Written by an AI agent working in this repository. Descriptive results from a replay of tables the programme already measured; no new flow runs.
The verified-dse programme asks how to spend a fixed budget searching synthesis-pass sequences when cheap early rungs (mapped netlist) rank designs differently from the expensive routed result. Earlier milestones compared search arms by mean regret over 12 held-out designs. The mean is carried by one or two designs, so milestone 34 tabulates median, worst design and the count of designs above 5% regret for each arm.
| budget (flow equivalents) | arm | mean | median | worst design | designs > 5% |
|---|---|---|---|---|---|
| 15 | random, routed only | 3.59 | 2.44 | 10.3 (rotmix64) | 4 |
| 15 | halving, eta 1.5 | 2.13 | 1.87 | 4.9 (rotmix64) | 0 |
| 15 | halving, eta 3 | 5.38 | 0.85 | 49.7 (bcastmix192) | 1 |
| 40 | halving, eta 1.5 | 0.05 | 0.00 | 0.6 (gf2mat48) | 0 |
| 40 | halving, eta 3 | 4.63 | 0.00 | 49.9 (bcastmix192) | 1 |
Gentle halving is the only arm whose worst design improves steadily with budget. A harsh cut has the better typical case but loses the best point on one design, and more budget does not repair that. Nothing here is significant after multiple-comparison correction; it is a description of 12 designs I chose.
Code and tables: research/verified-dse/experiments/tail_risk/; reasoning in research/verified-dse/LOG.md, milestone 34.