BHANUJBHANUJ
v1.0.3

Persistence

AI Governance control plane separates authoritative relational state from the projected graph and an append-only audit log, each optimized for its role.

Storage Layers

Persistence is split across three layers with clear ownership: the relational store holds authoritative records, the graph store holds a projection for traversal, and the audit log records every state change.

AI Governance Platform / governed event stream
decision / decision-991event order
  1. 01
    commit relational state

    The authoritative policy, subject and decision write completes transactionally.

  2. 02
    append audit event

    Actor, version and outcome are captured as immutable operational evidence.

  3. 03
    project graph state

    The graph receives a deterministic downstream projection for traversal.

  4. 04
    reconcile projection lag

    Any temporary graph lag is detected and converged without blocking the decision.

governed outcome

COMMITTED

Relational truth preserved · audit immutable · graph rebuildable

Relational Store

AI Governance Control Plane currently uses durable SQLite databases for authoritative control plane records and shared MCP execution audit records. All writes are transactional, and the graph remains downstream of those authoritative stores.

Dataset Artifact Store

Dataset metadata lives in the authoritative registry, while uploaded dataset bytes live in a separate S3-compatible object store. Each upload is validated as a stream, hashed, and written under a deterministic tenant-scoped key. The registry stores the immutable URI, checksum, record count, schema version, and lifecycle state—not the file body.

SeaweedFS provides the local development implementation and Filer UI athttp://localhost:8888. Production deployments can use AWS S3 or another compatible object store without changing asset semantics.

Graph Store

Neo4j holds the deterministic projection of relational state. It is rebuildable at any time and never accepts writes that do not originate from projection, keeping it strictly downstream of the relational store.

The synchronization worker records projection work durably, reconciles only semantic drift, and retains soft-deleted entities for history. Read the Neo4j Operations guide to run, inspect, and recover the projection.

Audit Log

Every mutation appends an immutable record capturing who, what, when and the resulting version. The log is never updated in place, making it a reliable basis for compliance and forensic review.

Rebuildable by design

Because the graph and audit projections derive from relational state, they can be rebuilt after any incident without data loss.

Consistency Model

Relational writes are strongly consistent. The graph projection is eventually consistent and converges through continuous reconciliation, so a brief lag in the graph never blocks a decision from committing.