Level 3 · The framing — two regimes and 'watch, don't run'

§1 — One system, two regimes

Levels 1 and 2 were the spine: a high-risk AI system answering to the EU AI Act and nothing stacked above it. Level 3 is your first overlay — a system that is high-risk under the AI Act and carries a sectoral regime on top.

The system is a diabetic-retinopathy (DR) screener: it reads a fundus photograph and decides refer to an ophthalmologist / do not refer. That decision touches a patient’s sight, so it is high-risk under the EU AI Act. But notice how it becomes high-risk. Student selection was high-risk because of its use (Annex III §3, education). A retina screener is high-risk for a different reason: it is a safety component of a medical device, and the AI Act’s Art. 6(1) + Annex I route says that an AI system which is a safety component of a product already regulated by EU product-safety law (the MDR) is high-risk. So the very same fact — “this is a medical device” — does two things at once: it makes the system AI Act high-risk (via Art. 6(1)), and it makes the system an MDR medical device in its own right (Class IIa). One system, two regimes.

Why two regimes, and what each one is

The EU AI Act governs the system as an AI system: the Articles 9–15 spine (risk management, data governance, documentation, transparency, human oversight, robustness). The MDR (Reg. EU 2017/745) governs the system as a medical device: General Safety and Performance Requirements (GSPR, Annex I), a clinical-evaluation duty (Annex XIV), and a conformity-assessment route (MDR Art. 52) that — for Class IIa — requires a notified body. These are two different legal frameworks, both in force, both applying to the same screener. The overlay is the stack: the AI Act spine, with the MDR layered on top. (If “overlay” is new, the foundations page below walks the picture.)

More than the AI ActOne system, N regimes: the AI Act spine plus a sectoral regime stacked on top. The same signed bundle projects onto multiple catalogs — governed standards get real verdicts, MDR/DORA get a signed crosswalk and an honest GAP.

§2 — Which laws apply (and which don’t)

Be precise about scope — honesty cuts both ways.

• EU AI Act — high-risk, via Art. 6(1) + Annex I. Nerea declares tier: high via the classify-system portal mission, with the basis spelled out in froga.yaml: “Art. 6(1) (safety component of a medical device — SaMD MDR Class IIa)”. That pulls the Articles 9–15 spine into scope.
• MDR (Reg. EU 2017/745) — Class IIa medical device. Declared in applicable_standards as eu/mdr@2017. This is the overlay. But — and this is the crux of the level — the MDR is binding law, not a harmonised AI-risk-management standard cited for Article 9. The engine treats it differently: it emits a signed crosswalk mapping the system’s risks onto the MDR’s GSPR clauses, but it does not print an MDR conformity verdict and the MDR confers no presumption through the engine. Never say “MDR-governed”. (§6 shows exactly what the crosswalk looks like.)
• No DORA. A credit model run by a bank would carry DORA (finance) on top of the AI Act. A retina screener does not: the operator is a clinic, not a financial entity, so DORA simply does not apply. Declaring it would be over-claiming a regime the system isn’t under. You meet DORA in the finance overlay (a later level).

The MDR being law, not a harmonised standard is why a notified body enters the picture at all. Self-declaration is only an option for the lowest device class (MDR Class I). A Class IIa device must go through the MDR conformity-assessment procedure (Art. 52), which requires an accredited, independent notified body to assess the evidence and certify it. The engine is not, and cannot be, that body.

Notified body vs self-declarationWhen you may declare conformity yourself versus when a regime requires an accredited, independent third party — a notified body (MDR Class IIa, MDR Art. 52). The engine signs evidence and assembles documentation, but it is not a notified body and cannot stand in for a clinical evaluation.

---

§3 — Watch, don’t run

This level is 'watch, don't run' — and that's an honest choice, not a shortcut

Levels 1 and 2 ran on a laptop CPU, so you ran them yourself. Retina is different. Re-training the screener extracts ResNet-18 features over the ODIR-5K fundus images — that needs a GPU and the ~2 GB image cache (FROGA_ODIR_CACHE), which is not a laptop job. Rather than fake a “run it yourself” that most readers can’t actually run, this level asks you to read the committed evidence and follow the tag links to the exact repository state at each beat.

What makes this honest rather than hand-wavy: a nightly reproduce gate in CI clones each tag and re-derives the verdict from the committed outputs with the current engine — no GPU, no dataset. The committed bundle.json already carries the measured numbers; the gate confirms the current engine still reads them to the same colour (red / green / green). So the evidence you read is machine-checked to still reproduce — the “Watch the evidence” block at the end shows you how to verify it yourself with only git and froga.