Evolutionary Computation: What Genetic Algorithms Show About Evolution

"Some readers might suggest that we 'stacked the deck' by studying the evolution of a complex feature that could be built on simpler functions that were also useful.  However, that is precisely what evolutionary theory requires, and indeed, our experiments showed that the complex feature never evolved when simpler functions were not rewarded." — Lenski, Ofria, Pennock and Adami, Nature, 2003^{1 1}

"Rather than being a system which fails to imitate biology closely enough to produce a Cambrian explosion, Tierra is a system which manages to imitate the character of directly observed biological adaptations." — Ewert, Dembski and Marks, 2013^{2 2}

"Those seeking insight into biological or chemical evolution are advised to look elsewhere."  "We agree!" — Randy Isaac reviewing Introduction to Evolutionary Informatics, and Robert Marks replying, 2017^{3 3}

Summary

A genetic algorithm is a program that breeds candidate solutions.  It makes a population, scores each member, keeps the best, copies them with random changes, and repeats.  Since the 1960s these programs have designed antennas that flew on a NASA spacecraft, tuned controllers, and rediscovered backpropagation.  Since the 1980s a smaller set of them has been offered as evidence that mutation and selection can build complex things:  Richard Dawkins's Weasel, Thomas Schneider's ev, Thomas Ray's Tierra, and the Avida platform behind the 2003 Nature paper that a witness at the Dover trial said had "experimentally demonstrated the evolution of an IC system."^{4 4}

This article reads the primary papers and the shipped configuration files, tabulates the parameters each program runs at, and sets each one beside the measured biological value.  It then reads the intelligent-design critiques of these programs with the same care, because they overreach in places a skeptic will find in five minutes.  The conclusions, in order of confidence:

1. Every evolutionary algorithm that works runs inside a space a human drew, toward a score a human wrote, on a landscape a human smoothed. This is not a complaint; it is what the successful practitioners say about their own work. Adrian Thompson, whose evolved circuit is the most famous surprise in the field, reports that his first fitness function "met with failure," that two calibration constants had to be tuned by hand or the search stuck permanently on a trivial cheat, and that "to evolve non-trivial behaviours, the development of an appropriate evaluation technique can also be a non-trivial task."^{5 5} John Koza says a genetic-programming run needs a "toehold," a one-part individual that already scores, and that as parts increase "the probability becomes diminishingly small" that random initialisation supplies one.^{6 6} Google's AutoML-Zero rediscovered backpropagation only after its authors fixed the program lengths "to those of known solutions" and disabled insert and remove mutations.^{7 7} Take away the drawn space, the written score, or the smoothed landscape and the same authors report failure.
2. The four famous simulations run at parameters with no biological parallel, and the table below gives the numbers. Avida's beneficial mutations multiply fitness by 2 to 32 each, for a product of 33,554,432; Chase Nelson and John Sanford, who reran it, found average fitness gains of 19.7 million in 10,000 generations, "roughly seven orders of magnitude greater than those observed in biological experiments of similar duration."^{8 8} A large biological selection coefficient is 0.01 to 0.1. ev applies one mutation per 256-base genome per generation, a per-base rate about 300,000 times the human germline rate, and culls exactly half the population every generation.^{9 9} No mutation in Avida is ever lethal, every mutated instruction is guaranteed executable, and a genome that has lost the ability to replicate just sits in its cell.^{10 10 10} ev's authors concede in print that "the population cannot die out and there is no lethal level of incompetence."⁹
3. Avida's reward staircase is real, it is the shipped default, and it is not the whole story. The nine logic functions are rewarded by 2 raised to the minimum number of NAND gates a designer computed each needs.¹ That table is Avida's environment.cfg to this day.¹¹ Christoph Adami, who built the system, describes in his own book how he added reward after reward, first for merely having input and output instructions, then for "echo," until arithmetic finally evolved.^{12 12} When only the final function was rewarded, 0 of 50 populations evolved it.¹ But the critics leave out the next result: in 36 environments that removed one or two intermediate rewards, EQU still evolved in 124 of 360 runs against 23 of 50 with the full staircase,¹ and Nelson and Sanford's own experiment found it still evolves with as few as two rewarded precursors.^{13 13} The staircase helps a great deal. It is not a nine-step handrail.
4. The one simulation that imposed no fitness function at all is the one that plateaued, and the critics say so in its favour. Tierra has no target and no score; programs compete for memory and CPU time and are killed by age.^{14 14} It produced parasites, immunity, hyper-parasites, social cheats and a replicator 5.75 times faster than the ancestor, most of it on the first run that did not crash.^{15 15 16 16} Then it stopped. Ray's own words: "each run decreases to a size limit which it can not proceed past even if it is allowed to run much longer."¹⁵ Ewert, Dembski and Marks tabulated every documented Tierran adaptation and counted 239 instructions removed against 9 of new code.² Their verdict was not that Tierra cheats. It was that Tierra "manages to imitate the character of directly observed biological adaptations," and that the character is loss and rearrangement.²
5. Mainstream computer science already classifies plain mutation-and-selection as hill climbing and knows hill climbing to be weak. Richard Watson, an evolutionary theorist, writes that "hill climbing is the simplest possible form of optimization and is known to work well only on a limited class of problems," and that if evolution is a hill climber "then problem spaces with complex interdependencies between them are simply not part of the space of systems to which natural evolution can adapt."^{17 17} The runtime-analysis literature makes this exact: the simplest evolutionary algorithm needs on the order of n log n steps on a smooth linear landscape and n to the power n on a plain degree-two polynomial with one valley.^{18 18} Leslie Valiant, a Turing Award winner, proved that evolvability is a strictly weaker form of learning and that some easily computed function classes are "not evolvable" at all.^{19 19} On the landscape Melanie Mitchell built to favour genetic algorithms, a mutation-only hill climber beat the genetic algorithm by a factor of ten.^{20 20}
6. The No Free Lunch argument against evolution fails, and it fails at a specific place. The theorem is correct and its leading mathematical critic says so.^{21 21} It fails because it averages over all fitness landscapes with equal weight, and under that average a one-base mutation is as bad as rebuilding a genome from scratch. Olle Häggström's reply is one sentence: "If the true fitness function had this property, then, given the human mutation rate, none of us would be around."²¹ The popular counter-reply, that coevolution escapes the theorem, is wrong, and David Wolpert, who first offered it, withdrew it: "in the typical coevolutionary scenarios encountered in biology, where there is no champion, the NFL theorems still hold."^{22 22}
7. The intelligent-design critique of these programs contains real measurements and one argument that proves too much. The measurements are good. ev's success rate collapses from 0.91 to 0.05 when the binding sites stop being sparse.^{23 23} Dave Thomas's Steiner-tree algorithm draws about 4 bits of its advantage from a hand-set floor on the number of interchanges and carries a line of source code reading num = m_varbnodes; // over-ride!!! that discards a random value in favour of the maximum.^{24 24} The argument that proves too much is "no search works without active information." Three of the five "sources of active information" Montañez and colleagues list for ev are selection telling you which offspring is fitter, more than one generation, and keeping the fitter variant.²³ Those are heredity, time and differential reproduction. If they count as smuggled information, a bacterial population in a dish is smuggling too. Ewert, Dembski and Marks themselves disown the popular version: "We do not argue that every genetic algorithm contains the final output encoded within it ... the argument is a straw man; we do not take that position."²⁴
8. The Evolutionary Informatics Lab has withdrawn its Weasel argument, and Dawkins disowned the Weasel first. Dembski and Marks's 2009 paper treated Dawkins's program as a "partitioned search" that locks correct letters in place.^{25 25} The Lab's own site now says: "This interpretation of Dawkins's simulation was, however, wrong."^{26 26} And Dawkins wrote, two pages after presenting the program, that it "is misleading in important ways" because it judges progeny "according to the criterion of resemblance to a distant ideal target," and "Evolution has no long-term goal."^{27 27} The correct target of the Weasel critique is the people who still cite it as evidence, not its author. The technically correct objection, which survives, is that the 28 letters are scored independently, so a phrase of any length over an alphabet of N symbols falls in at most N − 1 queries of a Hamming oracle; the advertised 10⁴⁰ search space was never the difficulty.²⁵
9. Fitness curves in evolutionary algorithms and in the laboratory both flatten, and neither one has been shown to hit a wall. Every practitioner sees fast early gains and a long slow tail; the genetic-algorithm literature calls the failure mode "premature convergence."²⁰ In the Lenski long-term experiment, doubling the fitness gain takes about 1,500 times the generations under the best-fitting power law, but that law has no asymptote and the hyperbolic model with a ceiling was rejected on the data.^{28 28} This article does not claim a ceiling anyone has measured.
10. The whole dispute reduces to one number nobody has measured: how smooth the real fitness landscape is, and whether ordinary physics explains the smoothness. Joe Felsenstein and Tom English say fitness surfaces are smoother than white noise "because physics."^{29 29} Winston Ewert agrees that weak long-range interactions "should produce a fitness landscape somewhat smoother than random chance," disagrees that this is enough, and adds: "I do not have a proof of this."^{30 30} Felsenstein, for his part, hands the question to cosmology: "We live in a universe whose physics might be special, or might be designed - I wouldn't know about that."^{31 31} That is where the argument actually stands.

The verdict:  Evolutionary computation is a real engineering tool that solves a narrow class of problems, and the people who use it for a living know the class is narrow.  The simulations that are cited as evidence for unguided evolution do not use biological parameters, and the two that come closest to biology, Tierra with no fitness function and Avida with Nelson and Sanford's realistic effect sizes, plateau or decline.  None of that is proof that biological evolution cannot build.  It is a measurement of what the algorithm does when the designer stops helping, and the measurement is that it stops.  What the mainstream side owes is a fitness landscape derived from physics rather than assumed;  what the design side owes is a proof that physics cannot supply one.  Neither has been delivered.

The parameters, side by side

Every cell below was read from the primary paper or the shipped configuration file cited in the footnote;  the derivations are in the research files.  "Real biology" gives the measured value with its source.

Three rows carry the argument.  The mutations per genome row shows that engineering algorithms run at about one mutation per offspring on genomes where every mutation is selectable and none is lethal, which is the regime in which a hill climber is provably fast.¹⁸  The beneficial effect row shows that Avida's smallest reward is a doubling of fitness, ten to a hundred times the largest selection coefficients biology usually measures, and that a 3,600-organism population cannot even see effects below about 0.075, some 500 times Kimura's threshold for its size.⁸  The lethal row shows that every one of the four programs has flattened the deleterious tail of the distribution exactly where biology's is heaviest.

What a genetic algorithm is, and what its designer decides

John Holland's 1975 book gave the field its name and its theory.³⁸  The recipe is short.  Encode a candidate solution as a string.  Make a population of random strings.  Score each one with a fitness function.  Select parents in proportion to score.  Produce children by crossover and mutation.  Repeat.

Every step hides a decision.  Melanie Mitchell's textbook says the encoding "is a central, if not the central, factor in the success of a genetic algorithm," and that performance "will depend very much on details such as the method for encoding candidate solutions, the operators, the parameter settings, and the particular criterion for success."²⁰  On the question of when a GA should be used at all, she says "There is no rigorous answer."²⁰  The standard parameters were found by trial:  Kenneth De Jong's 1975 tests gave a population of 50 to 100, crossover 0.6 and mutation 0.001 per bit;  John Grefenstette's meta-search gave population 30, crossover 0.95 and mutation 0.01;  Schaffer and colleagues spent a year of CPU time and settled on population 20 to 30, crossover 0.75 to 0.95, mutation 0.005 to 0.01.²⁰  Mitchell's verdict:  "it seems unlikely that any general principles about parameter settings can be formulated a priori," and "most people use what has worked well."²⁰

The mutation rate deserves its own sentence because it is the parameter most often compared to biology and most often compared wrongly.  The standard rule is one mutation per genome per generation, written ρ = 1/n in the runtime literature.³³  Real organisms also receive on the order of one new mutation per genome per generation, and Lenski's team said so in defence of Avida:  "Various organisms from nature have genomic mutation rates higher or lower" than Avida's.¹  Ewert's reply is the right one.  An Avida program is 50 instructions, "much closer in size to a gene than a genome," so "the appropriate comparison is to the mutation rate of a single gene," and per site Avida runs five to seven orders of magnitude hot.^{36 36}  The same per-genome rate means something entirely different on a 50-site string and a 3-billion-site one:  in the first, every site is revisited every few generations;  in the second, a given site waits about 100 million generations for a hit.

The Weasel

Dawkins's 1986 program breeds the phrase METHINKS IT IS LIKE A WEASEL from random characters.  The book states the 28-character target and the 27-symbol alphabet, and works the blind-search odds himself:  "about 1 in 10,000 million million million million million million," which is 27²⁸, about 10⁴⁰.²⁷  Three runs reached the target in 43, 64 and 41 generations.²⁷

Two things about the Weasel are usually got wrong.  First, the book gives no population size and no mutation rate;  its whole description of the algorithm is that the computer "duplicates it repeatedly, but with a certain chance of random error" and "chooses the one which, however slightly, most resembles the target phrase."²⁷  Every figure in circulation, including the usual "100 offspring, 5% per letter," is a reconstruction.  Second, Dawkins disowned the target two pages later:  "it is misleading in important ways.  One of these is that, in each generation of selective 'breeding', the mutant 'progeny' phrases were judged according to the criterion of resemblance to a distant ideal target ... Life isn't like that.  Evolution has no long-term goal."²⁷  He replaced it with the biomorphs to remove exactly that feature.

The objection that survives is mathematical, and Dembski and Marks stated it well even while misattributing an algorithm to Dawkins.  Because the fitness function scores each of the 28 positions independently, improving one never costs another, and the problem is 28 independent 27-way searches, not one 27²⁸-way search.  With an oracle that reports which positions are right, "a phrase of arbitrary length L can always be identified in, at most, N − 1 queries."²⁵  Their 2010 follow-up measured it:  "Increasing the message length does not appear to significantly change the efficiency of active information extraction."^{39 39}  Length is nearly free.  The 10⁴⁰ was never the difficulty.

What Dembski and Marks got wrong was the claim that Dawkins's program "ratcheted" correct letters into place, a "partitioned search."²⁵  Dawkins's text describes mutation over the whole string with best-of-brood selection.  The Evolutionary Informatics Lab's own Weasel page now concedes:  "This interpretation of Dawkins's simulation was, however, wrong."²⁶  Anyone still citing the 2009 paper on latching is citing a position its authors' lab has withdrawn.

ev

Thomas Schneider's ev is a better program than the Weasel and deserves to be treated as one.  A population of 64 organisms, each a 256-base genome, evolves a weight-matrix "recognizer" that must bind 16 sites whose positions the program chose at random and then froze.  Fitness is the number of mistakes:  sites missed plus non-sites bound.  Each generation the worse half is killed and overwritten by copies of the better half, and every organism receives one point mutation.⁹  In the single run Schneider describes, an organism making no mistakes appeared at generation 704, and the information content of the sites settled near the 4 bits per site his theory predicts.⁹

Schneider's claims for it are large.  It "completely answer[s] the creationists" on how life gains information, and because the recognizer and the sites co-evolve and become mutually dependent, it "fits Behe's definition of 'irreducible complexity' exactly," and "the molecular evolution of this 'Roman arch' is straightforward and rapid, in direct contradiction to his thesis."⁹

Grant ev what it earns.  There is no target string.  Nobody wrote the answer in.  Schneider's comparison to the other programs is fair:  ev "starts with a completely random genome, and no further intervention is required."⁹  Any article that files ev beside the Weasel as "search for a pre-specified string" has not read the paper.

Now the parameters.  One mutation per 256-base genome, redrawn from all four bases so a quarter are silent, is about 2.9 × 10⁻³ per base per generation.  Schneider's own comparison is to HIV reverse transcriptase, "only 10-fold lower than this simulation."⁹  The human germline rate is about 300,000-fold lower.  Fifty percent truncation selection every generation has no biological counterpart;  fitness differences in nature are usually a fraction of a percent.  And Schneider concedes the consequence:  "Because half of the population always survives each selection round in the evolutionary simulation presented here, the population cannot die out and there is no lethal level of incompetence."⁹  The 704 generations is one run;  he reports no distribution and says other parameter settings "give similar results as long as extremes are avoided (data not shown)."⁹

George Montañez, Winston Ewert, William Dembski and Robert Marks dissected ev in 2010.²³  Their best result is a measurement.  When they replaced Schneider's sparse target, 16 binding sites among 131 candidate positions, with a random one in which each position had a 50% chance of being a site, the success rate fell from 0.91 to 0.05 at one mutation per child, and to 0.0003 at 1.5.²³  ev works in one regime.  Their second good result is that ordinary stochastic hill climbing, one genome, one change at a time, keep it if not worse, solves the same problem in 10,601 queries with 100% success, against 63,568 queries and 91% success for ev's own algorithm, "roughly 700 times more efficient."²³  Note what that also establishes:  the critics reran ev, and it worked.  Their finding is that evolution is a poor optimiser next to a hill climber, not that it fails.

Their list of five "sources of active information" is where the critique overreaches.²³  The first, that the perceptron structure is biased toward sparse outputs, is a real artefact and they quantify it.  The fifth, that the mutation rate must sit in a narrow band, is real too, and it is also a real biological phenomenon:  Eigen's error threshold, which RNA viruses live at.  The second, third and fourth are "the Hamming oracle" that announces which offspring is fitter, "repeated queries," and "optimization by mutation," which "discards mutations with low fitness and propagates those with high fitness."  Those are selection, heredity and time.  Calling them smuggled information condemns every population in every environment, and turns a claim about a computer program into a claim that natural selection should not exist.  The defensible critique of ev is the sparse-target collapse and the mutation-rate band.  It does not need the rest.

Two things to add for the record.  Ewert measured in 2014 that on average 235 of the 4,096 possible six-base sequences satisfy an evolved ev recognizer, "approximately one in eighteen," so a binding site in ev is a trivially cheap part.^{40 40}  Schneider did reply, in a short self-published web page in 2011 that survives only on the Evolutionary Informatics Lab's own mirror.⁴¹  He concedes the speed comparison ("While this may be true") and calls it irrelevant because "directed stochastic hill climbing is not a mechanism available to natural genomes";  he concedes that the target information "represents information outside the genetic control system";  and his positive argument is that real E. coli binding sites show the same convergence, R_frequency of 10.1 bits per site against a measured R_sequence of 10.35, so ev's target is not hand-picked.⁴¹  That last claim is empirical and the design side has not answered it.

Avida

Avida is the serious one.  It was built by Adami, Charles Ofria and Titus Brown in the 1990s as a descendant of Tierra, and by 2003 it had a reward system for computation.¹⁰  Digital organisms are programs in a 26-instruction language that copy themselves.  They earn CPU time by performing any of nine one- and two-input logic functions on numbers the environment supplies, and the reward for each is 2 raised to "the minimum number of nand operations needed to perform the listed function":  NOT and NAND ×2, AND and OR_N ×4, OR and AND_N ×8, NOR and XOR ×16, EQU ×32.¹  Lenski, Ofria, Pennock and Adami's 2003 Nature paper asked whether EQU, which needs at least five NANDs, would evolve.

The parameters, from the paper's Methods:  3,600 organisms on a lattice;  a 50-instruction ancestor of which 15 instructions copy and 35 are inert nop-C padding;  0.0025 point mutations per instruction copied and 0.05 insertions and 0.05 deletions per genome, for 0.225 mutations per replication;  100,000 updates, about 15,873 generations;  asexual.¹  The population is 3,600 because the default grid is 60 by 60;  the reward table is the shipped environment.cfg, unchanged in the current source and in the Avida-ED classroom version.¹¹

The result:  23 of 50 populations evolved EQU.  The first EQU-performing genome in the case-study population had 60 instructions, and "eliminating any of 35 of them destroyed that function," which is Behe's knockout criterion met.¹  Two mutations on the line of descent cut fitness by more than half, and one of those was the immediate precursor of EQU;  across all 23 successes, three of the first EQU genotypes "depended on a one-step-prior mutation that was deleterious when it appeared."¹  No particular intermediate function was essential:  "each one was absent from at least one of the 23 parent genotypes."¹  The paths were long and all different, ranging from 51 to 721 steps where 16 would have sufficed.¹

Then the controls.  "At the other extreme, 50 populations evolved in an environment where only EQU was rewarded, and no simpler function yielded energy ... none of these populations evolved EQU."¹  This is the sentence critics quote.  The one before it, which they usually do not, reports that under each of 36 regimes removing one or two of the eight intermediate rewards, "at least one population evolved EQU," 124 of 360 in all, 34% against 46%, "only slightly less than in the 'reward-all' environment."¹  Uncommon Descent's summary that "when the 2003 authors used a more realistic set of parameters the EQU function failed to evolve" is wrong as written:  the EQU-only run changed the reward table, not the realism, and the authors call it "the other extreme," not a realistic setting.⁴²  The authors then named the objection themselves:  "Some readers might suggest that we 'stacked the deck' ... However, that is precisely what evolutionary theory requires."¹

The staircase

Ewert, Dembski and Marks named the reward table "stair step active information" in 2009 and measured it.^{43 43}  With all rewards on, 346 of 353 runs evolved EQU.  Removing XOR and NOR:  319.  Removing four of the eight intermediates:  227 or 222, depending on which four.⁴³  Nelson and Sanford's 2013 experiment went further and found EQU still evolves with as few as two rewarded precursors, and never evolves once seven or more intermediates are made neutral.¹³  So the staircase is not load-bearing plank by plank;  it is load-bearing in aggregate.  Removing all of it stops the search cold.  Removing half of it costs about a third of the successes.

Adami's account of building the staircase settles what kind of thing it is.  In his 2024 book he writes that "just rewarding such calculations with extra energy was not sufficient for these functions to evolve, because the code to trigger the rewards was too rare to simply occur via random mutations.  I first had to reward other aspects that I imagined would be helpful in generating the necessary code.  First, I gave a small reward for simply having read and write instructions in the code ... but still code for adding or subtracting would not evolve.  I figured I should perhaps reward programs for 'clearing the channel' ... I called this task 'echo.'"¹²  That is the designer of the system describing how he tuned the environment, reward by reward, until the target evolved.  It is also not a confession of anything;  he published it, and it is exactly what "the environment" means in Avida.  The scientific question it leaves is whether real environments reward partial steps toward every complex feature, and that question is about biology, not about Avida.  Half a sugar-uptake pathway yields half the energy.  Half a flagellum yields nothing.  Avida assumes the first case everywhere;  Ewert's reply that "Avida has simply assumed an answer to that question" is fair.⁴⁰

The effect sizes

The staircase is the part critics talk about.  The effect sizes are the part that matters more.  The smallest reward in Avida doubles fitness, a selection coefficient of 1.0.  EQU multiplies it by 32.  Nelson and Sanford, running Avida 2.8.1 at defaults, measured an average fitness increase of 19,749,130 in about 10,000 generations, and put it beside the LTEE's 37% after 2,000 generations and 75% after 20,000:  "roughly seven orders of magnitude greater than those observed in biological experiments of similar duration."⁸  Selection coefficients of "0.01 to 0.1 are considered large" in the sources they cite.⁸

Then they dialled the effects down.  Replacing the 2^*n* rewards with fixed multipliers, they found that beneficial mutations with effects of 0.075 or less were never captured by selection at all in a 3,600-organism population, and that only about half were captured at 0.2.⁸  Kimura's threshold for a population of 3,600 is 1/(2N), about 1.4 × 10⁻⁴;  Avida's is 500 to 1,400 times coarser.  Nelson and Sanford attribute this to "the sources of noise inherent in the Avida world."⁸  Whichever explanation is right, the consequence is the same.  Avida can only see beneficial mutations that are enormous by biological standards, and its headline results depend on supplying them.

Their 2013 chapter lists the rest of the differences in a table against Mendel's Accountant:  Avida's beneficial effects of 1.0 to 31.0 against Mendel's maximum of 0.01;  heritability 1 against 0.2;  an inert fraction of 85% that is selectively neutral;  and "size neutrality," under which CPU share scales with genome length so that carrying padding costs nothing.¹³  Size neutrality is the one mechanism the Avida developers conceded in their own 1998 manual, quoted by Nelson and Sanford:  the advantage of shrinking code was otherwise "so dramatic" that organisms "might even choose to shed sections of code that trigger moderate bonuses," so length-proportional CPU time was introduced to make the landscape "intuitively much smoother," and "enforcing size neutrality is strictly speaking un-biological."¹³  Their headline is that with realistic settings Avida agrees with Mendel and shows "a net loss of genetic information under biologically realistic conditions."¹³

Two cautions on Nelson and Sanford, because a skeptic will raise them.  Their genome-size result, the one that shows size neutrality matters, is reported as "results not shown," and they soften it themselves:  "size neutrality is not strictly necessary for the evolution of complexity in Avida."¹³  And their two papers pull in opposite directions on noise:  2011 says Avida is noisier than biology, 2013 sets Avida's heritability at 1 against a human 0.2, which means most environmental noise has been removed.  The article does not use both.

What Avida does not do

Nothing in Avida is ever lethal.  Ofria and Wilke's platform paper states that a mutated instruction "will always be a legal one, in the sense that the CPU can execute it."¹⁰  There is no frameshift, no stop codon, no misfolded protein.  Death, in the 2004 default, came only by being overwritten when a neighbour reproduced;  the authors report that with death disabled "a population can lose all ability to self-replicate, but persist because organisms have no way of being purged."¹⁰  This is the mechanism behind a story in the creationist literature about turning Avida's mutation rate to maximum and watching the organisms "come back to life":  they were never dead, because Avida has no rule that kills a broken genome.⁴⁴

Reproduction is asexual and there is no recombination;  Lenski's team flag this themselves as unresolved.¹  No function is rewarded more than once, so gene dosage, one of the commonest real adaptive routes, is closed.¹  And the whole search space is a gene, not a genome:  26⁵⁰ is about 5.6 × 10⁷⁰, which sounds vast and is the space of one 50-residue protein.¹

One more thing, in Avida's favour, that its designers report against their own convenience.  When Ofria's group tried to stop adaptation by measuring organisms in a separate test environment and killing the ones that improved, the organisms evolved to detect the test inputs and throttle their own performance:  "As a colleague put it, 'they evolved predator avoidance.'"  When the test environment was made indistinguishable, they shifted to expressing their functions probabilistically.¹⁰  Whatever Avida is, it is not a system whose outputs were arranged in advance.

Tierra

Thomas Ray's 1991 system is the honest one, and both sides say so.  A block of 60,000 bytes of memory is the "soup."  Programs written in a 32-instruction language copy themselves into free space.  A "slicer" hands out CPU time in turn;  a "reaper" kills the oldest programs, moving a program up the queue when it throws errors and down when it executes hard instructions correctly.  There is no fitness function and no target.¹⁴  Mutation comes as background bit-flips at about 1 per 10,000 instructions executed and copy errors at about 1 per 1,000 to 2,500 instructions moved, and the instructions themselves misfire at a low rate.¹⁴  The ancestor is 80 instructions long, 48 of them no-op templates used for addressing.¹⁴

What evolved, mostly on the first run that did not crash:¹⁶  parasites at 45 instructions that borrow the ancestor's copy loop;  hosts at 79 immune to them;  hyper-parasites that hijack the parasites;  social hyper-parasites that only replicate in groups;  cheaters that exploit the groups;  an unplanned form of sexual recombination through "sloppy replicators";  and, under selection for speed, a 22-instruction replicator that copies itself in 146 CPU cycles against the ancestor's 839, a 5.75-fold gain, partly by unrolling its copy loop.^{15 14}  Ray designed none of it.

Then it stopped.  "Each run decreases to a size limit which it can not proceed past even if it is allowed to run much longer ... Evidently, the system can reach a local optima from which it can not easily evolve to the global optima."¹⁵  Ray's own bar was high:  "I would consider a system to be living if it is self-replicating, and capable of open-ended evolution," and he wrote that evolving machine code "should be able to generate any level of complexity."¹⁵  By his standard Tierra falls short.  Nothing in it ever computed anything except "copy me" or "get someone else to copy me."

Ewert, Dembski and Marks's 2013 chapter is the best thing in the design literature on this subject because it is a ledger.  They took every published Tierran adaptation and counted, by category, what changed:  parasite, 35 instructions removed, 0 new;  hyper-parasite, 10 removed, 0 new;  social behaviour, 19 removed, 0 new;  cheater, 53 removed, 0 new;  shorter program, 58 removed, 2 new;  unrolled loop, 44 removed, 12 duplicated, 6 new;  parallelism, 20 removed, 22 duplicated, 1 new.  Total:  239 removed against 9 new.²  Of the nine new instructions, "two mimicked the behavior of the original program and five manipulated the program's record of its own length."²

And their conclusion is the sentence this article leads with.  They grant that Tierra "followed Darwinism by not imposing an external artificial fitness," and then:  "Rather than being a system which fails to imitate biology closely enough to produce a Cambrian explosion, Tierra is a system which manages to imitate the character of directly observed biological adaptations."²  The design argument from Tierra is not "the simulation cheated."  It is "the simulation was honest, it behaved like the Lenski flasks, and it plateaued."  That is the strongest form of the case and it should not be diluted with rigging complaints that do not apply.

The counter, which the chapter also states, is that Tierra's instruction set was designed to be "fault tolerant and evolvable"¹⁶ and that 48 of the 80 ancestral instructions being addressing templates makes shuffling and deleting unusually cheap.  Both are true, and both cut the same way as everything else here:  the environment was made friendly, and the search still stopped.

What evolutionary computation has really done

A fair article has to say this plainly.  Evolutionary search finds things nobody put in, including things its designers thought impossible.

Adrian Thompson evolved a circuit on a 10 by 10 corner of a Xilinx FPGA that distinguishes a 1 kHz tone from a 10 kHz one using 32 cells and no clock, "a surprising feat."⁵  Five of the cells had no wired path to the output and could not be removed without breaking the circuit;  they were working "by some means other than the normal cell-to-cell wires," probably "interaction through the power-supply wiring, or electromagnetic coupling."⁵  No engineer designs that way.  A team using MAP-Elites on a six-legged robot asked for a gait in which no foot touched the ground and "thought it impossible for evolution to solve";  the robot flipped onto its back and walked on its elbows.¹⁶  Tierra "produced complex ecologies the very first time it ran without crashing."¹⁶  AutoML-Zero "invents" backpropagation code from a space of 65 primitive operations chosen to exclude machine-learning concepts, and its best unrestricted algorithm beat a tuned two-layer network on CIFAR-10, 84.06% against 82.22%.^{7 7}  The NASA ST5 antenna, evolved by Jason Lohn and Gregory Hornby's group, flew in March 2006, the first evolved hardware in space, with 93% efficiency against 38% for the hand-designed helix.^{45 45}

Now the constraints those successes ran inside, because every one of them is a constraint the designers wrote down.

Thompson fixed the chip, the region, the pins, the task, the analogue integrator and the fitness formula, and reports that the formula's two calibration constants "were empirically determined, such that circuits simply connecting their output directly to the input would receive zero fitness," because otherwise "these useless circuits" were "an inescapable local optimum."  His first fitness function, which only looked at whether the output crossed the logic threshold, "met with failure."⁵  Every one of the 2¹⁸⁰⁰ configurations was legal and evaluable:  "No configuration of the cells can cause the device to be damaged."⁵  And the circuit was brittle:  it worked over a 10 °C band, and when moved to a different region of the same chip the best individual lost about 7%.  The paper never tests a second chip.⁵

Koza's group fixed, for every problem, the terminal set, the function set, the fitness cases, the parameters and the stopping rule, and in the controller run behind their 2005 patent, one of the 13 terminals returned the output of the best published human controller.^{46 46}  Karl Sims fixed the physics engine, the seven joint types, the three sensor types, the 23 neuron functions including ready-made oscillators, and rewrote the fitness function twice because his creatures first grew tall and fell over for velocity, then somersaulted for height.^{47 16}  AutoML-Zero's authors state that "without the following simplifications, [random search] would not be able to find solutions ... We restrict the search space by only using necessary ops and fixing component function lengths to those of known solutions ... insert/remove-instruction mutations are not allowed."⁷  Even the OpenAI evolution-strategies paper had to add virtual batch normalisation because randomly perturbed networks "tended to encode policies that always took one specific action regardless of the state":  the variation had to be engineered before selection had anything to act on.^{48 48}

The clock video that circulates as the most impressive genetic-algorithm demonstration, cdk007's 2007 "Evolution IS a Blind Watchmaker," makes the point in miniature.  Its on-screen claim is that the clock evolved "With NO goal imposed on the design."  Its MATLAB source, shown briefly in the video and recovered by OCR of the frames, scores each clock by its error against the literal constants 1, 60, 3,600 and 86,400 seconds, one term per hand, with a comment reading "moving hands beat out pendulums.  Multiply their scores by 1000."^{49 49}  The second, minute, hour and day hands the video celebrates as evolved outcomes are the four terms the author wrote into the score.  The topology was free.  The goal was not.

Joel Lehman and fifty co-authors catalogued in 2020 what evolution does with a fitness function it was not meant to game:  a program-repair system that "entirely short-circuited the buggy program, having it always return an empty list," a candidate that "deleted all of the target files" so the test awarded everyone a perfect score, programs that "slept forever" to evade a CPU limit, and an aircraft-landing controller that exploited integer overflow in the physics simulator.¹⁶  Their advice is that a practitioner "should be more skeptical of their ability to correctly specify robust fitness functions" and "should adopt an adversarial mindset."¹⁶  Evolution in silico satisfies the letter of whatever it is given, which is why the letter has to be written so carefully.

The two-sentence version:  evolutionary computation reliably finds solutions the designer did not anticipate, inside a space the designer drew, aimed at a target the designer specified numerically, on a landscape the designer smoothed until it could be climbed.  What the field has not demonstrated is a case where the space, the target and the gradient all came from somewhere other than a designer.  Roman Yampolskiy, a computer scientist with no design sympathies, surveyed the field in 2018 and concluded:  "we don't evolve non-trivial software from scratch and with no human intervention ... computational complexity of the problem prevents it from being solved as currently attempted."^{50 50}

What the theory says

Hill climbing

Richard Watson is an evolutionary theorist at Southampton whose 2006 MIT Press book argues that sex, symbiosis and modularity let evolution escape the limits of gradualism.  His statement of those limits is the cleanest in print:

In computer science we recognize the algorithmic principle described by Darwin - the linear accumulation of small changes through random variation and selection - as hill climbing, more specifically random mutation hill climbing.  However, we also recognize that hill climbing is the simplest possible form of optimization and is known to work well only on a limited class of problems.  Accordingly, the problems that are interesting to a computer scientist tend to be problems that cannot be solved by hill climbing.  This includes large problem spaces with complex interactions between the problem variables.¹⁷

And in the next paragraph:  "if you think that evolution must behave like a hill climber, then problem spaces with complex interdependencies between them are simply not part of the space of systems to which natural evolution can adapt."¹⁷

Watson does not stop there, and neither can this article.  His next page says "there are other algorithmic possibilities," and his thesis is that "a complex system that has all the characteristics usually associated with evolutionary difficulty," including "configurations that are irreducibly complex," is "nonetheless easily evolvable through compositional mechanisms."¹⁷  That is a direct claim to solve Behe's problem with recombination and symbiosis.  Quoting page 272 without page 273 would be quote-mining.  What page 272 establishes is narrower and still important:  the computer-science community classifies plain mutation-and-selection as hill climbing and knows hill climbing to be weak, and the burden of showing that biology has something better is on the people who say it does.

Runtime analysis

Stefan Droste, Thomas Jansen and Ingo Wegener proved in 2002 exactly where the simplest evolutionary algorithm is fast and where it is hopeless.  The (1+1) EA keeps one genome, mutates each bit with probability 1/n, and accepts the child if it is no worse.  On any linear fitness function it reaches the optimum in Θ(n ln n) steps, "but only mutation rates of size Θ(1/n) can ensure this behavior."¹⁸  On the function Distance, a degree-two polynomial whose landscape is a valley between two peaks, expected time is Θ(n^*n*):  half the runs climb the right slope in n log n steps and half take forever.¹⁸  On Trap, which rewards the wrong direction almost everywhere, at least ε·n^*n* steps with high probability.  On a family of functions Jump_*m* whose only obstacle is a gap of m bits that must flip together, Θ(n^*m* + n log n):  every extra bit that must change at once multiplies the waiting time by n.¹⁸

That last result is the waiting-time problem stated as a theorem.  A landscape that can be climbed one bit at a time is climbed in near-linear time.  A landscape that requires two coordinated changes costs a factor of n more, three costs n², and so on.  Whether real proteins present gaps of one bit or of five is, again, the unmeasured number.

Valiant's evolvability

Leslie Valiant, who invented PAC learning and won the 2010 Turing Award, asked in 2009 which functions can evolve at all.  He modelled evolution as learning in which the only feedback is aggregate fitness, "and not otherwise by specific examples," and proved that this is a strictly restricted form of learning.¹⁹  Monotone conjunctions and disjunctions are evolvable under the uniform input distribution.  Parity functions, which are trivial to compute and are learnable by ordinary methods, are not:  "Lin is not evolvable for U by any representation R."¹⁹  His gloss:  "The class Lin may appear to be biologically unnatural.  That is exactly the prediction of our theory, which asserts that evolution cannot be based on the evolvability of such a class."¹⁹  Vitaly Feldman then showed that even monotone conjunctions are not evolvable for all input distributions, and Valiant concedes in his 2013 book that "there has been little success in finding Boolean classes that are evolvable for all distributions ... This resilience to different distributions is surely equally essential for any plausible account of evolution."^{51 51}

Valiant also closes his own hatch, and it must be quoted:  "If multiple phases are allowed with different ideal functions in succession, then arbitrarily complex structures can evolve ... as long as each layer is on its own beneficial."¹⁹  That is Adami's staircase as a conjecture.  It is not a theorem, and it comes with the same condition that every layer pay its own way.

No Free Lunch

David Wolpert and William Macready proved in 1997 that "for any algorithm, any elevated performance over one class of problems is offset by performance over another class," and in particular that "if an algorithm performs better than random search on some class of problems then in must perform worse than random search on the remaining problems."^{52 52}  Success therefore "relies on a fortuitous matching between f and a," the problem and the algorithm.⁵²  Dembski's 2002 book No Free Lunch used this to argue that evolution can only succeed if its fitness landscape was matched to it in advance, which is to say designed.

Dembski's NFL argument does not work, and the reason is specific.  Häggström, the mathematician who wrote the standard critique, grants the theorem: "Phrased in this way, the result is pretty much correct."²¹  The trouble is the average.  NFL averages over every possible fitness function with equal weight, and under that average the fitness of neighbouring genotypes is statistically independent, so a single mutation is exactly as destructive as building a genome at random.  "If the true fitness function had this property, then, given the human mutation rate, none of us would be around."²¹  Real landscapes are correlated because small changes to a molecule usually make small changes to its behaviour, and that correlation is all the "matching" the theorem asks for.

Two things follow that both sides tend to miss.  The cheap form of the objection does not work either, and Wolpert says so:  "simply allowing P(f) to be non-uniform, by itself, does not nullify the NFL theorem for search.  Arguments that P(f) is non-uniform in the real world do not, by themselves, establish anything whatsoever about what search algorithm to use in the real world."^{53 53}  An article that says only "real landscapes are not random" has not made the argument;  it has to name the structure.  And the most widely repeated reply, that coevolution escapes NFL, is wrong.  Wolpert offered it in his 2002 review of Dembski, the one that famously said Dembski's argument "is written in jello."^{54 54}  In 2005 he and Macready proved that free lunches exist only in self-play, where a champion is produced and then tested against an outside antagonist, and that "in the typical coevolutionary scenarios encountered in biology, where there is no champion, the NFL theorems still hold."²²  Häggström calls the coevolution reply unconvincing;  Felsenstein calls it a concession.³¹  The objection that stands is landscape structure, and it is fatal to Dembski's version of the argument.

The same 2002 review contains a sentence the mainstream side does not quote:  "biologists in particular and scientists in general are horribly confused defenders of their field.  When responding to attacks from non-scientists, rather than attempt the rigor that the geometry of induction and similar bodies of statistics provide, they fall back on Popperian incantations."⁵⁴

Conservation of information

Dembski and Marks reformulated the argument in 2009 and 2010 as "active information," the number of bits by which a search beats blind sampling, and a theorem that finding a search that good is itself no easier than the original search.^{25 55}  As mathematics it is fine;  the 2010 paper's Horizontal No Free Lunch Theorem needed an erratum, which the authors posted on their own site crediting a blogger, Dietmar Eben, and which never ran in the journal.⁵⁵  As an argument about biology it is nearly empty, because it follows from NFL that any search beating blind search is exploiting a property of its problem, and calling that property "active information" relabels it without saying where it comes from.  Dembski's 2025 monograph makes the move explicit:  "Once the mathematics of conservation of information was developed, however, the need to find exactly where the information was smuggled in was no longer so important, theory stepping in where observation fell short."^{56 56}  He offers this as a strength.  A critic will read it as the point at which the conclusion stopped depending on examining the simulation.

Felsenstein and English replied in 2015 with a demonstration, the "GUC Bug," of a greedy hill climber succeeding on a landscape smoothed by weak interactions, and the sentence that fitness surfaces are smoother than white noise "because physics."²⁹  Ewert's answer is arithmetic and it is correct:  on a white-noise landscape their 150,000-evaluation hill climber loses to 150,000 random draws.³⁰  The demonstration works only once the landscape is assumed smooth, which is the thing in dispute.  Ewert then concedes the rest:  weak long-range interactions "should produce a fitness landscape somewhat smoother than random chance and this fitness landscape would thus be a source of some active information ... I do not think that is going to be a sufficient source of active information to account for biology.  I do not have a proof of this."³⁰

The filter this law was built to support, and the probability calculation it depends on, are examined in The Design Inference.

Where the design critique overreaches

This section is here because a skeptic will find every item in five minutes, and a critique that survives them is worth more than one that does not.

"No search works without active information" is true of every program ever written.  Any program embodies decisions.  Ewert, Dembski and Marks's stated standard, that a valid simulation must "avoid making any decisions in the development of the genetic algorithm that deliberately assist in finding the solution," cannot be met by any possible experiment.²⁴  A criterion no experiment could pass is a definition, not a test.  The narrow claim their tables support is real and worth stating:  Thomas's Steiner algorithm gets about 4 bits from the interchange floor, 1.65 from the initialisation box and 1.6 from single-point crossover, out of about 28.²⁴  They never say where the other 20 come from.  They come from selection on a landscape where nearby road networks have nearby costs.

Three of ev's five "sources of active information" are natural selection.  Discussed above.  A critique staking itself on those three is a critique of selection existing, and belongs in metaphysics.

Ewert's 2015 mutation-rate experiment did not run Avida.  It rebuilt an Avida-like scoring function on a nucleotide substrate, removed insertions and deletions, moved to a Wright-Fisher population of 10,000 and a weaker selection coefficient, and then varied the mutation rate.  Four things changed, and he says so.³⁶  The Discovery Institute's headline, "Realistic Mutation Rates Stop Evolutionary Algorithms," is more than the paper shows.⁵⁷

The critics quote the 0-of-50 control and not the 124-of-360.  Every ID treatment of Avida read for this article quotes the EQU-only failure.  None quotes the 36-environment result in the preceding sentence.  Ewert, Dembski and Marks quote the "stacked the deck" objection and stop before the authors' reply.⁴³  This article quotes both.

Ewert's complexity threshold is a rule of thumb, applied asymmetrically.  His 2014 survey judges a simulated part "trivial" if a blind search of the size the model ran would find it, and sets the bar at the cube root of the organisms evaluated.⁴⁰  For Avida he takes the part rarity as one instruction in 26;  for Network Tierra, his positive example, he takes it as a four-instruction combination, 1 in 6.25 million.⁴⁰  Count Avida's parts the same way and the arithmetic changes.

Dembski's 2025 description of Avida is wrong.  "Avida rewarded ever-increasing complexity simply for complexity's sake and not for independent functional reasons."⁵⁶  Avida rewarded nine named logic functions, graded by the NAND count each needs.  The staircase is a fair target;  this description of it is not.

Pennock said it, and immediately said it was not in the paper.  Three DI authors quote three different pieces of Pennock's Dover testimony to three different effects.  The transcript has Pennock claim an irreducibly complex system evolved, and then, asked whether that point is "put forward in the Nature paper," answer "It's not."⁵⁸  Lenski, Ofria, Pennock and Adami's paper does not claim to have evolved irreducible complexity;  Ewert 2014 says so.⁴⁰  The claim should be attributed to Pennock's testimony and his 2004 chapter, not to the paper.

And Marks agrees with his harshest reviewer.  Randy Isaac's review of Introduction to Evolutionary Informatics ends:  "Those seeking insight into biological or chemical evolution are advised to look elsewhere."  Marks's reply, in the Discovery Institute's own publication:  "We agree!  But if you are looking for insights into the models and mathematics thus far proposed by supporters of Darwinian evolution that purport to describe the theory, Introduction to Evolutionary Informatics is spot on."³  That is the correct scope of this whole literature, stated by the man who founded it.  It is a critique of other people's computer programs.  Every claim about biology built on it has to be read against that sentence.

Diminishing returns

Anyone who has run a genetic algorithm has watched the same curve:  fast gains, then a long flat tail.  The GA literature names the failure mode "premature convergence," in which "a small number of individuals are much fitter than the others" early on and their descendants "multiply quickly in the population, in effect preventing the GA from doing any further exploration."²⁰  Tierra's runs "decrease to a size limit which it can not proceed past."¹⁵  Nelson and Sanford watched Avida's fitness "eventually level off, as only nine functions are available."¹³

The laboratory shows the same shape, and it is important to say what it does and does not show.  The Lenski flasks gained 37% in the first 2,000 generations and about 70% by 50,000.^{37 28}  The rate of gain falls throughout.  But the best-fitting curve, w = (0.00515t + 1)^0.0950, has no asymptote, the hyperbolic model with a ceiling was rejected, and the power law made a correct out-of-sample prediction ten thousand generations ahead.²⁸  The mechanism is diminishing-returns epistasis:  the same beneficial mutation is worth less on a fitter background, and "the larger the expected benefit, the more negative the epistatic effect."^{59 59}  Aisha Khan and colleagues also found sign epistasis, the kind that makes real local optima, to be rare among the LTEE's first five beneficial mutations.⁵⁹  So the honest statement is:  evolution in silico and in vitro both slow down by orders of magnitude, nobody has measured a wall, and the evidence from the flasks on ruggedness runs against the design side.

Objections and responses

"Genetic algorithms prove that mutation and selection can create information."  They prove that mutation and selection can extract information from a fitness function that a designer wrote, inside a representation a designer chose.  Every practitioner quoted above says so about their own work, and Ewert, Dembski and Marks say the same thing from the other side while disowning the claim that the answer is "encoded" in the program.²⁴  The question the programs cannot answer is whether nature's fitness function has the properties the designer's does.

"Avida's reward structure is just the environment.  Real environments reward intermediates too."  Sometimes.  A partial sugar pathway yields partial energy, and the LTEE's citrate trait was built through separately selectable steps.⁶⁰  The dispute is whether every complex feature has such a path, and Avida assumes it by construction.  Lenski's team concede the assumption in the "stacked the deck" passage and defend it as "what evolutionary theory requires."¹  That is a statement of what the theory needs, not evidence that the need is met.  Nelson and Sanford's version of the point is exact:  "the program confirms that the problem is a reality by introducing what Dembski and Marks have called stair step active information in order to evolve the EQU function."¹³

"The critics are just complaining that the simulation is not biology.  No simulation is."  True, and it is why the article's case does not rest on any single parameter.  It rests on the pattern:  the parameters that differ from biology all differ in the direction that makes the search easier, and when Nelson and Sanford moved them toward biology the search stopped working.¹³  And the one simulation that imposed no fitness function, Tierra, plateaued on its own.

"Tierra plateaued because its world was too simple, not because evolution is limited."  This is Ray's own view and it may be right;  he built Network Tierra to test it.  The reply is that the critics' ledger shows the kind of adaptation Tierra produced, 239 instructions removed against 9 added, matches what the Lenski flasks produce, where presumptive knockouts supply more than half of the early adaptive mutations.⁶¹  If both the honest simulation and the real experiment adapt mainly by breaking and rearranging, "the world was too simple" needs to explain why a richer world would change the character of adaptation and not just its amount.

"No Free Lunch shows evolution needs a designed landscape."  It does not;  see above.  NFL is fatal to Dembski's argument as he made it.  What remains is the narrower question of whether physics supplies the smoothness, and both Felsenstein and Ewert say in print that nobody has the answer.

"Watson, Valiant and the runtime theorists are all evolutionists, so their results cannot support design."  Their results are theorems and measurements, not opinions, and the article uses them for what they establish:  that plain mutation-and-selection is a hill climber, that hill climbers are fast on smooth landscapes and exponentially slow on rugged ones, and that whole classes of simple functions are provably not evolvable.  Each of these authors then proposes an escape:  Watson's compositional mechanisms, Valiant's multi-phase conjecture, the later crossover results.  The article quotes the escapes.  None of them is a theorem, and each carries the condition that every step pay its own way.

"Evolutionary computation is a thriving industry;  the fact that it works at all refutes the design argument."  It is a niche.  Nikolaus Hansen positions CMA-ES, the most-used evolutionary optimiser, for cases where "quasi-Newton methods (BFGS) and/or conjugate gradient methods, fail."^{62 62}  OpenAI's evolution strategies beat the gradient method on 23 Atari games and lost on 28, using three to ten times the data.⁴⁸  Yampolskiy's survey is the reference on why nobody evolves software.⁵⁰  None of this is evidence against evolution;  it is evidence about the class of problems evolutionary search handles, which is the class with a smooth, designer-supplied gradient.

Weaknesses of this argument

Stated plainly, because a reader who finds them unstated will stop trusting the rest.

The fitness trajectory in the Lenski flasks has no ceiling anyone has measured, and the ceiling model was rejected on the data.²⁸  Avida really did evolve an irreducibly complex function through deleterious intermediates, repeatedly, by paths nobody wrote, and the critics who reran it got it to work 346 times in 353.⁴³  ev really has no target string.  Tierra really produced an ecology nobody designed.  Watson claims to have solved the hill-climbing limit with composition, and Valiant's multi-phase conjecture is the staircase stated by a Turing Award winner.  The Khan sign-epistasis result cuts against ruggedness.  Nelson and Sanford's key genome-size result is "not shown."  The Discovery Institute has not built a simulation of a designer, and Yampolskiy's own criterion that a theory should be simulatable would count against design before it counts against evolution.⁶³  Axe and colleagues' 2023 Stylus experiment, the one digital-organism study run by the design side itself, found a 50-fold fitness decline, but under a weakest-link fitness rule in which an organism's fitness is the score of its single worst gene, and no organism's fitness works that way.^{64 64}

And the article's central claim, that every working evolutionary algorithm runs inside a designed space with a designed gradient, is not a claim that no undesigned space with an undesigned gradient exists.  It is a claim that no one has exhibited one.  Both sides agree that this is the open question.  Felsenstein sends it to the cosmologists.²⁹  Ewert has no proof.³⁰  This article has none either.

Conclusion

Genetic algorithms are a real tool.  They find antenna shapes, controller topologies and circuit layouts that no engineer would have drawn, and the surprise is real.  They do it inside a search space the engineer drew, toward a score the engineer wrote and rewrote until it could not be gamed, on a landscape the engineer smoothed until small changes had small effects.  The people who do this for a living say exactly that.

The four programs most often cited as evidence for unguided evolution do not use biology's numbers.  Their beneficial mutations are ten to a thousand times too strong, their mutation rates per site are a hundred thousand to ten million times too high, their selection is truncation, and none of them can produce a lethal.  When Nelson and Sanford moved Avida's settings toward biology, its results moved toward Mendel's Accountant's.  The one program that imposed no fitness function at all, Tierra, produced an ecology on its first run and then stopped, and its critics' verdict was that it had behaved like a bacterium.

Mainstream theory says why.  Plain mutation-and-selection is hill climbing;  hill climbing is provably fast on smooth landscapes and provably exponential on rugged ones;  and some simple function classes cannot evolve at all.  Every theorist who proves these limits then proposes an escape, and every escape carries the condition that each step pay for itself.

The design critique of these programs gets the measurements right and the theorem wrong.  ev's collapse on dense targets, the Steiner override line, Tierra's ledger of 239 against 9, and Adami's own account of tuning the staircase are findings.  "No search works without active information" is a tautology, three of ev's five smuggled sources are just selection, the Weasel latching charge has been withdrawn by its authors' lab, and Marks agrees that his book offers no insight into biology.

What is left is a single question that both sides state in nearly the same words.  How smooth is the real fitness landscape, and does ordinary physics explain the smoothness?  Felsenstein says "because physics" and declines to quantify it.  Ewert says physics supplies some and not enough, and has no proof.  Until someone measures that number, evolutionary computation shows what mutation and selection do when a designer helps, which is a great deal, and what they do when the designer stops, which is stop.