Can a statistical guarantee make a cheap proxy safe in chip design search? In the verified-dse programme we replayed 20 design-by-clock groups (10 small designs, about 50 ABC pass sequences each, run through Yosys, ABC and OpenROAD on SKY130) and asked: if a 10 to 16 point routed pilot calibrates a one-sided conformal bound on routed area x delay, does the shortlist still contain the best point at the stated rate?
Result. With the mapped-netlist rung, the optimum was kept 0.99 of the time at a nominal 0.90 (0.90 to 0.92 at nominal 0.80), routing 42 to 57% of the pool; random search with the same number of routed runs found it 0.53 to 0.65 of the time. With the generic-gate rung, coverage was 0.65 at nominal 0.80, no better than random search of equal cost. The per-point guarantee does not transfer to the argmin, which is selected for having a low residual.
Caveats. One corner, 20 groups from ten small designs, wide intervals, global-route estimates rather than detailed routing. No theory yet for the undercoverage. The code (experiments/conformal/replay.py, 35 s, no flow re-run) and the log with the plan written before the run are in research/verified-dse. Written with AI assistance.