We are testing whether a search over open-source chip flows (Yosys, ABC, OpenROAD, SKY130) can save time by scoring candidates on cheap fidelity rungs and routing only the survivors. Last step showed that a one-shot screen does not beat random search at equal wall-clock. This step replays successive halving (keep the top third at each rung) on the same 1,200 measured runs, 10 designs, 400 paired seeds per cell.
- Starting at the generic-gate rung is the worst strategy: about 10% routed regret at the tight clock against 0.94% for random search at 45 s. That rung ranks candidates like noise (Kendall tau-b median -0.04), so each cut throws away the best point two times in three.
- Starting at the liberty-mapped rung wins on 6 to 8 of 10 designs at a relaxed clock at small budgets, but loses 5 to 12% on two designs, so its mean is still worse than random.
- Spending 6 routed evaluations as a pilot to choose the starting rung recovers the designs plain halving loses (alu8 16.9% to 0.8%), but most of that gain is just skipping the generic-gate rung.
Caveat: the pool is about 50 points per design, so halving stops using its budget beyond roughly 30 s and the later columns favour random search. The designs are small, so a routed point costs only about 8x a generic one. A larger design set is the next step. Code, log and tables: research/verified-dse/.