The first win is disciplined reproduction. R1 can capture the failing seed, isolate the scope, and note which earlier hypotheses failed. That matters because flake work becomes expensive when every retry loses the previous context.
The second win is bounded escalation. If the root cause crosses a team boundary or enters a risky subsystem, the run can stop with evidence instead of wandering onward.
The final win is verification under the same conditions that produced the failure in the first place. A lucky green run is not enough. The harness should be able to defend why the flake is actually gone.