Key points
- Set retention by evidence class, business purpose, sensitivity, and review obligation rather than applying one indefinite archive period.
- Delete or expire primary records, indexes, exports, caches, backups, and recovery copies under one documented lifecycle.
- Export a self-describing manifest with schema versions, checksums, integrity links, scope, exclusions, and verifier instructions.
Classify evidence before assigning a retention period
Separate operational telemetry, security events, model and tool evidence, approval records, and protected source references. They can carry different value and risk. Define the purpose, owner, minimum and maximum period, deletion trigger, access class, and any hold process for each category.
A longer period is not automatically safer. Keeping full prompts, retrieved documents, or personal data after the review need has ended expands exposure. Retain canonical evidence and references where they are sufficient, and keep protected source material under its original access policy.
Apply lifecycle rules to every recoverable copy
Primary tables are only one copy. Search indexes, analytics stores, export buckets, caches, dead-letter queues, snapshots, backups, and temporary review workspaces need the same lifecycle inventory. Deletion completion should be recorded without recreating the deleted payload in the deletion log.
If backups expire later than active records, document that delay and prevent ordinary users from restoring expired evidence. A restore operation must reapply tenant isolation, legal holds, and the current access policy before records become queryable.
Create an export that can stand on its own
An export should state the tenant or case scope, time window, query criteria, event count, schema versions, files, checksums, sequence range, integrity roots or signatures, generation time, and exporting identity. Include a machine-readable manifest rather than relying on a folder name or a PDF cover page.
Document redactions and exclusions explicitly. A reviewer should be able to tell the difference between an event that never existed, an event outside the selected scope, and a protected field intentionally omitted from the package.
- Verify file checksums and manifest signatures in a clean environment.
- Recalculate sequence links or Merkle proofs from the exported records.
- Reject unsupported schema and canonicalization versions.
- Record the reviewer, verifier version, result, and any exceptions.
Test holds, expiry, restore, and failed export paths
Exercise retention expiry during an active case, a hold applied just before deletion, partial export failure, duplicate export request, restored backup, revoked reviewer access, and a deliberately modified package. Reconcile counts between the evidence ledger, export manifest, and object storage.
The retention schedule and legal hold process must be reviewed by the organization responsible for the data. Technical controls can enforce an approved policy, but they do not determine the applicable legal requirement on their own.
References and review sources
- AI Risk Management Framework (AI RMF 1.0)NIST
- AI RMF Playbook: MeasureNIST AI Resource Center
- Building Evaluation Probes for Agentic AINIST