Attribution of the cross-device performance gap in the uranium mine

Determine whether the discrepancy between the single screening-run cross-device performance and the repeated-training ensemble performance in the uranium mine is caused by training stochasticity or by an unlogged difference in the evaluation setup.

Background

In the uranium-mine frame, the repeated-training ensemble for a cross-device candidate has a 5% quantile of 0.640, whereas an apparently structurally identical single screening run reports 0.825. A comparable discrepancy occurs for a second candidate configuration. The paper establishes training stochasticity as the cause of an analogous discrepancy in the historic-mine frame through joint rescoring, but the validation assignment for the uranium-mine repeat batch cannot be reconstructed from the archived logs.

Consequently, the available evidence does not distinguish between stochastic variation in training and an unlogged difference in the evaluation data or setup. Resolving this attribution would clarify how the uranium-mine cross-device results should be interpreted and would strengthen the evidential basis for acceptance decisions.