When a cheap synthesis score ranks designs for a multi-stage search, can a small routed pilot tell you the search will not drop the best design? The verified-dse programme has a rank-displacement certificate for that (T < 1). Earlier milestones described it from the optimum's side. This one reads it from the practitioner's side, on the 40 cached cells (2 clocks, 10 designs, 2 halving starts) with 300 pilots and 200 tie-breaks each.
At a pilot of 16 routed points, 22% of pilots are certified. After a certificate, successive halving had zero regret in 87.3% of draws, against 31.5% after no certificate, and the mean excess over the best design was 1.1% against 7.0%. I predicted a floor of 90% zero-regret and a ceiling of 5% on any cell's mean loss. Both were wrong: the 90% figure is reached only at 24 pilot points (90.7%), and the tight-clock adder_op16 cell from the R2 start is not guaranteed, yet half its pilots are certified and its mean regret is 13.9%.
So the certificate is a good screen and not a bound; a regret certificate stays open. All from a cached table, no new flow run. Code and results: research/verified-dse/experiments/theory/pilot_cert_regret.py.