Prediction, Action Ranking, and Closed-Loop Control: A Replay-Verified Audit of Visual World Models in PushT

With the same forecaster and per-decision budget, a CEM objective change raises success from 11/50 to 33/50; a task-specific objective effect, not a new dynamics mechanism or generalization claim.

Later visual-control confirmation draft, distinct from the early synthetic paper; external publication status was not verified from these local files.

Why we tried it

Lower prediction error need not preserve consequential same-state action differences.

What the evidence shows

In 50-scene confirmation, changing the CEM objective with a shared forecaster and per-decision budget raised success from 11/50 to 33/50; a separate reranking mean peak-coverage interval crossed zero.

Where the work stands

This is a task-specific objective effect; development mechanism gates failed, upstream scene exposure is unknown, and new dynamics or unseen-domain superiority is not established.

Possible next steps

Freeze complete state sets and prospectively separate teacher/self-conditioned drift, action ranking and executed payoff.

Suggested next steps are untested research questions, not new experimental results. The archive records manuscript provenance, licensing and status.