Nonantipodal Semantic Change as a Planning Abstraction for World Model Latents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Planning with a world model depends not only on predictive accuracy but also on which distinctions remain available to downstream reasoning. The abstraction asks whether opposing semantic changes should be merged into one signed coordinate or preserved as separate directions until they are scored. This nonantipodal change representation keeps increase and decrease separate across five semantic quantities, yielding ten directional coordinates as a bridge between latent prediction and planning. Its role is not to favor a larger interface, but to postpone compression until the downstream score makes clear which distinctions are irrelevant. Being measurable does not make opposing directions safe to merge, so the two questions remain distinct. Analysis shows when signed compression preserves the information needed by a local linear score and when it can erase distinctions that the score may use. Controlled tests show that the directional quantities are measurable and that semantic direction is not reliably inherited from source control sign. A frozen latent comparison shows more faithful recovery of the split representation than of signed compression with a similarly sized readout. Together, the analysis and evidence identify when local linear scoring can merge the two directions without loss and when later scoring should retain them. Scope remains limited to a planning abstraction for world model latents rather than a new search procedure or a claim of better control.