=== TIMING ANALYSIS === Per-mode timing — pendulum, N=20 plants, K=8, steer_steps=150, oracle=16×100 (gen on CUDA, LMI/steer/oracle on CPU) stage costs: raw generation 2.59 ms / controller ( 20.7 ms / plant for K=8) certification (LMI) 5.85 ms / controller steering (warm, 150 steps) 1101.4 ms / plant / objective cold per-plant oracle 15684.9 ms / plant PIPELINE (per plant) what you get time vs oracle 1. raw generation (K) K candidate controllers 20.7 ms 757× 2. + certification a certified controller 26.6 ms 591× 3. + steering (1 objective) optimal for a chosen objective 1127.9 ms 13.9× 4. cold per-plant oracle the per-plant optimum 15684.9 ms 1× Reading: unguided synthesis (modes 1–2) is ~591× faster than a cold per-plant solve; the STEERED capabilities (mode 3) cost more but are still ~14× faster than cold-solving the objective from scratch, because the diffusion sample is a warm start (§4.4: ~7× fewer steps) and one model is reused across the whole plant family.