So given that the future estimation could be trained on data from a delayed assumption state from the past prediction of the present, then what is missing? The missing seems to be based on the time factorization process NP problem and innovation stimulus which would cover things that are unknown within the net as well as time relevance which was not compensated for (the delay has an opportunity to sample lesser pasts for greater present prediction but produces nearer futures without doing Monte Carlo assumptions for a spread).
A subnet could be trained to do the estimations of the best assumption for such a predictive engine, leading to a trainability for an expected spread entropy (a situational requirement of MUST and or ANY as GOOD) given a similarity measure of an output of training to a random network spread RND classifier. https://www.youtube.com/watch?v=z4lAlVRwbrc is an interview with an author on an interesting paper about AI exploration. This covers the RND idea in a use case. Training a post RND latent space map to merge lingual or other equivalent factorizations of the novelty could be part of this.
The reevaluation of situational state novelty then can become a post addition of a trained residual based on the expected future estimation and the purpose to which the predicted estimator is to be put. Imagine on a stage pretending or on a real battlefield. The eventual motor actions of production to have for benefit?