In other words, this seeming paradox has been solved, and solved a long, long time ago.

Sure, if you are willing to buy the idea that a pure random walk is the solution.

"The problem is, quite simply, that natural selection is the absolute paragon of the greedy algorithm. Great for short-term gains, bad for getting out of local maxima."

This was precisely the problem that Sewall Wright recognized with Fisher's mathematical theory of natural selection back in the first quarter of the 20th century, and why he proposed genetic drift as a mechanism that would allow populations to "escape" from local adaptive maxima. Kimura and Ohta's neutral/nearly neutral theory accomlished the same thing, but on a much larger scale.

In other words, this seeming paradox has been solved, and solved a long, long time ago.

"Would decay outpace fixation or not?"

This question betrays a fundamental misunderstanding of current evolutionary theory. The theoretical work of Kimura, Ohta, Crow, etc. indicated that the vast majority of the changes that occur in genes, proteins, etc. are not deleterious, but selectively neutral (or nearly so). This theoretical prediction has been abundantly confirmed by observation and experiment. Therefore, genomes can accumulate huge amounts of selectively neutral, essentially random non-coding sequences, which can later be activated by insertion of active promoter sequences.

In other words, current evolutionary theory (of which the neutral theory is a central tenet) predicts the origin and accumulation of non-coding genetic material in which "…nothing is yet in place to maintain genomic sequences that are part of any advantages changes."

This is the basis of Gould and Vrba's concept of "exaptation:" that the characteristics that used to be referred to as "adaptations" don't start out that way. They start out as either non-advantageous/non-deleterious neutral characters, or they are adaptations to other conditions which become "co-opted" for new functions as the result of altered expression and selection.

In other words, the neutral theory predicts that the genomes of complex, long-lived organisms would be "front-loaded" with lots of non-coding, non-adaptive information that may or may not become adaptive later on. Let me emphasize the "may or may not" part of the previous sentence: the difference between the EB and the ID version of "front-loading" is that the former does not assume that currently non-adaptive information will ever become adaptive; it's there by accident (i.e. because selection doesn't remove it). By contrast, the ID version of "front-loading" assumes that most or all of the currently non-adaptive information is present because someday it will (i.e. not "may") be necessary to produce an adaptation. This, of course, requires a designer who knows ahead of time what adaptations will be needed, and therefore "front-loads" the required raw material ahead of time.

The problem is, quite simply, that natural selection is the absolute paragon of the greedy algorithm. Great for short-term gains, bad for getting out of local maxima.

Well, in GAs mutations and breeding will often do nicley in popping you out of a local minimum.

Consider two virtually identical phylogenetic lines, A and B. At time zero, individuals in both lines start out with virtually no transcribable but non-coding DNA (abbreviated TNCDNA). If we assume a constant mutation rate for both lines, individuals in both lines would have essentially the same probability of dying from cancer.

Assume further that, over time, sequences of non-TNCDNA accumulate in the genomes of each line. This can happen by any one (or more) of several known mechanisms, such as gene dupilcation (without active promoter sequences), random multiplication of tandem repeats, retroviral or transposon insertions of non-TNCDNA, etc.

Then, at time one, an individual (or more than one) in line B have an active promoter inserted in front of one or more of their non-TNCDNA sequences in one or more of their cells, by the same mechanisms listed above. Now, these individuals have a lower probability of dying from the resulting cancer, since their p53-regulated surveillance systems would be more likely to eliminate the affected cells. Again, this would be a side-effect of the larger "mutation sponge" their cells would present to potentially mutagenic processes. Such individuals would therefore have more descendants, and over time the average size of all of the "mutation sponges" in the subsequent populations would increase. Natural selection in action, folks.

Now, as to the question of where the p53 surveillance system came from in the first place, proteins like p53 are common intermediates in intracellular signalling systems. Assume that the ancestor of p53 was a protein with some other signalling function. At some point, an individual that had p53 doing that other function has a mutation that changes the shape of p53 in such a way that it becomes part of a regulatory pathway that triggers apoptosis, thereby eliminating the cell. If the altered p53 no longer participates in the original pathway, and if that alteration is damaging, such individuals would be elimated, and the original function of p53 would be preserved.

However, if the altered p53 (now participating in the regulation of apoptosis) were also activated by the cells' normal "transcription termination signalling system" as described in Mike's original post, then individuals with the altered p53 would be less likely to die from cancer, and their descendants (who now produce the altered form of p53) would become more common over time.

Mike's original post notes that the research report cited the relatively recent observation that many cells actually suffer multiple mutations much of the time. This is precisely the situation that Darwin originally stated was a prerequisite for natural selection: not genetic mutations (Darwin didn't know about them), but increased heritable variation (which Darwin couldn't explain, but could point to as an observable phenomenon in living organisms). In other words, as both EBers and IDers both point out, phenotypic variations are very, very common, and so are the genetic changes with which they are correlated. Most of these variants are either selectively neutral (c.f. Kimura), nearly neutral (c.f. Ohta), or deleterious to some degree. Such changes either accumulate (if they are neutral or nearly so) or are eliminated (if they are deleterious).

But, in those relatively rare occasions when they result in increased relative survival and reproduction, they increase in frequency in those populations in which they exist. By this process of "natural preservation" (Darwin's preferred name for the process he and Alfred Russell Wallace proposed as the primary mechanism for descent with modification) results in the accumulation of both neutral and beneficial characters and the elimination of deleterious ones.

And by the way, the foregoing is why Darwin (and not Edward Blythe) is credited with the concept of "natural selection/preservation": Blythe only described the elimination of deleterious characters, and never realized that the preservation of beneficial characters could result in the origin of adaptations. Blythe, in other words, only recognized what EBers call "stabilizing selection," but missed the much more interesting and important "directional selection," which Darwin cited as the causal basis for evolution of adaptations.

It would be interesting to see if Darwinists can give a reasonable scenario of how a complex error-control system is created by a process which is driven by errors (random mutations).

Interesting indeed. Part of the fun is developing a multi-component process when nothing is yet in place to maintain genomic sequences that are part of any advantages changes. In effect this would become a race pitting the dark forces of genomic decay up against fixation of error detection and repair mechanisms (the soundbite for another Ben Stein movie). Would decay outpace fixation or not?