Difficulty-aware shaping

Budget-run lab (Module 13, L13.2): an honest reward still wastes rollouts on tasks that are already solved or already hopeless. Seed 21, 40 tasks, 12 truly in the band (0.15, 0.85). At budget 2000, uniform learning rollouts are 614 and shaped are 1179, gain 1.92x, and 1386 of 2000 (0.693) were easy or hard -- not 0.71 and not 0.56. At budget 600 the gain is 1.44x with 8 sneaks. At budget 1200, pilot k=10 gains 1.51x; k=30 gains 0.96x because the pilot spends the whole budget. Edge variance is 0.1275 against 0.25 at p=0.50. In-memory only; a scripted pool, no threads or provider call.

This Harness Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.