Chip-design search loops score candidates with a cheap, early result and trust it to rank the same way a slower, more faithful flow would. In the verified-dse programme the reference is "global route plus timing repair". A fair objection is that real chips are detail-routed. We reran five designs (50 ABC pass sequences each, 250 points) through detailed routing and OpenRCX parasitic extraction at a clock 1.5x the baseline period.
What came back, with OpenROAD 26Q3 and Yosys 0.70:
- Global-route ranking agrees with the detail-routed ranking: Kendall tau-b 0.81 to 0.90 on all five designs. Its best pick costs 0.0% on four designs and 2.3% on the multiplier.
- The mapped-but-unplaced rung is the weak one: its pick is 10.1% off on the ALU and 9.2% off on the barrel shifter, which was 0.0% at the tight clock.
- Detail-route slack is a median 0.325 ns more optimistic than the global-route estimate, so absolute periods in the earlier tables are pessimistic. The ranking is what carries over.
- The tool version matters more than the router. The same points rerun under the newer tools correlate with the cached older-tool scores at tau 0.07 to 0.66.
Limits: five small designs, one corner, and no predictions fixed before this run, so read the numbers as descriptions. Details are in research/verified-dse/experiments/detail_route/results_relaxed.txt and the programme LOG.md (milestones 47 to 50).