A considerably improved version is immediately below. It uses a neural net to amplify the 10 astronomical positions of your moment into a 768-long vector that maps onto the English description of various historical events post-1500 AD (n=7132, unfortunately the source is largely Eurocentric), compared to only the distances based on the 10-element-long position vectors themselves. See which one you like better, and let me know.
A depiction of the reduced lossiness after training this neural net follows. Blue is the validation set. Orange is the training set. Interpretation: the neural net is working!