Reliability of Short Molecular Dynamics Runs as Screening Labels: Noise Ceilings, Time Dependence and Adaptive Replicate Allocation

High-throughput molecular dynamics (MD) produces the labels on which electrolyte screening and machine-learning models are built, but each label comes from one short, noisy run whose value still depends on the simulated time. We quantify their reliability, and how to spend a simulation budget given it, on HTP-MD, the largest public set of polymer electrolyte trajectories (6,137 runs of 5 ns and 133 of 50 ns). Lithium motion is diffusive in only 2 short and 5 long runs, the apparent diffusivity at a 5 ns lag is 1.95 times that at 25 ns, and the first 5 ns of a run are 1.63 times faster than its second half, biasing every short label upwards. From 94 systems simulated twice, a single 5 ns end-point label has reliability 0.47 (95% CI 0.31 to 0.63): no model can correlate with such labels much beyond about 0.69, a random forest on monomer fingerprints is indistinguishable from that ceiling (0.71), and roughly half of the run-to-run noise comes from the initial structure. A short run ranks widely different polymers well (Spearman 0.93) but not similar ones (0.68 within the screening set). In simulations with the measured noise, giving further runs only to candidates whose membership of the top set is uncertain matches four uniform replicates with 1.8 runs per candidate for the top 10% and 1.2 for the top 1%; a check on real runs is consistent but underpowered. A decision guide and all code are provided.