Part 3 of 4 · 2 min read
Exercises
Easy
Easy 1: Run it (30 minutes)
Run python -m pytest -q and python run_all.py in code/mini-compliance.
Done when: you list which sample systems fail and why.
Easy 2: Classify your own (30 minutes)
Describe a system you know as a System and run classify.
Done when: you note one place where the simplified rules and your judgement differ.
Easy 3: Date check (30 minutes)
Find the Official Journal text of the AI Omnibus and compare it with the date table.
Done when: you correct any date and note the source.
Hint
Official Journal and the Commission or AI Office pages are primary sources.
Medium
Medium 1: Evidence quality (3 hours)
Replace boolean evidence with a file path and a check that the file is non-empty and recent.
Done when: a stale model card fails the gate.
Hint
compare file modification time with the last model or prompt change.
Medium 2: CI integration (3 hours)
Run the gate in CI from a manifest in the repo.
Done when: a pull request that removes evidence is blocked with named reasons.
Hint
print one line per missing item so reviewers see what to supply.
Medium 3: GDPR rows (3 hours)
Add data-subject-request checks (erasure reaches derived copies) using module 6's lineage.
Done when: a system with untracked copies fails.
Hard
Hard 1: Counsel comparison (6 hours)
Have counsel classify a real system and list obligations. (I did not run this, so your numbers are the ground truth.)
Done when: you report the differences and update the rules.
Hard 2: Impact assessment (6 hours)
Fill an AI impact assessment for a real agent. (I did not run this.)
Done when: each risk has an owner and a mitigation tied to a code check.
Hard 3: Drift monitoring (6 hours)
Trigger the gate when a new tool, model or domain is added. (I did not run this.)
Done when: a tool addition changes the tier or the evidence list in a test.
Hint
hash the tool list and domain in the manifest and compare.