Contrastive Pretraining on Real LHC Jets: Transfer, Scorer Dependence and Mass Distortion

We compare contrastive pretraining on real AspenOpenJets and simulated LHC Olympics jets for background-based anomaly ranking, using matched random backbones and jet observables as controls. Equal training budgets comprise 100,000 jets per domain, one real-data shard and three seeds. Only 24 real-data training jets lie in the simulated reference sample’s central 90% transverse-momentum range, preventing a causal attribution to the pretraining domain. On partitions withheld from development, transverse-momentum balancing gives mean nearest-neighbor AUCs of 0.541, 0.564 and 0.517 for real-data, simulation and random representations on two-prong signals; three-prong values are 0.419, 0.553 and 0.426. All real-data seeds rank three-prong signal below background. Random backbones lead the original two-prong comparison with mean AUC 0.636. Signal–background momentum and multiplicity differences are consistent with score correlations and ranking changes, without establishing their cause. Regularized Mahalanobis and whitened nearest-neighbor scores do not make real-data representations competitive with girth; whitening benefits random features more. At background-calibrated operating points, girth retains more signal but strongly distorts background jet mass, whereas real-data-trained scores retain little signal. This bounded comparison characterizes representation, scorer and population dependence. Prior access to these public benchmarks limits independence; scorer and physics extensions are post-evaluation diagnosticsThis version supersedes the earlier preprint v1.0.0 (https://doi.org/10.5281/zenodo.20827792), which described a different analysis with different conclusions. Code and saved numerical results: https://github.com/Animesh-Parashar/aspen-jet-anomaly (archived at https://doi.org/10.5281/zenodo.23140427).