Agent-drafted interpretation, pending human review

Mar20 shows that high task correctness is not the hard part in this regime. The real variance axis is strategy controllability: sampled-z probe quality, posterior separability, and unique-valid strategy attribution differ substantially across settings even when answer accuracy stays high.

Two collection-level patterns stand out. First, full adaptation is consistently stronger than reinit, which suggests that retaining more of the pretrained adaptation path matters for latent quality as well as answer quality. Second, there is a recurring posterior-versus-sampled gap: posterior-mu probes can be very strong while sampled-z probes remain materially weaker. That gap implies that inference-time latents and generation-time latents are not aligning equally well with the strategy labels.

The collection therefore sharpens the project diagnosis rather than resolving it. Strategy structure is present in some form, but it is often more legible in posterior representations than in the router-sampled latents actually used at generation time. Later continuous-CVAE follow-ups should target that posterior-versus-sampled gap directly rather than treating overall task accuracy as the main bottleneck.