Level 1 · Iris — your first risk loop, no law involved

🟡 Partial — The risk loop is runnable today on CPU with froga and the public vldemo-iris repo (one tag per beat). The portal review runs on platformpre. The four commit-links resolve once vldemo-iris is published to the Venturalitica org.

⚡ TL;DR

A botanical garden classifies iris flowers by species — a system that harms no one and is subject to no law. We apply ISO 23894 anyway, because it is good engineering. Three people, one story: 🧭 Nerea, Product Owner, works from the portal (declares the system and identifies the risk via authoring missions — the mission edits froga.yaml and opens a PR); 🛡️ James, Quality Manager, works from the portal (reviews and approves the declaration PR, reviews the evidence and approves the release); 🧑‍💻 Martha, engineer, works from the froga CLI and git (treats, signs, pushes — she cannot approve). The first model sees only the sepal measurements and confuses versicolor with virginica: the macro-recall gate comes out RED (0.81, below the > 0.90 line). Martha treats it by adding the petal features — the ones that actually tell the two species apart — and the gate turns GREEN (0.95). Git closes the loop: the signed evidence travels to the portal, James reviews it and approves. Clean RED → GREEN, no AMBER, no law.

Why does a botanical garden care? A living collection is only as good as its labels. Each iris in the beds is tagged with its species so that a visiting botanist studies the right plant, a seed-exchange request ships the right specimen, and the catalogue stays trustworthy. Misread one versicolor as a virginica and you have quietly corrupted the record — the researcher draws conclusions about the wrong plant, the exchange sends the wrong seeds, and every later user inherits the mistake. No one is harmed and no regulator is watching, yet the work still has to be right. That is exactly where risk management earns its keep on merit alone — which is why the course starts here.

If you have not yet met the words likelihood × impact, appetite, residual and gate colours, read the five-minute on-ramp first — this workshop assumes them:

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.

This system is deliberately minimal-risk under the EU AI Act

Classifying flowers decides nothing about any person. It is not in Annex III (the high-risk uses), it is not a safety component, and it triggers no transparency duty — so under the EU AI Act it falls in the minimal-risk tier: the residual category that carries no legal obligations at all. (Minimal risk is a real tier of the Act, alongside prohibited, high-risk and limited / transparency risk — it is not a vague “not high-risk”.) We run a full ISO 23894 risk loop on it anyway — because it is good engineering, not because a law forces us. That is the entire point of Level 1: you learn the value of managing risk before any legal obligation enters the picture. The law arrives in Level 2; here it is just you, a model, and an honest measurement. Think of Iris as the simulator before the real track: you fly the loop here with flowers because no one is harmed if you get it wrong. The exact same loop later decides who gets into university (Level 2), whether a patient is screened in time (Level 3), and who gets the job (Level 5). The method is identical at every level; only the stakes change.

The modality map — where this level sits

There is more than one way to treat a risk, and more than one kind of control that gates it. This course makes both explicit at every level. Here:

• Treatment modality: a code change V1→V2. Martha edits the FEATURES list in train.py (from sepal-only to sepal + petal) to add the petal features. (Other levels treat with organizational measures or with a conscious-acceptance override.)
• Gate control: a performance control — macro recall, a quality metric. It is not a fairness control. Fairness arrives in Level 2 (student selection), where the data has a protected attribute and the honest result is AMBER, not green. Iris has no protected attribute, so there is nothing to be fair about; the risk here is simply accuracy.

📋 Prerrequisitos

None. No prior knowledge of AI or compliance is required. Three material requirements, split by role:

• 🧭 Nerea (portal): an email account to receive the sign-up code (a one-time password, OTP). Runs the authoring missions from the portal.
• 🛡️ James (portal): an email account to receive the sign-up code (OTP). Reviews and merges the declaration PR; later reviews the evidence and approves the release.
• 🧑‍💻 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 a ~5 KB dataset.

If you do the workshop solo, you play all three parts: keep the browser and the terminal within reach.

🎯 By the end you'll be able to…

• See how responsibility is split: 🧭 Nerea declares the system and identifies the risk from the portal via authoring missions (the mission edits froga.yaml and opens a PR); 🛡️ James merges the declaration PR, reviews the evidence and approves the release (portal); 🧑‍💻 Martha treats, measures and signs (CLI + git). The approval is James’s; the evidence is Martha’s. Martha literally cannot approve — the cryptographic four-eyes by key and the signed-act ledger (.froga/acts/) enforce it.
• Walk one complete ISO 23894 risk loop on a system with no legal obligation at all, and explain why you’d still bother.
• Watch a gate go from RED to GREEN because Martha treated a control with a commit in git (git closes the loop, ISO 23894 §6.5) — the treatment is a versioned parameter change anyone can reproduce.
• Read the evidence trail: the signed act .froga/acts/NNNN-approve.json is the approval, recorded as a DSSE-signed file in the ledger, not a click in a database.

Lessons in this level

The arc, by handoffsThe four handoffs of the workshop: Nerea declares (mission), James approves the PR, Martha measures (RED), treats (GREEN) and signs, James approves the release.

Practice and recapRun the full arc on your laptop, review what you built, and self-check what you learned.