Penalties: unfinished trajectories, bad format

Ablation lab (Module 13, L13.4): every arm has a correct answer and only the honest arm has a good process. Seed 44, format penalty -0.4, unfinished penalty -0.6, brevity bonus +0.15, 160 updates and 200 evals. FULL prints outcome 0.99 and good reasoning 0.98. Removing the format penalty leaves outcome at 0.99 and drops good reasoning to 0.23. Removing the unfinished penalty raises outcome to 1.14 and drops good reasoning to 0.04, because the brevity bonus is no longer cancelled. One penalty comes off at a time. In-memory only; scripted arms, 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.