> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cortex-labs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Understand retention, deletion, and legal holds

> Distinguish rolling expiry, protected evidence, removal from selection and exceptional erasure.

Retention determines what remains available. Removing a resource from selection, deleting content and erasing protected bytes are different operations.

## Retention and protected references

Dataset time and byte limits govern rolling source content. Required retained evidence has a separate storage obligation.

Saved work can protect exact inputs and dependencies. A visible citation or world-model mention alone creates no hold.

| Action or state | Meaning |
| - | - |
| Remove last collection membership | Stops future workspace selection. It does not delete the dataset. |
| Archive a definition or app | Removes it from ordinary new use while preserving applicable history. |
| Reprocess | Creates current derivations from selected source bytes. Retained earlier evidence remains distinct. |
| Rolling expiry | Removes eligible temporary source content and unprotected derivatives. |
| Protected reference | Keeps required evidence resolvable while retained work depends on it. |
| Legal hold | Blocks retention expiry and erasure for its covered subject until lifted. |
| Exceptional erasure | Uses a separate grant and audited procedure, leaving a provenance tombstone. |

Physical deletion requires both the applicable retention condition and no protected reference. A refused delete must not remove stored bytes.

## Inspect blockers before deletion

Open the resource's supported deletion or retention controls. Read the named references and holds. Identify the owner of each retained dependency.

A workspace deletion removes retained work owned by that workspace. It does not remove another workspace's retained work. A dataset shared through collections is one dataset, rather than a copy per workspace.

Retain the intended output and evidence before changing rolling limits. Reports cannot depend solely on temporary windows or an unverified external archive.

## Legal holds and erasure

A legal hold can cover a dataset, artefact or review item. Use the authorised hold procedure and record its subject and reason.

Erasure requires separate granted authority and an impact assessment of invalidated evidence. It reaches retained derivatives, including platform-held previews, embeddings, exports and index projections.

The tombstone preserves provenance and security classification. A published result keeps its publication record and shows the erased evidence limitation.

## Audit retention and event replay

Audit retention governs how long immutable audit events remain. Durable-event replay settings govern processed operational events. Neither setting grants authority to erase retained customer evidence.

An expired or inaccessible source limits reproducibility. State the gap explicitly; a retry or restored index cannot recreate missing original bytes.

Use [Streaming data and retention](/vantage/bring-in-data/streaming-and-retention) for collection windows and [Grant privileged authority and audit scope](/administration/people-and-access/privileged-authority) for erasure authority.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.