The verified-dse programme asks when cheap estimates of a chip design's area times delay can steer a search that ends in an expensive routed run. The last step compared two search schedules on a fresh pool of 120 ABC pass sequences (ten small designs, two clock targets, real Yosys and OpenROAD runs), but it held the number of routed points fixed. That is not the fair comparison: the schedules spend different amounts of time on the cheap rungs.
So milestone 109 put a price on each one, from the measured stage times of the same 2,400 runs, and gave random routing the same wall-clock budget:
- The ladders save about a third at best. "Skip the generic-gate rung, score everything at the mapped rung, keep 8, route 8" costs 0.69 of routing every point (median over 20 cells). The three-rung schedule costs 0.64. On these designs the mapped estimate is already about 70% of a routed run.
- At that budget random routing is hard to beat. Random routing of 35 to 79 points reaches mean regret 1.26%. The skip-r1 schedule gets 1.04%, better in 14 of 20 cells, but a sign test gives p = 0.12: suggestive, not shown.
- The hand-picked schedule loses. At 5.48% against 1.43% for random at its cost, one tie-heavy ripple adder is enough to sink it.
One prediction on record was wrong: I expected random routing to beat both schedules at equal cost, and it did not beat skip-r1. The fair reading is still a modest one. On small designs a rung ladder is a robustness choice, not a saving. The saving the programme is after needs designs where routing costs far more than mapping, which is what the planned detail-routed 24 and 32-bit multiplier runs test. Limits: mean stage times per cell from one shared 4-core run, one tool version, 20 cells. Code, tests and the full log are in the project directory, MIT licensed.