Explanation of the unexplained K=8 composition residuals

Determine why all five K=8 language-model ecosystem arms exhibit positive residuals relative to the fitted relationship between pool travel propensity and five-generation collapse drift.

Background

The paper proposes a travel-propensity index based on how readily the contributing models are carried along by recursive training. Across nineteen experimental arms, this index explains substantial variation in five-generation drift, but more than thirty percent of the variance remains unexplained.

The unexplained residuals are systematically positive for all five K=8 arms, indicating an ecosystem-size-specific deviation rather than ordinary seed noise. The authors explicitly state that they do not know the cause of this pattern, leaving its mechanism unresolved.