Some people object to (this extension of) Marshall's theory because they assume that it involves a reversal of the arrow of time - that events in the future have an influence on events taking place now. They feel that this would be unparsimonious. For example, they assume that when I flip a coin to decide whether to send a letter to a politician it is impossible that the flipping outcome could be influenced by resonance (eidopoic influences). They feel that the nerves and muscles in my body involved in flipping could not possibly know that I am going to post a letter in the future, or that the politician will be influenced by it.
Rather than looking at future events, I think that the right way to think about this is to consider the situation immediately after the event. Let us say that there are two possible outcomes, and therefore two resulting situations. The event might be something infinitesimally small - a quantum fluctuation at the wall of one of my nerve cells for example. (Or maybe the fluctuations are relatively large, involving millions of atoms. We don't know because we can't look at them - there is likely to be a strong observer effect here.) The possible outcomes are that the nerve fires or does not fire. A macroscopic event such as a coin landing with the head up can perhaps be broken down into a chain of minute events.
What is needed now is that the system is able to compare the two situations immediately after the event to see which one is likely to give rise to the greatest change, and to use the information to help to determine the which outcome is chosen.
I will try to explain below what I have in mind when I refer to "change" and a "situation" in this context.
In the diagram above the vertical axis represents time. The two horizontal axes represent the state of the universe, or a part of it such as the city of London on planet Earth. Obviously a myriad of variables are needed to describe even this small part of the universe, but since I can't show a space with more dimensions than three, these have been projected somehow onto only two axes. The arrows at the top of the diagram represent four of the myriad of variables. For example V1 might be the number of cars in London. V2 might be the number of blades of grass in the city. V3 might be the physical distance that the Chancellor of the Exchequer is from the Archbishop of Canterbury, and V4 might be the temperature in the McDonald's at Clapham Junction. The variables have all been scaled to fit onto the diagram.
As time goes forward these variables tend to change. If we imagine averaging the values of all the variables we can have a concept of a distance from the starting position when time = "now". I have shown two "contours" which represent the probability of getting to a particular point. If you start at "now", for any given time after this there is a 90% probability that you will pass through one of the points inside the inner contour, and a 99% probability of passing through one of the points inside the outer contour. The actual path followed is a random kinked line represented by the purple line.
Now, let us consider two changes, P and Q, which take us to points p and q respectively. Event P results in moving to a position inside the 90% probability line. Q is an unusual event, and it results in moving to a point outside the 99.999% probability line.
There are many routes to point p. There are fewer routes to q. Event P might be deciding to go down to the pub. Event Q might be deciding to let off an atom bomb in central London. If it helps, you can imagine that in some sinister experiment I decide to toss a coin to decide between the two.
Clearly, the situation after P is little changed. After Q, all of the points that had been probable are no longer probable, while all of the points that are now probable were previously improbable.
Marshall's theory suggests that a system is able to detect the difference between the starting situation and the situation after two possible outcomes, and to favor the outcome which gives the smaller difference. This would be achieved by comparing the resulting probabilities of points at subsequent times in the two cases with the probabilities before the event.