A candidate spends two minutes on requirements, then draws a CDN, a cache, three microservices and a queue — all before doing any arithmetic — and can't justify the queue when asked why it's there. Which phase actually failed, and why can't the phases run in a different order?

The visible symptom — no number behind the queue — shows up at the end, but the failure happened two phases earlier: estimate was skipped. The candidate sketched from habit ("scalable systems have a queue") instead of from a number that named a bottleneck the queue solves. The five phases — clarify, estimate, sketch, deepen, defend — can't be reordered because each one consumes what the previous one produced: you can't judge whether a sketch is appropriately simple without a scale to be simple at, and you can't defend a decision's cost before the decision was made concretely enough to have a real cost.