Audit-log integrity, data residency and sovereignty routing, and supply-chain security for model artifacts.
Governance for an AI system answers a different question from access control. Access control decides who may call the model; governance decides whether you can later prove what happened, where the data went, and that the model you ran was the one you approved. The three pages here cover those proofs: tamper-evident audit logs, region-aware routing for data residency, and integrity checks on model artifacts before they are loaded.
The controls are orthogonal and compose. Residency routing decides which region processes the request, supply-chain verification decides whether the model in that region may be loaded at all, and the audit log records both decisions together with the request metadata so an auditor can reconstruct them.
Audit Log Integrity moves from "we log every query" to "we can prove the log was not edited": append-only storage, hash chaining of entries, periodic anchoring of chain heads to a separate system, and what to record without logging sensitive prompt text. Data Residency and Sovereignty Routing adds a second routing axis next to sensitivity tiers, so that a request from a jurisdiction is processed by an endpoint in that jurisdiction and failover never silently crosses a border. Supply-Chain Security for Model Artifacts treats weights like any other dependency: pinned revisions, checksums, signatures, safe serialization formats instead of pickle, and verification at load time.
| Risk | Control | Evidence produced |
|---|---|---|
| Log entries altered or deleted after an incident | Audit log integrity | Verifiable hash chain |
| Personal data processed outside its permitted region | Residency routing | Per-request region decision |
| Tampered or substituted model weights | Artifact supply-chain checks | Recorded hash and signature at load |
| Unclear which model version answered | Artifact checks + audit log | Model digest stored with each request |
An audit log is only as good as the guarantees you can make about it under adversarial review. "We log every query" is a starting point; "we can
Sensitivity-tier routing decides which model handles a request (cloud vs. on-prem). Residency routing is a second, orthogonal axis: which region processes…
On-prem inference reduces the exposure of prompt and response data — but only if the model you are running is the model you think you are running. Model…