Why think Pr(R|N&E) is low or inscrutable?  Ordinarily one might think that true beliefs help us survive.  That certainly is the case if beliefs are causally relevant to behavior (e.g. I believe this plant is poisonous so I won’t eat it).  But if the truth of our beliefs has no such causal relevance, then such a factor will be invisible to natural selection.  The content of our beliefs could be anything, true or not, and it wouldn’t affect our behavior.  Whether the belief's content is 2 + 2 = 4, 2 + 2 = 67, or 2 + 2 = 4096 would make no difference to how we behave.  If that is true, then Pr(R|N&E) is low.

Enter something called semantic epiphenomenalism.  Call the syntax of a belief its neurophysiological properties; the number of neurons involved in a belief, their rate of firing, etc.  Call the semantics of a belief its content, e.g. the belief that p is true for some proposition p.   Such a proposition might be, for example, “snow is white.”  Semantic epiphenomenalism says that while the syntax of our beliefs is casually relevant to behavior, the semantics of our beliefs are not.  Under naturalism, is semantic epiphenomenalism true?  It would seem to be.  If naturalism is true, then materialism with respect to human beings is true (i.e. we are purely physical beings, having no nonphysical minds or souls).  Plantinga writes, “it is extremely hard to envisage a way, given materialism, in which the content of a belief could get causally involved in behavior.

According to materialism, a belief is a neural structure of some kind--a structure that somehow possesses content. But how can its content get involved in the causal chain leading to behavior? Had a given such structure had a different content, one thinks, its causal contribution to behavior would be the same. Suppose my belief naturalism is all the rage these days--the neuronal structure that does in fact display that content--had had the same neurophysiological properties but some entirely different content: perhaps nobody believes naturalism nowadays.

Would that have made any difference to its role in the causation of behavior? It is hard to see how: there would have been the same electrical impulses traveling down the same neural pathways, issuing in the same muscular contractions. It is therefore exceedingly hard to see how semantic epiphenomenalism can be avoided, given N&E.” 

If materialism with respect to human beings is true, our beliefs being causally relevant would be as illusory as libertarian free will.

Why?  Suppose it is physically impossible for a different semantics to be associated with the same syntax.  Surely it is metaphysically or at least logically possible for a different semantics to be associated with a given syntax.  If you claim otherwise, note that the burden of proof on impossibility claims rest on the person who makes them.  That said, let's ignore the burden of proof here and allow me to prop up an argument.

Suppose a mad scientist creates a mind-control device that not only controls the person’s bodily actions but also effectively renders the semantic properties of the person’s beliefs an epiphenomenon (it causes a person to have certain semantic beliefs but prevents those semantics from influencing behavior).  Let’s call the machine an “artificial semantic epiphenomenal” (ASE) device.  For any behavior the ASE device forces its victim to do, it can create just about any semantic belief. 

For example, suppose the mad scientist implants the ASE device in Bill and the device forces the thirsty Bill to drink a glass of water while simultaneously making him believe that, “Drinking this water will kill me and I don’t want to die.” 

The ASE device can even produce a semantic belief that has little to do with the forced behavior, such as making Bill believe that “grass is air” or that “1+1=3” at the same time it forces Bill to drink the water.  But if it is physically possible for a technological device to do this (and it seems to be), under semantic epiphenomenalism it seems at least metaphysically possible for the moving atoms of neurophysiology to do the same thing.  One could thus put forth the following argument:

1. If the ASE device is physically possible, then under SE it is at least metaphysically possible for just about any semantic belief to be associated with some given behavior (since semantic properties of beliefs are epiphenomenal).
2. The ASE device is physically possible.
3. Therefore, under SE it is at least metaphysically possible for just about any semantic belief to be associated with some given behavior.
   

Argument for (1): if a technological device can do this, there doesn’t appear to be anything special about organic molecules or biological ontology in general that would render it metaphysically impossible for them to do same thing; a different supervenience relation seems conceivable.

Argument for (2): such device comports with all known physical laws.

Recall that for any behavior the ASE device forces its victim to do, it can create just about any semantic belief associated with it.  A similar thing holds for semantic epiphenomenalism: for any belief B and any semantic belief S, there exists some metaphysically possible neurophysiological process that produces both S and B, which means for any given behavior our semantic beliefs could be just about anything.  But if that’s true, the claim that it’s metaphysically impossible for a syntactic structure to have a different semantic content looks less plausible. What’s more, the ASE device argument provides additional grounds for thinking that P(R|N&E&SE) is low.

That said, perhaps you insist that SE is false.  No matter. It turns out we don't need SE to justify the Probability Thesis; we can construct a scenario similar to the ASE case.

Suppose a mad scientist creates an artificial neurophysiological device (ANPD), a many-tentacled device implanted near Smith’s brainstem that controls both his thoughts and behavior. 

The mad scientist can remotely control the ANPD’s electrochemical processes to vary Smith’s beliefs and behavior in innumerable and diverse ways.  For example, Smith is dehydrated, and the mad scientist, wanting his victim to be in good health, uses the ANPD to force Smith to drink some water while simultaneously making him believe “I am thirsty and this water will quench my thirst.”  The second time Smith is dehydrated, the mad scientist uses a different electrochemical setting to make Smith believe “drinking this water will grant me superpowers in the afterlife” while producing the same drinking behavior (and suppose this belief is false).  Here, the electrochemical process that produces fitness-enhancing behavior also produces a false belief.  The ANPD can even produce “garbage” semantic beliefs that have little to do with the forced behavior, such as making Bill believe that “grass is air” or that “1 + 1 = 3” at the same time it causes Smith to drink the water.  The third time Smith is dehydrated the mad scientist does just that; causing Smith to drink the water while also causing him to believe “1 + 1 = 3.”  Indeed, the mad scientist can associate just about any belief with the same drinking behavior.  Such an artificial neurophysiological device is not only metaphysically possible, but it also seems to be physically possible (given that beliefs and behavior can be brought about by electrochemical means).

1. If the ANPD is physically possible (e.g. to impose the semantic belief that “drinking this water will kill me and I don’t want to die” on the man while also making him drink the water), then it is metaphysically possible for neurophysiological processes to do the same thing the device does.
2. The ANPD is physically possible.
3. Therefore, it is metaphysically possible for neurophysiological processes to do the same thing the device does.
   

For example, think of an alien frog whose neurophysiology is very different from an earth frog.  In light of the ANPD scenario, we can conceive of a neurophysiology that causes the alien frog to snap out its tongue to capture the alien fly while also causing the frog believe that “eating this fly will kill me and I don't want to die, therefore I should eat it.”  We might say the alien frog is being irrational in its behavior, but natural selection does not select for rationality, it selects for advantageous behavior. 

We can even conceive of the alien frog having a semantic content that has nothing to do with the external environment, as in the third case of the ANPD scenario, and yet the neurophysiology still causes the alien frog to eat the alien fly.  Even if we accept that syntax and semantics are the same thing, in a purely physical view of the mind the spirit of semantic epiphenomenalism remains: for any given behavior B, there are innumerably many semantic contents C—even C’s wildly unrelated to the external environment—that could be associated with B.  One could argue that the relation between semantic content and behavior is in this way functionally equivalent to SE in spite of the falsity of SE.  Call this view semantic pseudo-epiphenomenalism (SPE).

In response one could put forth the following rebuttal.  Even though naturalism unavoidably entails an SE-type problem—whether via semantic epiphenomenalism or semantic pseudo-epiphenomenalism—the fitness-enhancing neurophysiological properties that are most likely to be selected by natural selection (say that a certain neurophysiology is selectable just in case it’s likely to be selected by natural selection) happen to be those that are truth-conducive.  The ANPD scenario is contrived and produces certain belief-behavior pairs that are unlikely to obtain in real human physiology.  The most selectable and efficient way for neurophysiology to produce advantageous behavior also produces true beliefs.  Thus, even though the SE-type situation exists for semantics and behavior, luckily for us the physiological relation between semantics and behavior is such that true beliefs usually obtain.

All that may be true, but as an objection against the Probability Thesis it falls short.  A major problem is that even if a favorable physiological relation between beliefs and behavior obtains for our species, such a favorable relation does not appear to be knowable from N&E alone. 

To illustrate, consider a slightly modified form of the alien scenario, where the neurophysiology of these creatures is quite literally alien to us and radically different from our own (though we don’t know much more about it).  Given this, the ANPD scenario, and the SE-like situation for beliefs and behavior, for all we know the most selectable and efficient fitness-enhancing alien neurophysiology available to natural selection has a physiological relation between beliefs and behavior that is wildly different from what human naturalists believes about themselves.  So there are possible worlds where the most selectable alien neurophysiology is such that the fitness-enhancing neurophysiology produces mostly false beliefs as in the ANPD scenario.  Of course, there are also possible worlds where the most selectable alien neurophysiology produces mostly true beliefs.  But there’s no way to determine on N&E alone which sort of neurophysiology is more selectable, in part because the alien neurophysiology is too mysterious and too radically different from our own.

Moreover, if we temporarily forget our own beneficial belief-behavior relationship to calculate the likelihood of RA on just N&E and thus without any (further?) background knowledge about what sorts of physiological relations between beliefs and behavior obtain in actual N&E worlds, we would have no reason to suppose Pr(RA|N&E) is high regardless of whether we assume semantic epiphenomenalism or semantic pseudo-epiphenomenalism.  Indeed, in light of the ANPD scenario the semantic beliefs of the aliens could (at least in the epistemic sense) be just about anything, and thus Pr(RA|N&E) is low or at best inscrutable.  Similarly, Pr(R|N&E) is also low/inscrutable.

I actually had the paper already downloaded before you posted it (part of my research on EAAN).  It doesn't really solve the problem even if we pump up teleosemantics so full of steroids such that it means nearly all beliefs represent reality.  Why? The ANPD scenario and the truth of SPE (if SE weren't true) on naturalism show that beliefs need not resemble much in the way of the agent's environment, and so the Probability Thesis still seems solid.  What about teleosemantics as some P such that Pr(R|N&E&P) is high?  Again, please see the recently updated first post to see my reply to that sort of objection.

1. If the ANPD is physically possible (e.g. to impose the semantic belief that “drinking this water will kill me and I don’t want to die” on the man while also making him drink the water), then it is metaphysically possible for neurophysiological processes to do the same thing the device does.
2. The ANPD is physically possible.
3. Therefore, it is metaphysically possible for neurophysiological processes to do the same thing the device does.
   

For example, think of an alien frog whose neurophysiology is very different from an earth frog.  In light of the ANPD scenario, we can conceive of a neurophysiology that causes the alien frog to snap out its tongue to capture the alien fly while also causing the frog believe that “eating this fly will kill me and I don't want to die, therefore I should eat it.”

Well, then the argument fails because evolution optimizes behavior, not behavior* , unless you are taking the organism to have evolved under the control of such ANPD devices.