Some learning transfers; large feature gains can yield small image gains

The question

Distinguish improvement in the learned feature objective from useful decoded predictions, on training examples and separate development videos.

Why this experiment?

Missing-pixel prediction did not establish coherent construction. Video endpoints provide another concrete question: can a mechanism construct an intermediate scene while using the shared VSA interface?

How it works

Perform no new fitting. Run the frozen models on all 128 training cases and 32 cases from separate development videos, using endpoint-only inputs. Compare each learned output with its matched untrained blend in feature space and decoded images.

What happened

Original-feature transport lowers training image error 6.09% but raises development error 5.14%. Recovered synthesis lowers training feature error 32.06%, mostly in the coarse overview, while image error improves about 1% on both training and development. The audit passed all 960 comparisons. Fixed-panel visual review finds recognizable scenes and persistent blur or overlapping moving contours, not consistently coherent middle frames.

Limits and next question

Training examples were previously used for learning. Development videos are separate from constructor training but are existing diagnostic data, not a fresh test set. Original encoder features are privileged inputs. Feature error is not image quality or coherence; low error on dark or static regions is not temporal understanding. All outputs pass through VSA before the frozen renderer.

Next question. Compare a learning objective tied to useful output with the existing feature-only objective on the same broader data and frozen output path, using matched learned arms and unexamined source videos. No further fit begins automatically.