Skip to content

Remaining-Risk Register

Honest status after the knowledge-lifecycle work in issue #6. Exact provenance, deterministic reconciliation, authoritative-first writes, projection cleanup, hybrid retrieval, reproducible snapshots, the agent MCP contract, Confluence lifecycle handling, ontology governance, and governed community lineage are implemented and tested. The items below remain operational or migration risks; none is a substitute for a production readiness review.

Cutover and Deployment

  • Authoritative Fabric writes remain an explicit rollout decision. VERA_MEMORY__FABRIC_WRITE_MODE supports legacy, dual, and fabric, and defaults to legacy. fabric makes PostgreSQL facts authoritative and drives graph projection only from committed outbox work. Production rollout still needs per-scope verification, queue monitoring, and a tested rollback plan.
  • Database role enforcement remains deployment-gated. The read and worker paths assume vera_trusted and vera_worker with per-transaction SET LOCAL ROLE when VERA_DB__ROLE_ENFORCEMENT=true. The login role must be granted those roles before the setting is enabled. Cross-tenant security testing belongs to the production-readiness follow-up.

Retrieval

  • Vector retrieval remains provider- and rollout-dependent. When VERA_MEMORY__VECTOR_SEARCH_ENABLED=true, full-text and vector candidates are fused rather than selected as alternatives. Embeddings are versioned per chunk, provider, model, and model version, and new chunks are embedded on the live path. Existing chunks still require an operator backfill for the selected model, and production must supply provider credentials, validate dimensions, and monitor embedding failures.
  • Old vector snapshots require their pinned query embedder. Snapshot rows retain exact chunk fields and vectors. Replaying vector ranking also requires the provider, model, and model version recorded by the snapshot to remain configured. Vera fails closed when the active retrieval implementation differs from that pin.
  • Assembler contracts require an explicit version bump. A snapshot pins the assembler version, which covers its scoring algorithm, default weights, diversification, and packing. Changing any of those without bumping the version would make replay claims inaccurate.
  • The committed golden set covers deterministic fixtures at test scale. It gates hit@k, MRR, nDCG@k, and citation rate against deterministic fixtures. Domain coverage, latency, and ranking quality still need measurement with production-shaped data and traffic.

Graph and Communities

  • Community construction remains an operator-run, potentially expensive step. Each run rebuilds the active fact projection from PostgreSQL, records normalized lineage in community_fact_lineage, and publishes summaries marked derived; authoritative fact search never presents those summaries as evidence. Scheduling, LLM cost controls, and freshness targets remain deployment concerns. Saga construction remains deferred.
  • The graph is a rebuildable projection. PostgreSQL and S3 remain the recovery sources. Projection drift is detectable and stale facts are removed incrementally, but operations must alert on queue lag and drift and retain a tested rebuild procedure.

Historical Migration

  • Free-text legacy episodes remain queued for controlled re-extraction. Episodes without structured triples are counted as needs_review rather than assigned invented evidence. Recovering them requires controlled re-extraction from immutable source content.
  • Backfilled legacy objects remain scalar values. The backfill does not reconstruct the object side of historical entity-to-entity graph edges. Re-ingestion is required where that relationship structure matters.

Production Operations

  • No production-scale benchmark or recovery drill has been run here. The benchmark_fabric harness reports context-pack latency percentiles, but target volumes, concurrency, backup/restore, retention, observability, and incident recovery require a production-shaped environment.
  • Production security is a separate gate. MCP authorization, tenant-isolation adversarial tests, data-poisoning and prompt-injection defenses, and operational abuse controls are not claimed by issue #6.

These production concerns are tracked in the dependent production operations and security readiness EPIC.