The verified-dse programme searches over ABC logic-synthesis pass sequences in an open chip flow (Yosys, ABC and OpenROAD on SKY130). Its working rule, the clipped ladder, maps every candidate, ranks them by a cheap mapped score, and routes only the best 8. It switches on only when routing one point costs more than 3 times as much as mapping it (theta = 3); below that it routes random candidates.
The last run found the ladder beating equal-cost random routing on 21 of 25 new designs. It also noticed something after the table was read. All four losses were at the tight clock target, and three were the three cells whose cost ratio was lowest, just above theta. If that held up, the fix would be free: the clock is known and the ratio is measured on one point before the switch decides. Route randomly at the tight clock when the ratio is low.
A pattern you spot after reading a table is a hypothesis, not a finding. So this run wrote down six predictions and a test pool first, committed them, and only then computed anything.
The test
The test pool is every earlier cell where the ladder runs: 99 cells from milestones 119 to 137. None of them suggested the reading, and none tuned the ladder, which was fixed at milestone 115. Milestone 138's 25 cells, where the reading came from, are reported separately and count in no test.
What held and what didn't
The clock half is refuted. The ladder fails (loses to random by more than a point) on 14 of 51 tight cells and 11 of 48 relaxed ones (Fisher p = 0.39). Pairing the two clocks of the same design, the tight cell is the worse one on 19 designs and the better one on 23 (sign p = 0.64). With 13 of milestone 138's 25 cells at the tight clock, all four losses landing there had about a 6% chance by luck alone. That looks like what happened.
The threshold half holds. Failing cells have lower cost ratios than the others: the AUC is 0.67 (permutation p = 0.006), and 16 of the 25 failures have a ratio of at most 4.
But the guard it suggests saves nothing. The band just above theta holds the ladder's biggest wins as well as its losses. Over the 45 cells with a ratio between 3 and 4, the ladder and random routing tie on the mean (6.20% against 6.03% regret). Above 4 the ladder has 2.65% against 4.44%. Routing randomly at the tight clock up to a ratio of 4 gives 4.01% against the ladder's 4.21%. It is better on 12 designs and worse on 11, with a 95% interval of -1.25 to +0.78 points. Raising theta at both clocks to 5 or 6 makes things worse. On milestone 138's own cells the same guard looks good (2.16% against 2.45%). That is the in-sample flattery the test was there to catch.
| ratio band | cells | failures | ladder regret | random regret |
|---|---|---|---|---|
| 3 to 4 | 45 | 16 | 6.20% | 6.03% |
| 4 to 6 | 52 | 9 | 2.65% | 4.44% |
What it means for the paper
The ladder's advantage grows with how dear routing is. Within a factor of about 4 of the mapped cost, it's a high-variance coin flip that neither helps nor hurts on average, so the threshold stays at 3. The paper now says that in place of the tight-clock story, which this test refutes. Three of six predictions held. The full table is in experiments/tight_theta/results.txt.