Part 2 of 12 · 2 min read
Classification sets the tier, and the tier decides every obligation that follows
A use case's domain and practices, not its technology, decide whether it is prohibited, high-risk, limited-risk or minimal, and that single label drives the rest of the module.
Think of building codes. The rules say what you must show an inspector. The engineer's job is to keep the plans, records and logs that prove the building follows them. The building's use (a hospital, a shed) decides which code applies, not which bricks it is made from.
obligations scale with the potential harm of the use case. Its domain and the practices it uses -- not with which model or framework built it, so the first step for any system is a classification into prohibited, high-risk, limited-risk or minimal.
A worked example. Of ten invented systems run through the example's classifier, three land in employment or credit and are automatically high-risk regardless of how well they work; two use a banned practice (non-consensual imagery, social scoring) and are prohibited outright, no matter what evidence they have; three are chatbots or image generators with only transparency duties (limited); and two internal tools with no public-facing or sensitive-domain flag are minimal.
The key lines. From code/mini-compliance/compliance.py:
# from code/mini-compliance/compliance.py
def classify(s: System) -> str:
if s.practices & PROHIBITED_PRACTICES:
return "prohibited"
if s.domain in HIGH_RISK_DOMAINS:
return "high"
if s.flags & TRANSPARENCY_TRIGGERS:
return "limited"
return "minimal"
Read it as: the checks run in a fixed order, and the first match wins. A prohibited practice overrides everything else. A system in a harmless domain that also happens to do social scoring is still prohibited, not minimal.
Risk-based regulation, high-risk, limited-risk, prohibited
- Risk-based regulation: obligations scale with the potential harm of the use case, not with the technology used.
- Prohibited: practices banned outright.
- High-risk: use cases in listed sensitive domains with the heaviest obligations.
- Limited-risk: mainly transparency duties.
Example: loan-triage classifies as high purely from domain="credit", with no need to look at how the triage model was built.