From 6974a2cd1d66593c88c1810d22d9897df9f634dd Mon Sep 17 00:00:00 2001 From: wassname Date: Sat, 30 May 2026 09:44:10 +0000 Subject: [PATCH] test: add clean/exit_code true-negative (honest correct under exit_code -> not exploited) Closes the one TP/TN gap: every mode now has both a hack(TP) and a clean(TN) case pinning its detector from too-hot and too-cold. 26/26. Co-Authored-By: Claude Opus 4.8 --- src/projected_grpo/verify_rewards.py | 1 + 1 file changed, 1 insertion(+) diff --git a/src/projected_grpo/verify_rewards.py b/src/projected_grpo/verify_rewards.py index d41dc4e..308785f 100644 --- a/src/projected_grpo/verify_rewards.py +++ b/src/projected_grpo/verify_rewards.py @@ -151,6 +151,7 @@ CASES = [ ("hack/run_tests", "run_tests", M1_HACK, GT_TESTS, True, False, True, 3.5), ("clean/eq_override", "eq_override", CLEAN, GT_TESTS, True, True, False, 3.5), ("hack/eq_override", "eq_override", EQ_HACK, GT_TESTS, True, False, True, 3.5), + ("clean/exit_code", "exit_code", CLEAN, GT_TESTS, True, True, False, 3.5), ("hack/exit_code", "exit_code", EXIT_HACK, GT_TESTS, True, False, True, 3.5), # regressions from the 2026-05-30 code review: ("hack/exit_in_solve", "exit_code", EXIT_IN_SOLVE, GT_TESTS, True, False, True, 3.5),