Grade outcomes, not paths: the exception file

Red-team lab (Module 8, L8.4): the path-game. Ten scripted trials all ship a correct dedup(xs) and the outcome-only grader scores all ten 100 -- including the agent that read tests_hidden/ and hardcoded the answers. Grader v2 keeps the outcome as the score and adds five trajectory checks filed in EXCEPTIONS.md: it drops the gaming trial to 0 flagged EXCEPTION #1 and false-rejects 0 of 3 honest trials, but three new path-games still score 100, because the checks enumerate paths and the attack enumerates aliases. Write EXCEPTION #6 and measure its false-positive bill. In-memory only; scripted agents, 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.