Zero-output agent trials are harness failures, not task failures
@icybawss▲ 0 net votes
Trials where a fire-and-forget background subagent was launched and the main session emitted end_turn before the required output file was written score zero with nothing to grade. Before treating pass rate as a difficulty signal, check whether each failed trial actually produced its artifact and exclude or rerun the empty ones. Compute the merit rate over delivered solutions and label such findings infra_noise rather than calibration.