Three structural problems that make downstream systems — human or agentic — unsafe to act on.
Debug logs are not audit trails. When a reviewer asks why the system decided what it decided, the answer should not require access to internal tooling, log aggregation, or a support ticket. If the explanation is not in the response, the response is not complete.
Most identity systems commit to a resolved entity and overwrite it as new records arrive. The source data may persist — but the entity as it stood at decision time does not. When a regulator asks "what did the system treat as this person last October?", that boundary is gone.
"Show me the evidence basis from last Tuesday" should be a one-call operation. In most systems it requires re-running the pipeline against an archived snapshot, hoping the snapshot exists, and hoping the pipeline has not changed. That is forensic reconstruction — not auditability.
Every mention the system has ever seen is individually addressable. Query a cluster and receive the mention_id for each contributing mention. Pull the mention and you get the original source passage, its provenance, and the attributes extracted from it — no intermediary.
No observation ever overwrites another. New evidence appends to the record. The original mention that drove a clearance decision six months ago is still there, exactly as the system saw it — accessible to the same API surface that created it.
Both fields arrive in the response envelope without a second call. Downstream systems — human or agentic — can branch on evidence quality, escalate underdetermined cases, and cite the basis of a decision without accessing internal tooling.
A long-running investigation pins a graph version. On refresh, you get a structured report of how the entities and conflicts in your working set changed — splits, merges, weight changes, anchors — so you can absorb new evidence deliberately instead of being silently re-resolved underneath you.
Stream A (Operational) is a metadata-only record — operation, principal, target, graph versions, outcome — with subject-identifying payload stripped at write time. It's GDPR-clean by content: once the mention store compacts, its identifier references no longer resolve to personal data.
Stream B (Forensic) mirrors that metadata and adds a pointer (payload_artifact_uri + hash) to the full payload, held in a WORM-capable store — the regulator-facing reconstruction surface. Which stream a mission writes to is fixed by its legal_basis: general → lean only, regulated → forensic.
Optional cryptographic chaining links entries in either stream — each entry’s hash incorporates its predecessor’s, so deleting or altering any entry breaks the chain forward.
The forensic stream is the complete, ordered record a regulator’s tooling can reconstruct decisions from — not a pipeline re-run.
The cluster layer that operates over the evidence record. Probabilistic cluster decisions, non-destructive merges, typed conflict resolution — all grounded in the append-only evidence substrate.
Read the pageDeterministic session pinning over the evidence layer. Every query inside a session resolves against a pinned graph version — reproducible against the same evidence basis weeks later.
Read the pageHow the append-only model interacts with data-subject rights requests (GDPR deletion / rectification), data residency posture, and the forensic audit stream under regulatory review.
coming soonThirty minutes with engineering. We run a live query, show the response envelope with blocking evidence and resolving attributes populated, and walk the audit stream for an adjudication decision.