Early warning of failure would let costly episodes of large language model (LLM) agents be stopped early. Across five model families and five task domains (3421 usable rollouts), we study a black-box signal: the divergence of a rollout’s judged goal progress from sibling rollouts of the same task after k steps. Under our original judging protocol, cohort divergence predicted failure (pooled AUROC 0.73 in retail). A judge swap revealed that the judge had received a task identifier in place of the task objective. After re-judging with the correct objective, the cohort signal fell to chance in retail and airline (0.47 to 0.52), and a single-rollout judge score had the higher point estimate in eight of ten combinations of domain and horizon. In paired contrasts, before adjustment for multiple comparisons, cohort divergence was significantly worse in two and better in one. Of three pre-specified failure-onset profiles, one met its criterion under the original protocol and none under the corrected one. Task-preserving perturbations produced no detectable amplification of divergence. In an abort simulation, under the stated cost assumptions, the cohort monitor lost tokens for every model, and the judge monitor saved tokens for two of five and had no usable operating point for the other three. Judge instructions are thus a first-order threat to validity in agent-monitoring research.