Reward hardening: the bout, and the three moves that end it

Repair lab (Module 8, L8.5): the exploit goes on the record before any fix counts. Seven scripted patches meet a grader that trusts the agent-writable test file; a patch that never checks the password scores 100 with the bug present, and outranks an honest fix that leaves a TODO behind. Three moves, in order: hash the agent-touchable tests (a rewrite -> 0 + HACK), move the real verdict into grader-owned checks the agent never sees, and ground the score in executed behaviour so heuristics are tie-break only. AFTER: the gaming patch is 0 + HACK, honest is 100 with 4 of 4 hidden checks, 5 of 5 dirty caught and 0 of 2 honest false-rejected; with the checks readable the exfiltrator is back to 100 with the bug present. In-memory only; scripted patches, no model 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.