The consultant Shipped
Re-scores every finding 0–100 against an anchored rubric, blind to who raised it — and is the one actor allowed a global view, because duplicates are invisible from inside a lane.
After the pillars finish, one consultant re-scores every finding against a rubric and a false-positive catalog that live in team data. It writes a verdict and a validated confidence alongside each finding. It never deletes one.
The false-positive catalog it scores against
- Anything a linter, typechecker or compiler would already catch — the pre-checks own that ground.
- Findings about lines the change did not touch.
- Behaviour that is explicitly silenced or deliberately suppressed in the code.
- Speculation about code paths the reviewer could not see.
Why it is the exception to the blindness rule
Every other actor is scoped to one lane. The consultant is not, and that is exactly why marking one finding as a duplicate of another is its job alone — you cannot spot a duplicate from inside a single lane. The separation of scoring from finding came from reading how Anthropic's own review tooling is built; the difference here is that the score orders the report rather than filtering it.