Private inference boundaries
Review the architecture with one real request trace
Use a controlled request to compare the documented data path with the deployed one. Then exercise an error and a fallback to expose routes the main diagram may omit.
In this article
Choose a representative synthetic request
Include the types of data the application normally handles, such as a user question and retrieved passage, but use fictional content with a unique marker. Record the expected destinations and permitted copies.
Follow the request from the application into retrieval, inference and response storage. Inspect endpoint configuration, identity and payload scope at each stage. A network diagram should be supported by deployed evidence, not accepted as its own proof.
Check whether preprocessing or evaluation adds a separate model call. Those calls belong in the review even if the user sees only one answer.
Inspect the identity and routing rules
Verify the runtime uses the intended endpoint and private route where required. Test an identity that should be denied and an unapproved route if the environment permits that controlled check.
Review destination configuration for fallbacks. A primary endpoint can be correctly constrained while an emergency fallback sends the same data elsewhere under broader permissions.
Keep service availability and data-handling claims tied to current documentation for the actual configuration. Avoid assuming that a regional hostname establishes every required residency or retention property.
Trigger diagnostic paths
Cause a validation failure and a service timeout, then search the known logs and traces for the marker. Confirm content appears only where the review permits it.
Inspect access to those records and any export function. An operational dashboard can be restricted while its downloadable report is routinely shared more broadly.
Check retention settings and source-to-derived-record relationships. The team should be able to identify which copies need attention when a document is removed or a data owner requests investigation.
Record the gaps precisely
Separate verified controls from unresolved questions. A missing retention decision for an evaluation store is more actionable than a general statement that privacy needs improvement.
Assign each gap to the team that can resolve it and define the evidence needed. The review is complete when the documented path matches observed behaviour for representative and exceptional requests, with explicit limits on what the evidence establishes.
Primary sources
AWS: Bedrock private endpointsAWS: model invocation loggingReferences checked 11 September 2026.