Level 2 · Student selection — the law walks in

🟡 Partial — The risk loop is runnable today on CPU with froga and the public vldemo-student-selection repo (one tag per beat). The portal review runs on platformpre. The four commit-links resolve once vldemo-student-selection is published to the Venturalitica org. The honest close is a clean GREEN on the data-minimization gate, with an acknowledged explainability residual consciously accepted — by design.

TL;DR

An education system decides who gets admitted or placed — it sorts students by a predicted pass/fail. That decision touches a person’s future, so the law walks in: this use is high-risk under the EU AI Act, Annex III §3 (education and vocational training). There is no sector overlay on top — no DORA, no MDR — so this is the clean spine: just the AI Act, plus the two standards the engine can project truthfully, prEN 18228 and ISO 23894. Three roles: Nerea, Product Owner, declares from the portal; James, Quality Manager, reviews and approves from the portal; Martha, engineer, works the froga CLI and git. The subjects may be minors, and the GDPR demands minimizing their data (Art. 5(1)(c)): on real student data, the first model keeps each student’s exact age, and the k-anonymity gate comes out RED — 216 of 1044 students are uniquely re-identifiable by just three attributes. Martha treats it with a code change — generalizing age to a tercile, the classic Sweeney technique — and k-anonymity rises from 1 to 27, with a confidence interval that lands entirely on the safe side: the engine closes clean GREEN, no ambiguity. Demographic parity also improves in parallel, but on the real sample it was already small — it lands as a certified audit control, not the gate. James reviews what does remain a residual (the model’s opacity) and approves with that reason on the record. A clean gate is not “nothing left to watch” — that is the whole lesson.

If you have not met likelihood × impact, appetite, residual and the RED / AMBER / GREEN gate, read the five-minute on-ramp first — and Level 1, the no-law warm-up:

What is a risk?Likelihood × impact, the 5×5 matrix, appetite, inherent vs residual, and the RED/AMBER/GREEN gate — the vocabulary this workshop runs on.

Level 1 · Iris — your first risk loop, no law involvedThe warm-up: a flower classifier no law touches, closing cleanly GREEN. Level 2 keeps the three roles and adds the law, minor subjects, and a data-minimization gate.

This system IS high-risk under the EU AI Act — and that changes everything

In Level 1 we ran the risk loop on flowers because it is good engineering, with no law in sight. Here the system decides about people — and those people may be minors: who is admitted, who is placed. Under the EU AI Act that puts it squarely in Annex III §3 — the list of high-risk uses in education and vocational training. High-risk is not a label you opt into; it is a legal classification that pulls in real obligations: a risk-management system (Art. 9), data governance (Art. 10), technical documentation (Art. 11), transparency to users (Art. 13), human oversight (Art. 14) — and, because minors are involved, the GDPR raises the bar (Art. 5, 8, 22, 35). Level 1’s point was that the method pays off without a law. Level 2’s point is that, when the law does apply, the same method is how you meet it — and how you stay honest about what is measured and what remains an acknowledged residual.

Prerrequisitos

Level 1 (the risk loop) and What is a risk? (the vocabulary), plus Standards & presumption (the standards this level declares). Three material requirements, one per role:

Nerea (portal): an email account to receive the sign-up code (a one-time password, OTP) — she declares and opens the PRs.
James (portal): an email account — he reviews and approves Nerea’s PRs, then approves the risk.
Martha (terminal): git and uv (manages Python ≥ 3.11 and the environment; the workshop uses uv sync and uv run). No GPU — the whole thing runs on a laptop CPU in minutes, on the real UCI Student Performance dataset combining both its cohorts (mathematics + Portuguese, about a thousand rows).

If you do the workshop solo, you play all three parts: keep the browser (for Nerea’s and James’s portal actions) and the terminal (Martha’s CLI work) within reach.

By the end you'll be able to…

Classify a system as high-risk under Annex III §3 and explain why the law enters now (it decides about people, who may be minors) — and why there is no sector overlay here (the clean spine).
Run a complete risk loop on a data-minimization control (GDPR Art. 5(1)(c)) measured as k-anonymity, and read a gate that goes from RED to GREEN with a real generalization treatment.
Tell apart a certified control (demographic parity: real, measured, green, but not the gate) from a blocking control (data minimization: the one that decides the arc’s colour) — and why that distinction is the honest answer, not a narrative trick.
See a real code-change treatment with TWO mitigations in the same diff (generalizing age + a fairness-aware fairlearn learner) and recognise that a clean green gate can coexist with an acknowledged residual (the model’s opacity) consciously accepted at approval.
Read the one-bundle → N-standards projection: the same signed bundle reported against ISO 23894 and prEN 18228 with real verdicts, with presumption = false today — and understand why DORA/MDR are not declared here.

Lessons in this level

The framing — why the law enters and which law appliesWhy a student-selection system is high-risk (EU AI Act Annex III §3) and how ISO 23894 relates to prEN 18228.

The arc, by handoffsThe three-role workshop: the data-minimization gate for a minor's data comes out RED on real data, Martha treats it by generalizing age, and the engine closes clean GREEN — demographic parity lands as a certified control alongside it.

In depth — data minimization, the articles, and the projectionThe k-anonymity gate explained, Articles 5/9/10/13 briefly, the one-bundle-many-standards projection, and why Annex IV is partial by design.

Practice & recapRun the student-selection arc yourself, review the system that closes green with an acknowledged residual, and self-check what you learned.