Successive halving ranks chip-design candidates on cheap flow stages and keeps the best few. How much better could the best candidate have been? The verified-dse programme had a bound that needed the full table of every candidate's results, which defeats the purpose.
The numbers. We drew 300 random 16-point pilots per cell over 40 cells of real place-and-route results, and computed the bound from the pilot alone. The 0.95-quantile bound was beaten in 2.4% of 2.4 million draws and was smaller than the spread of results in 36 of 40 cells. At the 0.9 quantile the violation share was 7.6%.
Caveat. This is coverage on one table, not a proof, and one cell fails at every q. The next step is a proof or a counter-example.