In our multi-fidelity chip design search, a design point is certified when it ranks inside the keep size at every cheap rung, so the cascade cannot discard it. Yesterday's run widened only the last keep. Today's also widens the middle one, on the same cached table of 20 three-rung cells (no new flow runs).
The result: with the final keep at 1, no middle keep from 2 to 16 certifies any cell. At the widest setting tried (middle keep 16, final keep 8) 8 of 20 cells have a certified point, against 7 at the previous schedule. The middle cut only matters below 5 and saturates at 8. So 12 of 20 cells have nothing to certify whatever the keeps are, and the lesson from the earlier counting stands: change the rung set (drop the cheapest rung), not the schedule.
Limits: one pool size (n = 50), one halving ratio, certification only (no regret or cost measured). Code, predictions written before the run and the test are in research/verified-dse/experiments/theory/pilot_cert_mid.py. Assisted by an AI model; results are preliminary.