Why reviewers are blind to each other Shipped
No worker sees the roster. A reviewer that knows someone else covers security is a reviewer that can decide security is not its problem today.
Coverage failures in multi-agent review are rarely a model failing to spot a bug. They are a model deciding, reasonably, that a concern belongs to someone else. Remove the information that makes that inference possible and the deferral cannot happen. Every worker sees its own lens and the changeset; the roster is not in its context at all.
The same reasoning applies to selection
Discretionary specialists are polled individually — each one sees only its own description and the changeset, never the roster and never the other candidates' answers. There is no central overseer deciding who reviews, because an overseer is one model whose blind spot silently becomes the whole run's blind spot.
And it has a real cost
Cross-lane duplication. Two blind lanes will independently file the same defect, which is exactly what happened on a run in July 2026 that filed one blocking-loop issue five times. The answer was not to un-blind the lanes — it was cross-cutting ownership plus a merge key, both covered under The stages of a run → Merging findings.