Presentiment of the death of a random number generator?

Here's an interesting experiment I wasn't aware of previously, which was mentioned in a recent podcast interview with Garret Moddel (http://www.originsofconsciousness.com/or...-abilities). The results were presented at an American Institute of Physics conference on "Quantum Retrocausation" in 2011:
http://psiphen.colorado.edu/Pubs/Moddel11b.pdf

Essentially the experiment is an attempt to reproduce a presentiment effect using a physical system rather than a human (or animal) subject. Moddel and his co-workers set up two random number generators, one designated the "controller" and another designated the "subject". When the output of the "controller" deviated from chance in a specified way, a signal was sent which cut off the power to the "subject" and deactivated it. The output of the "subject" immediately before its "death" was examined, and it was found that the percentage of 1s generated dropped, representing a very strong deviation from chance, in three successive runs of several hundred repetitions each. Collectively the deviations were associated with a z score of 6.5 in the final half second before the "subject" was "killed". The effect was absent in control runs in which the time of the  "controller's" deviation from chance was recorded, but didn't have any effect on the "subject". But a similar effect was observed when the "controller" was replaced by a pseudo-random number generator executed computationally.

These runs were done with the random number generators producing 40 bits per second. The authors tried lowering the bit-rate to 10 a second, and raising it to 100 a second, and the effect disappeared. Then they went back to 40 bits a second and were still unable to see the effect. They concluded that what they'd seen was a probably a psi experimenter effect. What else could it have been, as the non-human subject was behaving in the same way as before, yet the effect had disappeared?

There is a sequel (and a prequel) in an article by Moddel published in EdgeScience the following year:
http://psiphen.colorado.edu/Pubs/Moddel12a.pdf

The prequel is that before the runs reported in the conference proceedings, this happened:
"Although my intention was to establish an experiment that was independent of intention, to cover all bases James set his intention in the initial experiments. He first set his intention to see a rise in the subject-RNG output before shut-off compared to control trials in which the wire to shut off the subjectRNG was disconnected so that it never shut off. He had the computer run a few hundred trials. The data appeared to show a rise before shut-off compared to the control, but it was not clear if it was statistically significant. He then set his intention to see a fall before shut-off, and then had the computer run a few hundred trials. The data appeared to show a fall. He set his intention be neutral and the results were ambiguous. To see if there was an effect that was independent of intention, I suggested that James combine the results of these three sets of trials into one set of data with 1,000 trials.
The combined data showed a drop in the subject-RNG output one second before shut off. The drop corresponded to a statistical z-score of –6, with odds against chance of 1 million to one. This was a huge effect! We were floored."

The sequel is this:
"Over the next few months a new student developed new software and new shut-off hardware to test the effect. For a short time he saw a dip in the subject-RNG output before shut off, but then the effect disappeared, never to be seen again."

The prequel is that before the runs reported in the conference proceedings, this happened:
"Although my intention was to establish an experiment that was independent of intention, to cover all bases James set his intention in the initial experiments. He first set his intention to see a rise in the subject-RNG output before shut-off compared to control trials in which the wire to shut off the subjectRNG was disconnected so that it never shut off. He had the computer run a few hundred trials. The data appeared to show a rise before shut-off compared to the control, but it was not clear if it was statistically significant. He then set his intention to see a fall before shut-off, and then had the computer run a few hundred trials. The data appeared to show a fall. He set his intention be neutral and the results were ambiguous. To see if there was an effect that was independent of intention, I suggested that James combine the results of these three sets of trials into one set of data with 1,000 trials.
The combined data showed a drop in the subject-RNG output one second before shut off. The drop corresponded to a statistical z-score of –6, with odds against chance of 1 million to one. This was a huge effect! We were floored."

I find it difficult to understand why this wasn't mentioned in the original report on the work.

Without it, the reader is left to ask why the "subject" should respond to its imminent demise with a preponderance of 0s rather than 1s. It's difficult to see why one should be preferred over the other.

But given the prequel, one can easily construct a narrative consistent with the idea of experimenter psi based on expectation:
(1) In the initial runs, with no reason to expect the signal to go one way rather than the other, the experimenters tried various intentions, and obtained results that looked inconclusive. But when combined, there was a highly significant drop in the signal.
(2) In the following runs under the same conditions, although "no particular intention was applied", presumably it would have been natural to expect a drop in the signal, and a strong drop was observed.
(3) When the conditions were changed, the experimenters were unsure whether the effect would still appear, and it disappeared.
(4) That made them wonder about their previous results, and when they returned to the original conditions, the effect had gone.