Part 3 of 12 · 1 min read
The evidence list comes from the tier, not from memory
A lookup table maps each tier to the documents and checks it needs, so a system's obligations are read off, never recalled from habit.
Think of a packing list that changes with the trip. A weekend away needs three items. A month abroad needs visas and vaccination records too. You read the list for the trip you are actually taking. You do not pack from memory.
each tier maps to a fixed set of required evidence. Documents such as a model card and checks such as human oversight -- and that mapping is looked up, never recalled from habit, so nobody has to remember which tier needs what.
A worked example. The minimal tier needs only 2 items (a model card and a retention policy). Limited needs 4 (adding an evaluation report and AI disclosure). High needs 8, including a data protection impact assessment, decision logging and human oversight, plus a separate two-person sign-off checked outside the lookup table entirely.
The key lines. From code/mini-compliance/compliance.py:
# from code/mini-compliance/compliance.py
OBLIGATIONS = {
"model_card": {"high", "limited", "minimal"},
"risk_assessment": {"high"},
"dpia": {"high"},
}
def required(tier: str) -> list:
return sorted(o for o, tiers in OBLIGATIONS.items() if tier in tiers)
Read it as: each obligation names every tier it applies to, and required() just filters that table. Adding a new obligation anywhere in the system means editing one dictionary entry. It never means hunting through code for every place a tier is mentioned.
Model card, impact assessment
- Model card: documentation of a model's purpose, data, evaluation and limits.
- Impact assessment: a structured analysis of who is affected and how.
Example: required("high") returns 8 items including model_card; internal-summariser-undocumented fails the gate specifically for missing that model card.