Agent-drafted interpretation, pending human review

The central pattern is a strong decoupling between task success and latent-strategy alignment. Final sampled-test accuracy is essentially saturated across the five completed tasks, but router-sampled and posterior-mean probe accuracy remain only moderate.

The logged objective decomposition helps explain the failure mode. The actual optimized loss in this collection is just normalized reconstruction plus 0.1 * KL prior, and that objective behaves well. The quantities that look explosive, posterior variance and router marginal KL to the prior, are only diagnostics because their weights are zero. So this collection provides a clean example of objective optimization succeeding without forcing z to become a linearly decodable strategy variable.

The most plausible read is that the pretrained backbone can already solve these tasks well enough that the easiest solution is to preserve answer quality without making z semantically clean. In that regime, latent structure can drift toward a nuisance code while the generator still produces correct outputs. This direct-pretrained diagnostic therefore motivates later interventions that either regularize the latent more aggressively or make the model rely on z more semantically.