitors more than 80,000 products for price anomalies	Unsupervised anomaly detection over time series. No labels, so it reports unusual
Poland	Investigates dark patterns using eye tracking and other neuromarketing methods	Human-subject experimental research. Barely machine learning at all, and a different legal animal entirely

The four jobs, and why collapsing them is the mistake

Deciding who to investigate — Spain, Brazil, Chile. These select targets, so the legal questions are selection bias, false positives, and whether the firm ever learns a model chose it. Analysing evidence already lawfully held — Greece. Here the questions are rights of defence, access to the file, and whether the firm can replicate the analysis. Internal knowledge retrieval — France. The question is corpus integrity, and whether a wrong retrieval silently contaminates a decision. Researching third-party conduct — Poland. Closer to consumer protection than enforcement. The literature treats all of this as *AI in competition enforcement*, as though it were one object. It is four objects with four distinct legal problem sets, and separating them is a real contribution.

What the public record does not tell you, and why that is the opening

The Stanford cross-agency reports are the only systematic account of any of this, now in a fifth edition covering thirty agencies — and they are agencies describing their own tools. You rarely learn the model class, the training data, the validation design, the decision threshold, or the measured error rates. For a line of work that selects firms for investigation, almost none of what this chapter says you should check is disclosed. That absence is itself the finding, and it is what the audit protocol project exists to fix. Agencies also name their own three biggest difficulties in those reports: secure cloud computing, the explainability of deployed tools, and improving human–machine interaction. The second and third are this chapter and chapter 9. They are not hostile to the question — they have already said it is their problem.

2. What learning actually is

Stripped of vocabulary, supervised learning is four decisions.

Decision	Plain meaning	In the example
Hypothesis class	Which shapes of rule you are willing to consider. You are choosing a family of candidate models before seeing which one wins	Only straight-line relationships. Or: any decision tree up to depth six. Or: any combination of 500 trees
Loss function	A number saying how bad a single prediction was. Smaller is better	The model said 0.9 and the tender was clean. That costs a lot. It said 0.52 and the tender was rigged. That costs a little
Empirical risk minimisation	Search the hypothesis class for the model with the lowest average loss on the data you have. This is what training is	Try parameter settings until the average cost across 40,000 tenders stops falling
Generalisation	Whether that model still works on tenders it has never seen. The only thing that actually matters	Trained on 2010 to 2020, and the question is whether it works in 2026

A loss function is just a formula for the cost of being wrong. The common ones:

Loss	What it penalises	Used for
Squared error	The squared distance between prediction and truth. Being wrong by 10 costs a hundred times being wrong by 1	Regression. Badly affected by outliers, which heavy-tailed contract values are full of
Absolute error	The plain distance. Being wrong by 10 costs ten times being wrong by 1	Regression where outliers should not dominate. Targets the median rather than the mean
Cross-entropy, log loss	How surprised the model should be by the truth it was given. Confident and wrong is punished very heavily	Classification. This is the standard one
Hinge	Only whether the prediction was on the right side of the boundary, by a margin	Support vector machines

\text{MSE} = \frac{1}{n}\sum_{i=1}^{n}\left(y_i - \hat{y}_i\right)^2

Mean squared error equals the average, over all cases, of the squared difference between the true value and the prediction.The hat on y means *predicted*. Squaring means an error of 10 costs a hundred times an error of 1, which is why outliers dominate it.

\text{Log loss} = -\frac{1}{n}\sum_{i=1}^{n}\Big[ y_i \log \hat{p}_i + (1 - y_i)\log(1 - \hat{p}_i) \Big]

Log loss equals minus the average of: the true label times the log of the predicted probability, plus one minus the label times the log of one minus the predicted probability.Also called binary cross-entropy. Only one of the two terms survives per case, because the label is either 1 or 0 — so it reduces to minus the log of the probability you assigned to the right answer.

Pair	Definition	Example
Parametric	The model has a fixed number of parameters whatever the data size. Learning means setting those numbers	Logistic regression with four features has about five numbers, whether you train on 400 tenders or 4 million
Non-parametric	The model's size grows with the data. It can keep adding structure	A decision tree adds branches as needed. Nearest-neighbour methods effectively store the whole dataset

3. Overfitting, underfitting, and the tradeoff

Two different ways a model can be wrong, and they need opposite fixes — which is why diagnosing which one you have is the first thing you do.

Term	Plain meaning	Cause
Bias	The model is too simple to capture the real pattern, so it is wrong in the same direction for everyone. Called underfitting	Hypothesis class too restrictive, too few features, too much regularisation
Variance	The model is so flexible it has fitted the accidents of this particular sample. Called overfitting	Model too flexible for the amount of data, too many features, too little regularisation
Irreducible noise	Genuine randomness no model can remove. Two identical tenders can have different outcomes	The world

\mathbb{E}\big[(y - \hat{f}(x))^2\big] = \underbrace{\big(\text{Bias}[\hat{f}(x)]\big)^2}_{\text{too simple}} + \underbrace{\text{Var}[\hat{f}(x)]}_{\text{too sample-specific}} + \underbrace{\sigma^2}_{\text{irreducible}}

Expected squared error decomposes exactly into bias squared, plus variance, plus irreducible noise.An identity, not an approximation. It is why you cannot reduce both terms at once by tuning alone — and why the only way to improve both is more data or better features.

Reducing one term often raises the other, which is why it is a tradeoff rather than a problem to be solved.

What you see	What it means	What to do
Both numbers poor and close together	Underfitting	More capacity, better features, less regularisation
Training number excellent, test number poor	Overfitting	More data, more regularisation, fewer features, a simpler model
Both reasonable and close	A defensible fit	Now check calibration and performance per subgroup before believing it

4. Regularisation, or deliberately handicapping the model

Regularisation means adding a penalty for complexity, so that training has to trade off fitting the data against staying simple. You are accepting slightly worse performance on the data you have, in exchange for better performance on data you do not.

To follow this you need one idea: a weight is the number a model multiplies a feature by. A large weight means that feature strongly moves the output; a weight of zero means the feature is ignored entirely.

Method	What it does	Effect
L2, ridge	Adds a penalty equal to the sum of the squared weights	Shrinks every weight smoothly toward zero without reaching it. All features stay in, none dominates
L1, lasso	Adds a penalty equal to the sum of the absolute weights	Pushes some weights to exactly zero, which removes those features entirely. It performs feature selection for you
Elastic net	Both penalties together	Sparsity, with correlated features handled more gracefully
Early stopping	Stop training when performance on held-out data starts getting worse	Caps complexity without changing the model
Dropout	Randomly switch off part of a neural network during each training step	Stops the model relying on any single path
Data augmentation	Train on altered copies of the input that should not change the answer	Teaches an invariance you believe in

\hat{w} = \arg\min_{w} \; \underbrace{L(w)}_{\text{fit the data}} \; + \; \underbrace{\lambda \, R(w)}_{\text{stay simple}}

The chosen weights are those minimising the loss plus lambda times a complexity penalty.Lambda is the dial. At zero there is no regularisation; large lambda forces every weight toward zero and the model underfits. Lambda is a hyperparameter, so a human chooses it.

R_{L2}(w) = \sum_{j=1}^{p} w_j^2 \qquad\qquad R_{L1}(w) = \sum_{j=1}^{p} \lvert w_j \rvert

The L2 penalty is the sum of squared weights. The L1 penalty is the sum of absolute weights.The whole difference is squared against absolute — and that is why L1 drives weights to exactly zero while L2 only shrinks them. Squaring makes the penalty on a tiny weight negligible, so L2 never bothers finishing the job.

5. The algorithms, explained rather than listed

6. Unsupervised methods

No labels, so the goal is structure. Two families: grouping similar things together, and compressing many features into few.

Method	What it does	Watch out for
k-means	Pick k centre points, assign every instance to the nearest, move each centre to the average of what it caught, repeat until nothing moves	You must choose k yourself. Assumes roughly round, similar-sized groups. Sensitive to feature scaling and to where the centres start
Hierarchical clustering	Repeatedly merge the two closest groups, producing a tree of nested groupings	No k needed in advance, and the tree is readable. Expensive on large datasets
DBSCAN	Groups points that sit in dense regions, and labels points in sparse regions as noise rather than forcing them into a group	Finds irregular shapes and identifies outliers directly. Struggles when some genuine groups are much sparser than others
Gaussian mixture model	Assumes the data came from several overlapping bell curves and works out which, giving each instance a probability of belonging to each group	Soft membership rather than hard assignment, which is often more honest
Principal component analysis	Finds the directions along which the data varies most and describes each instance by its position along those directions	Reduces 200 features to 10 while keeping most of the variation. The 10 are mixtures of the originals, so interpretability drops sharply
t-SNE and UMAP	Squash high-dimensional data to two dimensions so you can look at it	For looking only. Distances between clusters in the picture are not meaningful, and people over-read these plots constantly
Isolation forest, one-class methods, autoencoders	Anomaly detection: learn what normal looks like and score how far each instance sits from it	This is what an unlabelled screen actually is. It outputs unusual, which is not unlawful

7. Preparing the data

Most of the real work, and most of the real errors.

Step	What it means	Example
Scaling	Putting features on comparable numeric ranges. Standardising subtracts the average and divides by the spread	Contract value runs to millions, bidder count to single digits. Without scaling, distance-based methods see only the value. Trees do not care
Categorical encoding	Turning categories into numbers. One-hot creates one yes/no column per category	Country becomes 27 columns. For something with thousands of categories, such as buyer identity, one-hot is impractical and you encode by frequency or by average outcome instead
Missing data	Deciding what to do about blanks: drop the row, fill with the average, predict the value, or add a column recording that it was missing	In enforcement data, missingness is almost never random — the fact that a field is blank is itself informative, so an explicit missing flag usually beats filling it in
Transforming skew	Taking logarithms of quantities with a long right tail, so the few huge values stop dominating	Contract values. Model the logarithm, not the euro
Feature construction	Building the informative features by hand from the raw ones	Margin as a fraction of estimate. Bidder count relative to the sector norm. Gap between lowest and second-lowest bid. These are where the signal lives, and none of them is in the raw data
Curse of dimensionality	With many features, all points become roughly equally far apart and distance stops being informative	Why nearest-neighbour methods degrade as features are added

1. Scaling — watch a feature disappear

Two tenders. A: value 2,000,000 euro, 3 bidders. B: value 2,000,100 euro, 12 bidders. A distance-based method measures how far apart they are. Unscaled: the value differs by 100 and the bidder count by 9, so the squared distance is 100 squared plus 9 squared — that is 10,000 plus 81. The value contributes 99.2 percent of the distance and the bidder count 0.8 percent. Yet these two tenders are nearly identical in price and completely different in competitiveness. Standardised: divide each difference by that feature's spread across the dataset — say 5,000,000 euro for value and 2 for bidders. Now the value difference is 100 over 5,000,000, which is 0.00002, and the bidder difference is 9 over 2, which is 4.5. The bidder count now contributes essentially all of the distance. Same two tenders, opposite conclusion, and nothing changed but the units.

2. Categorical encoding — three ways, and the trap in the third

One-hot: country becomes 27 yes/no columns, so Spain is 1,0,0,… and France is 0,1,0,…. Fine for 27. Now try buyer identity with 8,000 buyers: 8,000 columns, almost all zero, and a buyer appearing once gets its own column the model can simply memorise. Frequency encoding: replace Madrid City Council with how often it appears — 1,240 of 40,000 tenders, so 0.031. One column instead of 8,000. Target encoding: replace it with the rigged rate among its tenders — 37 of 1,240, so 0.0298. Powerful, and here is the trap: compute that on the whole dataset and a buyer with exactly one tender, which was rigged, gets the value 1.0. That number *is* the label. The model will score perfectly in validation and fail completely in deployment. Target encoding must be computed inside each cross-validation fold, never before splitting.

3. Missing data — and why filling the blank is usually wrong

The estimated value field is blank in 6,000 of 40,000 tenders. Mean imputation: fill all 6,000 with the average, 1.8 million. Now 6,000 rows share one identical value, and the model can learn *estimate exactly 1,800,000* as a perfect proxy for *this field was missing* — you have not removed the information, you have disguised it. Then ask why it is missing. Small municipalities often do not publish estimates, so missingness tracks buyer size, which tracks everything else. Missing indicator: add a column that is 1 when blank. Check what it is worth: if the rigged rate among blank-estimate tenders is 1.2 percent against 0.4 percent overall, then the blank is three times more informative than most of your real features. In enforcement data the absence of a record is evidence, and imputation throws it away.

5. Feature construction — where the signal actually lives

The raw record: winning bid 1,960,000; published estimate 2,000,000; three bidders; second-lowest bid 2,310,000; sector construction. Each of those alone says very little. Margin against estimate: 2,000,000 minus 1,960,000, over 2,000,000, equals 2.0 percent — the winner came in only just under, which is odd when competing hard. Bidder count against sector norm: the construction average is 7.4, so 3 over 7.4 is 0.41, meaning well under half the usual competition. Gap to the second bid: 2,310,000 minus 1,960,000 is 350,000, which is 17.9 percent of the winning bid — a suspiciously wide margin of victory. All three numbers are informative. Not one of them exists in the raw data. This is why feature construction is most of the work.

Three leaks, with what each one costs

One, preprocessing before splitting. You standardise using the average across all 40,000 tenders, then split. The training step has now absorbed information from the test rows. Usually worth a few points of inflated accuracy — small, and enough to make a worse model look better than a good one. Two, a feature recorded after the outcome. *Days until case closed* exists only for tenders that became cases, so it is blank for every clean tender. The model learns *field is populated, therefore rigged* and reports 99 percent accuracy. Three, and the one specific to enforcement: *number of documents requested from this firm*. That number is above zero only because somebody already opened an investigation. The model appears to predict collusion with uncanny accuracy; it is detecting that the agency already looked. This is the first thing to check when auditing a screen, because a leaked model and a brilliant model look identical from the outside and only one survives deployment.

8. Evaluation, which matters most to you

Everything starts from a table of four counts. These four words appear constantly, so they are worth fixing properly.

Term	Definition	In enforcement terms
True positive	Flagged, and genuinely rigged	A cartel caught
False positive	Flagged, but actually clean	An innocent firm raided
True negative	Not flagged, and clean	Correctly left alone
False negative	Not flagged, but rigged	A cartel missed

Those four counts arranged in a two-by-two grid are called the confusion matrix. Every metric below is a ratio of them.

The confusion matrix

                        ACTUALLY RIGGED     ACTUALLY CLEAN
                      +-------------------+------------------+
      FLAGGED         |   TP              |   FP             |
                      |   cartel caught   |   innocent firm  |
                      |                   |   raided         |
                      +-------------------+------------------+
      NOT FLAGGED     |   FN              |   TN             |
                      |   cartel missed   |   correctly      |
                      |                   |   left alone     |
                      +-------------------+------------------+

      Precision  reads ACROSS the top row     ->  TP / (TP + FP)
      Recall     reads DOWN the left column   ->  TP / (TP + FN)

Precision reads across the flagged row; recall reads down the actually-rigged column. That is the whole distinction, and it is the one people reverse under pressure. Precision asks: of those I accused, how many were guilty. Recall asks: of those who were guilty, how many did I catch.

\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}

Accuracy equals true positives plus true negatives, divided by the total number of cases.Everything you got right, over everything. Useless under imbalance — see the base-rate box below.

\text{Precision} = \frac{TP}{TP + FP}

Precision equals true positives divided by true positives plus false positives.Of everything you flagged, the share that was genuinely rigged. This is the number the firm cares about.

\text{Recall} = \frac{TP}{TP + FN}

Recall equals true positives divided by true positives plus false negatives.Of everything genuinely rigged, the share you caught. This is the number the agency cares about. Also called sensitivity, or the true positive rate.

\text{Specificity} = \frac{TN}{TN + FP}

Specificity equals true negatives divided by true negatives plus false positives.Of everything genuinely clean, the share you correctly cleared. Its complement, FP over FP plus TN, is the false positive rate.

F_1 = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}

F one equals two times precision times recall, divided by precision plus recall.The harmonic mean, which punishes imbalance between the two — 0.9 and 0.1 gives 0.18, not 0.5. It also weights them equally, which is wrong whenever their real costs differ.

\text{Precision} = \frac{\text{Recall} \cdot \pi}{\text{Recall} \cdot \pi + (1 - \text{Specificity})(1 - \pi)}

Precision equals recall times the base rate, divided by recall times the base rate plus the false positive rate times one minus the base rate.Bayes' theorem written in confusion-matrix terms, where pi is the base rate. This is the formula behind the base-rate argument: hold recall and specificity fixed, vary pi, and precision moves anyway.

Metric	Definition in words	When it misleads
Accuracy	Of everything, the fraction got right	Useless when one class is rare. If 1 tender in 1,000 is rigged, predicting never rigged scores 99.9 percent
Precision	Of those you flagged, the fraction that really were rigged	This is the number firms care about. Low precision means innocent firms raided
Recall, sensitivity	Of those really rigged, the fraction you flagged	The number the agency cares about. Trades directly against precision
Specificity	Of those really clean, the fraction correctly cleared	Looks reassuringly high under imbalance even when precision is dreadful
F1	A single score blending precision and recall	Hides which of the two is bad, and weights them equally when their real costs are nothing like equal
ROC AUC	The chance that a randomly chosen rigged tender scores above a randomly chosen clean one	Over-optimistic when positives are rare, because the false positive rate has an enormous denominator
Precision–recall AUC	The same idea, but focused on the flagged set	The right summary for rare positives. Use this for screening, not ROC
Calibration	Whether a stated 0.7 really corresponds to 70 percent	Independent of everything above. A model can top every leaderboard and still be badly calibrated

9. Splitting the data honestly

You cannot judge a model on data it learned from, any more than you can mark your own exam. So you hold data back. How you hold it back is where the errors are.

Method	What it is	Why
Train, validation, test	Three portions. Fit on the first, tune choices on the second, and touch the third exactly once at the very end	If you tune against the test set you have learned from it, and the reported number stops being an estimate of anything
k-fold cross-validation	Split into k parts. Train on k minus one, measure on the one left out, rotate so each part is left out once, average the results	Uses all the data for both purposes and gives a spread rather than a single number. Usually k of 5 or 10
Stratified k-fold	The same, but each fold keeps the same proportion of rigged tenders	Essential with rare positives. Otherwise a fold may contain none at all
Grouped split	Keep all rows belonging to the same entity on the same side of the split	If a firm appears in 400 tenders, splitting by tender puts the same firm in both train and test and inflates the result
Time-based split	Train on earlier periods, test on later ones	The real task is predicting the future. A random split lets the model see tomorrow, which is leakage
Nested cross-validation	An inner loop for tuning, an outer loop for measuring	The correct procedure when you must both tune and report from the same data

10. When the interesting class is rare

Class imbalance means one outcome is far more common than the other. Cartels are rare, so a model can score brilliantly by never flagging anything. Five standard responses:

Approach	What it does	Cost
Class weights	Tell the loss function that getting a rigged tender wrong costs more than getting a clean one wrong	Simplest, and usually first choice. It distorts the output probabilities, so recalibrate afterwards
Undersampling	Throw away most of the clean tenders so the classes balance	Discards real information
Oversampling	Duplicate the rigged tenders until the classes balance	Encourages memorising the few positives you have
SMOTE	Invent new synthetic rigged tenders by interpolating between real ones	Can produce implausible cases. Averaging two real cartels does not give you a third
Threshold moving	Leave the model alone and choose explicitly how high the score must be before you act	Often the cleanest, because it makes the tradeoff a visible decision rather than a hidden default

11. Explaining a model

Two different ideas, routinely confused. Interpretability is a property of the model: you can read it directly. Explanation is something generated afterwards about a model you cannot read.

Method	What it gives you	Limits
Interpretable models	Read the thing itself: a linear equation, a printable tree, a rule list	May cost accuracy — though on tabular data far less often than people assume
Permutation importance	Shuffle one feature and see how much performance drops. Big drop means it mattered	Misleading with correlated features, which share the credit arbitrarily
Partial dependence	Vary one feature across its range, holding the rest, and plot the average effect	Assumes you can vary one feature independently of the others, which is often false
LIME	Fit a simple readable model in the small neighbourhood of one prediction	Unstable. Re-running with different random samples can change the explanation
SHAP	Attribute the prediction across features, with a principled grounding in cooperative game theory	Expensive to compute exactly. And faithful to the model, not the world — it says what the model used, not what caused anything
Counterfactual explanation	The smallest change that would have flipped the decision	Often the most legally useful form, because it tells the firm what would have had to be different

12. Bias and fairness, properly

*Bias* means two completely different things in this field, and conflating them wastes most discussions. There is statistical bias from section 3, meaning a model too simple to fit the pattern. And there is social bias, meaning systematically worse treatment of some group. This section is the second. It enters at five distinct points.

Source	Definition	In a screen
Historical	The world the data records was already unequal	Past enforcement concentrated on construction, because that is where the agency built expertise
Representation	Some groups are under-sampled	Small firms rarely bid for above-threshold contracts, so they are thin in the data
Measurement	The thing you recorded differs from the thing you care about	The label is was prosecuted, not did collude
Aggregation	One model for groups that behave differently	Construction and IT services have unlike normal bid spreads, so IT gets scored against construction's baseline
Deployment	Used differently from how it was validated	Validated on national tenders, run on municipal ones

Fairness criterion	What it requires
Demographic parity	Equal flag rates across groups
Equal opportunity	Equal recall across groups — you catch the same proportion of real cartels in each
Equalised odds	Equal recall and equal false positive rates across groups
Calibration within groups	A score of 0.7 means 70 percent in every group
Individual fairness	Similar cases treated similarly

13. Four questions you asked

14. How this connects to ATLANTIS

This chapter is the technical substrate of the fairness strand, and a good part of the accuracy strand too.

Schrepel's framing of the fairness problem	The machine learning content of it
AI systems inherit the biases of their training data	The five sources in section 12, and measurement bias above all: the label is was prosecuted, not did collude
When a model flags one market rather than another, the selection reflects choices about data and design that nobody outside the agency can examine	Feature construction, class weighting, the decision threshold, and variance under resampling. Every one is a design choice with a distributional consequence, and none appears in the decision
The duty to give reasons meets an explainability problem the case law never anticipated	Section 11: interpretable models against post-hoc explanation, and the fact that faithful, stable and causal are three different properties
The AI Act adds obligations but its application raises questions the text does not settle	Article 10 says examine for bias without saying how, and the impossibility result means no choice of metric discharges it neutrally

15. What the panel brings to this chapter

Panel member	Where machine learning meets their work
Thibault Schrepel	Built a knowledge graph of every Commission competition decision from 1977 to 2025. Constructed an openness metric with Pott and applied it across GPT-4, Llama 3, Gemini, Mistral and MidJourney, then argued policy from the scores. Co-authored with Alex Pentland at MIT on foundation model competition. The pattern throughout: build an instrument, measure something with it, then argue law from the measurement. That is the house style and it is what your position is for
Catalina Goanta	Founded the Maastricht Law and Tech Lab with computer scientists in residence at a law school, and runs classification and measurement studies at scale. She has lived the practical problem your appointment creates — how a lawyer and a modeller divide work without one becoming the other's service desk. Expect a question about it, and have a view
Georgiana Mirza	Digital ecosystems and market regulation, where the measurement questions are about structure and concentration
Tijmen Wisman	SyRI, where a state detection model was struck down partly because its risk indicators and its operation were not sufficiently knowable from outside. That is an interpretability finding handed down by a court

16. What is unexplored, and five projects

Project 1. An audit protocol for enforcement screens

The gap: there is no agreed procedure for auditing a screening tool, so each agency's self-assessment is unreviewable and no two are comparable. The method: build the protocol and validate it against a real or reconstructed screen. Leakage checks first, then split design grouped by firm and ordered by time, then precision–recall rather than ROC, the base-rate translation into expected false flags, calibration by reliability diagram, flag-rate disparity across firm size and sector and country, and stability under bootstrap resampling. Why it works: every component is standard practice somewhere and none has been assembled for this setting. Produces: an instrument agencies could adopt, with a real route to use through the Stanford network of 80-plus agencies.

17. Your CV, mapped onto this chapter

What you have	Where it lands	Why it fits
Supervised learning across the standard algorithms — logistic regression, trees, forests, SVM, KNN, boosting	Projects 1 and 3	Auditing a screen means being able to rebuild it, and tabular procurement data is where these methods live
Unsupervised fraud detection at Infosys — insurance fraud, identity theft, phantom billing	Projects 1 and 3	Rare positives, no reliable labels, and a flag meaning anomalous rather than unlawful. You have met the failure mode already
Model evaluation discipline — cross-validation, grid search, calibration, interval estimation at small n, agreement reporting	Project 1, which is essentially this skill written down as a protocol	A protocol is only credible from someone who has been burned by a bad evaluation. You have, and you published the post-mortem
The v1 defect analysis — a grader fed reasoning traces instead of output, empty responses scored as results, and an omitted control arm that invalidated the headline	Projects 1 and 3	A leakage and measurement-validity case study with your name on it, and the most persuasive evidence you own that you audit your own work
Feature engineering and preprocessing	Projects 1 and 3	The informative features in bid screens are constructed, not given
Production deployment at 28,000 monthly users with guardrails and edge cases found in live logs	Project 5	Documentation from someone who has operated a system differs from documentation by someone who has only specified one
Legal training in evidence and procedure	Projects 2 and 4	Both are normative arguments needing the doctrinal half done properly rather than gestured at

18. Where this shows up in the interview

If they ask	Reach for
How would you audit a screening tool?	Leakage first, then the split design — grouped and time-based — then precision–recall rather than ROC, then the base-rate calculation, then calibration, then subgroup flag rates, then stability under resampling
What model would you use?	Logistic regression as the baseline you can explain, boosting as the comparison, and the gap between them is the price of interpretability stated as a number
Why is this harder than ordinary classification?	Rare positives, no clean labels, labels that come from past enforcement, and a false positive that is a dawn raid on an innocent firm
Can you make a model fair?	Name the five sources, then the impossibility result. Fairness is not one property, and choosing between criteria is normative rather than technical
Explain overfitting to a lawyer	A high-variance model trained on slightly different history would have flagged different firms, so the choice of who gets investigated is partly arbitrary — and arbitrariness is a legal defect
Explain any of this to a non-technical panel	Use the threshold. One number decides how many innocent firms are disturbed per cartel caught, and nobody chose it deliberately

19. If you remember twelve things

A feature is an input, a label is the answer, parameters are found by training, hyperparameters are chosen by you.
Supervised has labels; unsupervised does not; self-supervised invents them; reinforcement has only a reward.
Error equals bias squared plus variance plus noise. Variance is the one with legal consequences.
The gap between training and test performance is the diagnosis of over- or underfitting, every time.
L1 produces sparsity, and sparsity is what makes a model explainable in a decision.
Logistic regression is the baseline worth defending in a regulated setting; boosting usually wins on tabular data.
Anomaly detection flags *unusual*, not *unlawful*.
Leakage is the first thing to check in an audit, and the classic case is a feature that exists only because an investigation happened.
Accuracy is useless under imbalance. Use precision–recall, not ROC, for rare positives.
The base rate destroys precision: 95 percent recall and 95 percent specificity at one in a thousand gives under two percent precision.
Split by group and by time, never randomly, on enforcement data.
Calibration within groups and equalised odds are provably incompatible when base rates differ.