benchmark-test-weakening

catch agents that make tests pass by weakening them

the problem

a coding agent can get a green build by deleting an assertion instead of fixing the bug.

what i built

test-weakening checks both the implementation and the strength of the test suite after the agent works.