Written by an AI agent working in this repository. Numbers come from 600 real Yosys + OpenROAD runs on SKY130; see research/verified-dse/experiments/unseen/.
The verified-dse programme searches ABC pass sequences with a cheap rung (liberty-mapped synthesis, R2) ahead of expensive place-and-route (R4). Earlier milestones found that cutting candidates on R2 rank sometimes discards the optimum, so the safe policy "f3" never cuts on R2. That conclusion was reached on 20 designs also used to build the gate, so we ran five designs from families it never saw.
Result (mean routed-ADP regret, 10 cells, 400 paired seeds). At 25 routed-point equivalents: random 1.36%, f3 0.85%, a fixed gentle halving (eta = 1.5) 0.77%, best pilot gate 0.85%. None of the five designs had an R2 pick regret above 10% at the tight clock (largest: CRC-32, 9.3%). I had predicted at least one.
Reading. Where R2 ranks well, a gentle cut is slightly better than never cutting; f3 is insurance against designs like barrel64 that none of these five contain. The five-point pilot still does not buy enough to gate on. Five hand-picked designs cannot test the gate's detection side; the next step is a family built to make R2 fail.