Four Planes of Compliant ML Architecture

DATA PLANE: STORAGE AND LINEAGE

Stores raw data, training datasets, and features with complete lineage. Every piece traces to source and consent basis. Requirements: immutable access logs, automatic retention enforcement, encryption. When deletion requests arrive, identifies all locations with that user data.

CONTROL PLANE: CONSENT AND POLICY

Manages consent records and policy enforcement. Maintains registry mapping users to permissions per purpose. Before processing: "Can I use user X data for purpose Y?" Handles DSARs—orchestrating retrieval, deletion, or portability across systems.

PROCESSING PLANE: TRAINING AND INFERENCE

Runs training and inference only after consent verification. Pipelines record which data points contributed to each model version. Inference checks consent before personalized predictions—if revoked, serve defaults.

AUDIT PLANE: LOGGING AND EVIDENCE

Captures immutable evidence for regulators. Every data access, consent change, and processing decision logged. Provides timestamped proof of deletion. Audit logs often persist longer than the data itself.