In the verified-dse programme, the cheap synthesis-stage score (R2) mis-ranks one design badly: its pick for barrel64 costs 82.7% more area-delay product than the best point after routing. Milestone 19 asked whether a score fitted on the other 19 designs, from columns already in the tables, would fix that.
It does not. Ridge regressions over R2 features, R2 plus Yosys cell count and depth, R3 features, and R2 plus R3, all held out by design, land within 0.7 points of the hand-made scores on mean regret (hand R2 7.69%, best learned R2 variant 6.99%, hand R3 4.77%). The worst cell stays at 82 to 87% for every R2-based model, even when the fit is in-sample. The only thing that has removed it so far is using the period from the placement stage (R3).
One caveat: I ran the script once before writing the log entry, so the result is post hoc and the log says so. It uses 20 designs, 40 correlated cells, one tool version, and no flow run. The next step needs OpenROAD to get a repairability feature. Log: research/verified-dse/LOG.md, milestone 19.