Mapping to governance frameworks
This taxonomy supplies technical vocabulary for describing model changes. It is not legal advice and does not determine whether a change is a substantial modification, a significant change, or a reportable device change. Those determinations rest with manufacturers, regulators, and notified or auditing bodies under the applicable framework.
What each dimension can document
| Dimension | Question it answers |
|---|---|
| D1 Mechanism | What changes? |
| D2 Goal | Why adapt? |
| D3 Data Requirements | What data is needed? |
| D4 Persistence | How long does the change last? |
| D5 Scope | How much of the model structure is modified? |
| D6 Model Type | What model is being adapted? |
Four frameworks
NIST AI RMF
Its lifecycle stages and core functions may be mapped to adaptation goals: the operate-and-monitor stage invokes drift remediation and continual adaptation, which can be separated by persistence (D4) and structural scope (D5). Data-stage obligations connect to D3.
EU AI Act
Risk-based requirements and post-market obligations may be informed by goal, persistence, and scope. A compute threshold can identify large fine-tuning events but may not capture low-compute interventions that materially alter behaviour. Derived artifacts with fused composition may heighten the need for provenance documentation.
EU MDR / IVDR
Post-market surveillance obligations may draw on D1, D4, and D5 for technical documentation, and on D2 and D3 for benefit-risk and clinical evaluation. A change driven by drift remediation can be distinguished from one driven by task specialisation even where mechanism and persistence are identical.
FDA PCCP
Modification description may map to D1, D4, D5 and D6; the modification protocol to D3 plus separately stated validation criteria; impact assessment to D2 and D4 alongside intended-purpose analysis conducted outside this framework.
Start from a technique
Every technique page carries its full profile, so the documentation dimensions above can be read off directly.