Path-conditioned Jacobian lenses for Mixture-of-Experts models.

The Jacobian lens reads what a dense model is disposed to say at each layer. Every frontier model shipped today is Mixture-of-Experts, and the lens’s fitting procedure breaks on them by construction — it averages across forward passes that routed through entirely different experts.

MoE doesn’t delay workspace formation: readability arrives at the same ~85% depth as dense models, once you’re past onset. Below onset, we tried the obvious repair — cluster by routing path, fit one lens per cluster — and it lost to a random-conditioned control. That negative result is the paper’s actual contribution: it says the small gain everyone would have shipped was overfitting, not signal.