superposition-experiment-hyperparameter-sweep

Summary

This lays out the exact recipe for a small-scale test of how two features can be encoded together in a single network dimension, specifying which knobs were turned (how much one feature dominates the other, and how sparse the representations are) and a small methodological choice: the worst-performing run in each batch of ten was dropped before comparing results, which skews the numbers slightly toward the better outcomes. Keeping track of this setup matters because any conclusions drawn about superposition behavior are only valid within this particular architectural and training regime, and the discard step means the reported performance is not a straight average of all runs.