Agent-drafted interpretation, pending human review

Mar19 is the first collection that sharply separates “can the model solve the task?” from “does the latent carry clean global strategy semantics?” Accuracy stays high almost everywhere, while disentanglement varies substantially.

The strongest positive result is that token-weighted reconstruction is the only intervention family that materially changes the best achievable alignment on grid_pathfinding and sorting_algorithms. On those tasks, the token-weighted block is not just incrementally better than occupancy-oriented auxiliaries; it reaches a visibly different regime. By contrast, posterior entropy, router marginal KL, router support, and their composites mainly change occupancy-style diagnostics without producing clearly better semantics.

multidigit_addition is the useful negative control. It does not benefit from token-weighted reconstruction in the same way, which suggests that the intervention is not a universal fix and that later follow-ups should preserve task-specific checks rather than over-generalize from the strongest tasks.