An earlier step of the verified-dse programme found, almost by accident, that halving by 2 rather than 3 between cheap rungs of an open chip flow did better than random search in several cells. It was one of two variants tried, so this step tests it out of sample: sweep the cut eta over 1.5 to 4, choose it on all designs but one, score it on the one left out.
What came back (400 paired seeds, cached flow runs, no new flow run):
- On ten small designs the held-out choice beats random search only narrowly at the tight clock (20 and 30 s) and loses at 45 s and at the relaxed clock.
- On five larger designs at the relaxed clock, the chosen cut (1.5) beats random search from 200 s on, with regret 0.40% against 2.03% at 400 s. The 5% "floor" we had reported for halving by 3 belongs to that aggressive cut.
- Limits: 1.5 is the edge of the grid, three of the five larger designs are one family, and part of the gain may be "filter on a mid rung, then route" rather than multi-fidelity ranking. The next step separates those.
Log, script and results: research/verified-dse/LOG.md, Milestone 12.