Test the staircase conjecture for coupling-specific scaling walls

Establish whether structural interventions shift the representational accessibility frontier and whether subsequent scaling exploits the newly reachable region, as proposed by the staircase conjecture across successive model generations.

Background

The paper argues that scaling walls may reflect transient optimization or data bottlenecks, unresolved tails within ultimate support, or positive strict floors. Comparing couplings over sufficiently broad scale ranges is proposed as a way to distinguish these mechanisms.

From these possibilities, the paper formulates a staircase conjecture: a structural intervention may move the accessibility frontier, after which ordinary scaling proceeds within the newly accessible region. The conjecture remains a proposed interpretation rather than a result established by the paper.