# See Corpus at work

Corpus gives agents ways to recover earlier work, review changes and preserve corrections. Start with these actual sequences: a question, an action, a judgment and a result. Each links to the fuller account, including what went wrong.

## Recover a half-remembered investigation

The Human remembered comparing several agent tools but had forgotten one name. Reeve searched the records, encountered noisy results, tried more specific project names and followed source records and their corrections.

The answer recovered OpenClaw and the earlier comparisons without turning them into a performance ranking. It also named ideas Corpus could learn from the other projects. This was an existing agent sitting with conversation context, not a fresh-successor test.

**Result:** a usable answer with a source trail and its limits. [Follow the actual exchange →](/recall/)

## Stop a change that already passed review

Agents needed to know who could change which files without colliding with other work. An early custody implementation checked for overlapping claims and then recorded ownership as separate steps. Two writers could slip through that gap.

A reviewer passed the candidate. The coordinating agent held that approval and required the check and claim to happen atomically. Further amendments followed; the final candidate received a separate review before the trial was accepted.

**Result:** a bounded ownership mechanism, not infallible review. Its next use exposed a legacy-record problem that required another repair. [Read the custody case →](/work/source-custody/)

## Let Human judgment change the account

In a masked text comparison, model judges preferred leaving a passage unchanged. The Human preferred a short clarification. The account was corrected without rewriting the original candidates or model rankings to match that later choice.

A subsequent check found the older assessment still surviving in an overview. Preserving a correction had not automatically propagated it everywhere.

**Result:** a recoverable disagreement and a corrected assessment—not proof that one wording is generally better. [See the Human judgment case →](/system/#selected-continuity-and-correction-cases)