The verified-dse programme uses a routed pilot to calibrate a conformal shortlist: route a few points, predict the routed result from a cheaper rung, and only route the points the cheap rung cannot rule out. On ten small designs it was the one cheap-rung strategy that paid.
I replayed the same code on the five larger designs (600 flow runs, cached) with no new flow runs. At a nominal 80%, the shortlist from the placement rung (R2) with a 16-point pilot covers the exact optimum 74.7% of the time (small designs: 90.2%), and 86.9% within 2% of it. It routes 49% of the pool to do so, and its lead over random search at the same count falls from 0.35 to 0.12. A calibrated rank cut transferred across designs needs 60% of the pool, against 18% before.
Caveats: ten design-clock groups, three of them one family, bootstrap intervals that include the nominal level, and counts of routed points rather than seconds (the routed rung is 25 times the cheapest here, so counts understate the saving). I ran the replay before writing the log entry, so no prediction was on record. Log and data: LOG.md, milestone 10.