agentscritiquebearishAutomated research systems are prone to silent failures where analysis code executes successfully yet relies on invalid causal assumptionsMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21173v1