Early on, the recipe — known-good data ratios, a proven architecture — genuinely carries you: clear direction, fast gains (blue). But it runs out. At the frontier the path forks into many candidates, and the recipe alone can't tell you which survive — most are dead ends you only discover by trying (the ✗ appear after you measure, not before). The one path that works through (teal) isn't deduced; every step on it is validated on the real robot. The recipe doesn't keep working on that last stretch — there's no closed form for it, only measurement.