A team gave frontier agents six days, $3,000 in API credits, GPUs, a VM and the central questions from two unpublished NeurIPS submissions. The agents wrote code and ran experiments, but the original paper authors judged that neither solved the underlying research problem. The failures were recognizable: weak judgment about what would count as an answer, uncreative reaction to negative results, poor backtracking, resource blindness and gradual drift. The value of this “shadow evaluation” paper is its uncontaminated setup and released logs, not a sweeping claim from two cases. It separates automation of research labor from selection of the right experiment—still the expensive human part.