Identification of the length-sparsity confound

Determine which component of the length–sparsity complex—document length, slot count, filler update density, or related covariates—causes the observed direction of the mechanistic readout.

Background

Slot count is correlated with redundancy in the main grid and cannot be varied independently of all other relevant properties using the basic generator parameters. A shortened-document arm matches slot count while changing document length and copy dispersion, whereas a fixed-length pupdate arm matches slot count through filler update density but leaves more of the copy-dispersion difference intact.

Because the two control arms alter different accompanying quantities and yield different patterns under different pooling statistics, the experiments establish only that realized redundancy does not explain the effect. They do not isolate the causal contribution of the remaining length–sparsity variables.