ADR 0006: Persistence boundaries

• Status: Accepted
• Date: 2026-10-03

ContextCopy Anchor Link

A database in every evaluation adds failure modes, latency, and mutable state. Correlation knowledge also needs an authoritative, reviewable source for field aliases, context terms, calibration, reviewed outcomes, and provenance. Australian address workloads additionally need relational and geospatial storage.

DecisionCopy Anchor Link

Include PostgreSQL from the foundation for correlation knowledge and provenance, then expand it for justified control-plane functions. Keep authored policy in Git, immutable artefacts in OCI when introduced, snapshots in evaluator memory, and audits in an external sink. Compile approved PostgreSQL state into immutable snapshots; do not query it in the normal evaluation hot path. Use a separately versioned PostgreSQL/PostGIS schema or database for G-NAF/address data.

ConsequencesCopy Anchor Link

• Evaluation survives database outages with its last-known-good snapshot.
• The initial schema and migration workflow are foundational infrastructure even while dynamic tenant bindings and approvals wait for a control plane.
• Correlation/reference data must compile or cache into snapshot-safe structures.
• pgvector is excluded until a measured retrieval need exists.