The verified-dse programme asks when a cheap flow stage can cut candidate synthesis sequences before the costly routed run. The last milestone found that on a fanout-96 design the routed clock period is set by a wire floor the mapped netlist cannot see, while area survives routing (Kendall tau 0.98). That suggested scoring the cheap rung by mapped area alone instead of area x period.
This run replays that idea on 12 designs never used to tune the schedule (24 design-clock cells, 400 paired seeds, no new flow runs). Three predictions were written down first.
- Held: on
fanao96, picking the argmin of mapped area drops regret from about 14% to 0%.bcastmix192goes from 49.9% to 0%. Both are the designs that motivated the rule, so this is not a held-out test. - Refuted: mean tau with the routed score rose (0.38 to 0.58) but mean pick regret got worse, 6.0% to 10.3%.
bcastadd64went from 0.4% to 95% regret at the tight clock. - Held, trivially: inside successive halving with a cut factor of 1.5, the two scorings differ by 0.00 points at 25 routed-run equivalents on every design. That cut keeps two thirds of the candidates, so almost any ordering keeps the best.
One side result: on 12 designs, halving beats random routing by 1.2 points at 25 equivalents, 10 designs better, and survives a Holm correction over 12 tests (p 0.044). The earlier ten-design test did not. Two of the 12 were added after that test, and all designs are ones I chose.
Next is the same replay with a harsher first cut, where the cheap rung's ordering should matter. Code and numbers: research/verified-dse/experiments/r2_area_arm/.