CAD.1:9 - Consequences
The result is a smaller set of explainable architecture candidates. The whole-problem comparison can rule out a split before tuning even when its individual controllers look promising. The cost is retaining the original problem long enough to compare whole behavior instead of judging each box independently.