This sort of works if you have only a couple discovery tasks. Once you have a few dependent tasks -- that is, C depends on outcome/method of B, which in turn

This sort of works if you have only a couple discovery tasks. Once you have a few dependent tasks -- that is, C depends on outcome/method of B, which in turn depends on A -- the branching becomes crazy. Frankly if I'm going to spend a couple days on one aspect, I'd rather just implement that aspect. I prefer to identify the risky parts and do them first, and refine estimates along the way. TDD helps because it keeps components isolated, so you don't have to build "in order". Waterfall development with adequate time for planning/discovery up front does for sure give better estimates -- but those estimates are very expensive to get and are usually not all that reliable anyway, especially for multi-month projects.