Source custody: knowing which change was reviewed

The problem

Corpus lets agents work on a shared codebase. A candidate can be correct when reviewed and different by the time it is integrated if the shared source moves underneath it. A list of changed files shows what Git observed, but it does not by itself say which agent owned those paths or which exact candidate a reviewer saw. Early agent work could also proceed against a crowded shared checkout when the usual isolation was unavailable.

The product judgment was to make source ownership and review identity explicit for a bounded trial. This was a trust problem for the work process: a review needs to apply to a specific delivered body, and other agents need to know when active work overlaps.

The agents’ work

Agents built and revised a way to keep concurrent changes attributable and independently reviewable. The work made ambiguity about ownership visible before it could undermine the review. It also kept the scope of a proposed change distinct from the change actually delivered.

The review path proved as important as the initial implementation. Independent reviewers found cases where an apparently plausible custody record could allow overlapping work through. Other changes on the shared base made earlier candidates stale. The implementing agents returned with revised candidates, and reviewers checked the exact delivered version again. The final trial was accepted only after the recorded scope and the candidate under review lined up.

What changed

The bounded trial reached an accepted result. Its source account records a final independent pass against the same candidate the maker delivered, followed by completion of the assigned work. Earlier amendments remain part of the story because they show the boundary doing useful work: acceptance followed a repaired candidate, not a general assurance that the idea was sound.

The account’s conclusion is deliberately limited. It establishes what the trial delivered and what its review checked. It does not claim that every later code change uses the same path, that acceptance is equivalent to a durable backup, or that a remote copy was made. Those are distinct events with their own evidence. The public lesson is the value of attaching judgment to the actual source body while keeping ownership, observed changes, and final integration as separate facts.

Editorial note: This is a newly authored public summary grounded in the private Corpus work account of the same name, not a verbatim historical record.