Retention and deletion

​

Retention depends on the record

Deletion runs in the background. Active tests may need to finish before their linked evidence can be removed. Exported files and copies in your own application are outside Bench’s deletion controls.

​

Delete a production trace

Open AI systems → your system → Production, inspect the trace and choose Delete trace. Read the confirmation before proceeding. Deleting a trace erases the trace, its spans and checks, and saved test cases that explicitly record that trace as their source. It also erases affected context snapshots and associated detailed evaluation artifacts. This can remove more than the visible trace because a summary or saved result may contain the same evidence. A hashed deletion marker prevents delayed ingestion retries from restoring the same trace identity without retaining the original identifier. New, separately identified events are separate records. Cases copied manually without the original trace reference are not automatically located by content similarity. Review those copies separately. Trace expiry follows this same provenance-based path. A case explicitly derived from a trace can therefore expire with the source trace. Do not assume that adding an incident to the library makes its original production payload permanent.

​

Remove context, datasets or connections

Use Manage sources to remove selected saved context. Source removal deletes its active and historical evidence, marks affected snapshots as erased and clears associated detailed run artifacts and reusable scoring-cache entries. Aggregate historical scores remain; erased evidence must not be presented as available for reinspection or reuse. Deleting a dataset also removes test-library cases linked to that dataset and applies source erasure to those cases. Disconnecting an imported-source connector removes its connection and associated evidence and applies the derived-artifact cleanup. It does not delete the upstream project’s original data. Because runtime evidence or future automatic-run settings may embed removed context, source erasure also removes the system’s saved runtime reports and automatic-runtime configuration, and clears retained hosted-job content. Configure a new suite using the remaining approved context before enabling automatic testing again. Relevant queued/running work or unsettled evaluation reservations can block these operations. Finish or cancel the work, allow reservation settlement, then retry. A blocked deletion must not be interpreted as a completed erasure.

​

Remove an AI system

Deleting a system removes its structure, context and attached datasets, and scrubs its detailed evaluation artifacts before retaining aggregate run history. Runtime records attached to that system are also removed or cleared. Minimal discovery markers prevent the same explicitly deleted static system from being recreated by the next scan. Raw production traces belong to a repository and branch. Deleting a system does not itself delete every trace from that repository; delete those traces separately or allow their retention process to run. It also does not delete the source repository or revoke the GitHub App installation. Manage GitHub access opens GitHub’s installation settings when you need to change that access.

​

Backups and account deletion

Deleting a record removes it from the live application through the applicable deletion process. Backup expiry is separate. Contact Bench for account-wide deletion or contractual retention requirements. See security and privacy controls to collect less data.