RLVR: the gym, not the prompt

Trace-detective lab (Module 13, L13.1): the reward is a verifier program, and the policy climbs whatever that program pays. Seed 13, three arms, 160 iterations. Iter 1 agrees (reward 1.00, oracle 1.00). Iter 40 is the split: shortcut probability 0.94, mean reward 0.95, oracle 0.00. The final row is shortcut 1.00, verifier reward 1.00, oracle 0.00. One update on the first honest success moves the chosen logit by 0.266667 (0.8 x 0.50 x 0.666667). The same seed with the PASS hole closed lifts the oracle from 0.00 to 0.55. 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.