Preregistered study: self-repair value comes from falsification, not re-exposure

arXiv·medium signal

Small frozen code models are routinely asked to fix a failed program after seeing their own failing output, treated as a retry mechanism. This internally-preregistered, placebo-controlled study finds the value of feedback comes from opening the conjecture to an external executable counterexample (a test violation), not from re-exposure to the failing code. For agent self-repair loops, it implies you should feed the failing test/oracle, not just the broken output.