Simplification: Evolution by Subtraction

I An Animal Made of Only a Few Cells

There is a group of parasites called myxozoans. In their parasitic stage, they usually consist of only a few cells, measuring tens to hundreds of micrometers across. No body axis, no tentacles, no mouth, no complete digestive cavity, no nervous system.

One species of Kudoa has a nuclear genome of only about 22.5 million bases, with approximately 5,533 protein-coding genes.

From the late nineteenth century into the twentieth, these organisms were classified as “protists.” In other words, they did not count as animals. The reason was precisely their simplicity: so few cells that almost nothing resembling a typical animal body plan could be seen.

In 1994, an analysis using a ribosomal gene provided the first strong evidence that these “myxozoan protists” were actually metazoans. Over the next twenty years, the evidence accumulated, layer by layer.

A 2015 phylogenomic study used 200 protein markers and 51,940 comparable sites. Its sampling covered several major myxozoan lineages, 22 cnidarians, 38 other animals, and nine unicellular opisthokonts. Both Bayesian and maximum-likelihood analyses placed “myxozoans plus Polypodium” within Cnidaria, with a posterior probability of 1.0 and 100 percent bootstrap support at the crucial node.

In other words, this thing made of a handful of cells is a relative of jellyfish and sea anemones. The fingerprint it retained is called a polar capsule. Inside the spore is a little organ capable of turning a threadlike structure outward. Its shape had long suggested that it was homologous to the nematocysts of cnidarians.

Then researchers found genes for minicollagens and nematogalectins, components specific to nematocysts. A resemblance in shape had become strong evidence of homology. The same pattern has a second, more immediately visible version: the rhizocephalans.

These are parasitic cirripedes, the group that includes barnacles. Once the parasitic female matures, she no longer has an ordinary crustacean’s segmented body, thoracic legs, or complete digestive tract. In their place is a rootlike feeding system extending through the body of a crustacean host. Outside, all that remains is a sac devoted to reproduction.

Looking at the adult alone, nobody would recognize a crustacean. Its identity as a cirripede rests chiefly on the larvae, not the adult. It still produces recognizable nauplius larvae, followed by larval stages with unmistakably cirripede features.

Those swimming appendages and sensory and attachment structures connect it clearly to other cirripedes. Between them, these two cases lay out every question this essay will ask.

This is no exercise in academic fussiness. Confuse the two explanations, and the root of the entire animal tree ends up in the wrong place. The reconstructed ancestral states, including when nerves, muscles, and guts first appeared, all go wrong with it. Section VII will show a version of this problem that remains very much alive today, right at the root of the animal tree.

When we see something simple, there are two possible explanations: it has always been that way, or it was once more complex and subsequently lost things. In morphological terms, the two explanations look exactly alike.

What separates them is not “how many organs are left,” but a position on the phylogenetic tree, remnants preserved during development, and the highly specific molecular components that remain.

One term needs defining at the outset. In this series, “secondary simplification” refers specifically to cases in which phylogenetic or developmental evidence has established that ancestors possessed more structures, and their descendants show a reduction on a specified scale. The term does not mean “simple, therefore primitive.” Nor does it carry any built-in claim about the mechanism.

This essay is about subtraction. It asks three questions: When can subtraction be demonstrated? How should we use the ruler called “complexity”? And does this series’ criterion for a phase transition still work when applied in reverse? That last question is the real destination. The first seven sections are assembling what we need to reach it.

II Different Rulers, Different Answers

First, why are there so many parasites in this essay? Parasitism has evolved independently at least 223 times among animals, across fifteen phyla. The ten largest parasitic lineages contain about 90 percent of described parasitic animal species. Include estimates for the many species not yet described, and parasites may account for roughly 40 percent of all animal species.

That makes parasitism a substantial part of animal life, not a sideshow. But 40 percent is not a majority, and parasitism is not synonymous with reduction. The interesting question is more precise: how can some lineages lose so much and continue to thrive?

Free-living does not mean self-sufficient in every respect: animals already obtain essential nutrients from food. Inside a host, however, some substances an organism previously had to make may become reliably available from outside. The enzymes that produced them may cease to be indispensable. Maintaining genes and producing proteins also have costs, though the balance differs among lineages and life stages.

When those functions matter less, selection against damaging mutations can relax. Mutations may persist through drift, eventually degrading a pathway; in some circumstances, saving costs may itself be favored. What a host supplies may become less necessary to manufacture. This is one route to parasitic reduction, not a universal explanation for every loss in this essay.

Now for the point this section is really here to establish: “simplification” is not a quantity. It is a statement made with a particular ruler. Change the ruler, and the same object can give answers pointing in opposite directions.

Take the myxozoan again. Measured by “number of bases in the genome,” it is much smaller than its cnidarian relatives that have not undergone such extreme simplification: about 22.5 million bases, against approximately 561 million in its sister, Polypodium. A difference of more than twentyfold.

But measured by “number of protein-coding genes,” it retains about a third as many: roughly 5,533 against 17,440. Two rulers. One says more than twentyfold; the other says threefold.

The difference lies in the noncoding space within genes. Average intron length has been compressed from 1,163 bases in Polypodium to about 82 in the myxozoan. The genes remain; their internal noncoding space has been squeezed dry. Parasitic plants offer an even more drastic example.

Balanophora himalayana does not photosynthesize. It spends most of its life buried in its host’s roots, sending a short projection aboveground only when it flowers, rather like a mushroom. Its nuclear genome is about 1.92 billion bases long. Nothing small about that.

Count its protein-coding genes, though, and there are only about 13,670. Compared with its nonparasitic angiosperm ancestor, it is inferred to have lost roughly 54 percent of the ancestral genes. The losses are concentrated in photosynthesis, chloroplast functions, the construction of a typical plant body, and several defense processes.

If the ruler is total DNA, it can look enormous. If the ruler is the proportion of ancestral genes retained, it is severely reduced. Both statements concern the same plant, the same sequencing effort, the same paper.

We can set down a general rule here, because the coming sections will keep running into this kind of trouble: whenever a number is used to support a trend, first ask what counting rules produced it. This has come up several times in the series. Episode IV examined mineral counts; Episode V, organelle gene counts; Episode X, the definition of an “explosion.” From here on, this essay will not rehearse the rule every time. It will flag it only where the counting rules can turn the conclusion upside down.

Now a different question: does every parasite undergo reduction? No. There are plenty of counterexamples.

Strongyloides and its relatives offer an unusually good comparison. The parasitic species and their free-living close relatives have genome sizes and gene counts of the same order, with differences far smaller than those seen in myxozoans. More important is what went the other way: about 1,075 new gene families are inferred to have appeared on the branch leading to five parasitic species, with substantial expansions in metalloproteases, secreted proteins, and acetylcholinesterases.

A lineage becoming parasitic can lose things and grow new ones at the same time. The jewel wasp, a parasitoid, has a genome of about 267 million bases and 14,221 protein-coding genes, sitting quite respectably within the ordinary insect range. So what did the lineages that really did shrink actually lose?

Apicomplexans provide the best example, because they have close photosynthetic relatives still living freely that we can use for comparison. Plasmodium falciparum, the malaria parasite, has a nuclear genome of about 23 million bases and around 5,300 genes. Two very closely related lineages remain free-living and photosynthetic, with gene counts on the order of 20,000 to several tens of thousands.

Different parasites lose different pathways, including some routes for making amino acids, lipids, or sterols. We should not apply the whole list to Plasmodium: it retains apicoplast fatty-acid synthesis needed at particular life stages. Changes in DNA repair cannot simply be described as the host doing the job. Meanwhile, gene families involved in host invasion, immune evasion, and life-cycle transitions can expand.

Microsporidia occupy the extreme end. Encephalitozoon cuniculi has about 2.9 million bases and around 2,000 protein-coding genes.

What does 2,000 genes mean? Escherichia coli has about 4,000, yeast about 6,000, and humans about 20,000. Here is a eukaryote getting by with less equipment than a bacterium.

Yet even as it compresses its metabolism and genome this far, it retains a highly specialized apparatus for invading hosts and a life cycle that is anything but simple. Reduction is selective, not across the board. Is there, then, a lower limit?

No. There is no universal “minimum number of genes” that applies to every parasite. What must be retained depends on what the host can provide, where in the host the parasite lives, whether it can import metabolites directly, and whether it still has a free-living stage. The supposed lower limit moves with the ecological network.

Bacteria lead to the same conclusion, with a reminder pointing the other way: free-living bacteria in the nutrient-poor open ocean can evolve very small genomes too. They parasitize nothing. They have simply streamlined themselves. Small size does not automatically establish parasitic reduction.

III The Same Traits, Lost Again and Again

Cave organisms provide nature’s best replicated experiments. The same thing has happened many times in unrelated places, and in each instance we can see the starting point.

More than 200 species of cavefish have been described. They did not radiate from a single shared “blind-fish ancestor.” In several different fish lineages, sighted surface ancestors independently entered subterranean environments. Reduced eyes and pigmentation are the most conspicuous convergences, alongside changes in circadian rhythms and some behaviors, and enhancements of taste, smell, the lateral line, and feeding systems. Nor does every cave species lose all the traits in question.

The Mexican tetra alone has more than thirty cave populations, with recent surveys putting the number at around thirty-four. Geographically and genetically, they cannot be traced to one simple origin. That lets us ask a very pointed question: is the same phenotype produced by the same set of mutations?

First, something easily misunderstood. A cavefish’s lack of eyes does not mean “the embryo never builds eyes.” Early embryos form eye primordia, lenses, and retinas. Then the lens undergoes programmed cell death, followed by secondary reduction of the retina and eyeball.

Formation first, regression afterward: an entirely different history from never having had eyes in the ancestral line. Why are eyes lost? There are three sets of mechanisms, and each has evidence behind it.

The first is neutrality and relaxed selection. In darkness, vision no longer contributes what it once did to fitness. Purifying selection on vision-related genes weakens, allowing harmful mutations to become fixed more readily.

This predicts that independent cave lineages will lose vision through different genes and different mutational routes. Recent genomic work has indeed found more sites showing signals of relaxed selection in obligate cave lineages.

The second is direct selection: saving energy. Measurements in Mexican tetras found that in a surface fish weighing about one gram, the eyes and the midbrain optic tectum, which processes visual information, together account for roughly 15 percent of resting metabolism.

As body size increases, that proportion falls to around 5 percent. Eye tissue alone accounts for about 8 percent in small fish. Include the entire brain, and eyes plus brain can consume about 23 percent in a one-gram surface fish, compared with roughly 10 percent in the cave form.

This demonstrates a substantial cost to maintaining vision. It does not demonstrate that selection for energy savings drove the historical reduction of eyes. It answers the question, “Is there a cost large enough for selection to act on?” Yes, there is. It does not answer, “Did the decisive historical selection coefficient come from the cost of eyes, or from the benefits of other traits?”

The third is pleiotropy, or indirect selection, and here the experimental developmental evidence is most direct. A signaling pathway active in the embryonic head operates over a wider domain in the cave form. It can suppress expression of a key regulatory gene associated with the eye field and promote lens-cell death. The same pathway also promotes the expansion of feeding structures, including jaws and taste buds.

In genetic crosses, eye size is inversely associated with jaw and taste-bud phenotypes. Experimentally enhancing or inhibiting the pathway changes both sets of traits. There is therefore clear molecular and developmental evidence for the proposition that “positive selection for better feeding can, through pleiotropy, reduce the eyes as a side effect.”

Quantitative genetics supplies another clue. An analysis of 539 second-generation hybrid individuals produced a genetic map spanning roughly 2,200 centimorgans. It detected numerous quantitative trait loci associated with eye and lens size, concluding that more than ten loci were involved at a minimum. This is no single “blindness gene.”

That study also gave pure drift a problem. At the eye-related loci detected, alleles from the cave parent almost uniformly pushed eye size downward. If the result came entirely from the random accumulation of neutral mutations, we would expect a mixture of mutations that enlarged eyes and mutations that reduced them.

Interestingly, the pigment-related loci in the same study showed a much more mixed pattern of directions. Pigment loss is therefore more readily compatible with drift or relaxed selection than eye loss. “Cave traits” cannot all be bundled into a single mechanistic explanation.

Nor do different cave populations always use the same genes. Cross blind fish from different cave populations, and their first-generation offspring can sometimes develop eyes larger than either parent’s, even partially recovering vision. That indicates defects at different, complementary loci in the two parents. At the same time, some genes and loci really are used repeatedly. Evolutionary routes contain both parallel and nonparallel elements.

The three mechanisms are not mutually exclusive. One plausible model layers them through time: direct or indirect selection initially drives reduction of the phenotype; once vision-related genes lose their function, selection on them relaxes. Their relative contributions may also differ between cave populations. Finally, reversibility, because it bears directly on Section VII.

The strongest evidence of “recovery” is not a cave population fully re-evolving its ancestral eyes in a natural surface environment. It is genetic complementation: cross blind fish from different cave populations, and the offspring can recover some eye structures and some vision. This is reconstruction through complementation.

The developmental evidence is stronger still. Transplant a lens from the surface form into the eye cup of a cave-form embryo, and the cavefish’s eye tissues can continue to grow, restoring a considerable degree of normal eye morphology. Transplant in the opposite direction, and a normal surface eye cannot be maintained.

This shows that “the phenotype no longer has eyes” and “every genetic module for building eyes has disappeared” are two different things. The phenotype has been subtracted. Some of the downstream parts are still there, waiting.

IV Neither Parasites nor Cave Dwellers

Subtraction requires neither parasitism nor caves. This section widens the field, because the judgment in Section VIII will carry weight only once we have seen enough different routes to loss.

A sessile life is the first route. Miniaturization is the second. Extreme environments, the third. Reduced predation pressure, the fourth. A return to water, the fifth. Five unrelated routes, all capable of producing net losses. That alone tells us that subtraction is not the by-product of any one way of life.

The following examples share a pattern. We will state it once rather than repeat it for each: loss in one respect can coexist with gains in another. What has been measured is the set of things lost. What was kept, and what was newly built, usually does not make it into the same headline.

The first route is settling down. A sea-squirt larva looks like a little tadpole: it has a tail, a notochord, a dorsal neural tube, a clear front-to-back axis, and the ability to swim.

After swimming for a while, it finds a rock, plants itself headfirst, sticks, and begins to metamorphose. The tail is reabsorbed; the notochord and most larval neural structures disappear. What remains grows into a filter-feeding bag fixed to the rock.

This is no “primitive animal without a notochord.” The larval notochord, dorsal nerve cord, and tail muscles, together with molecular phylogenies, place sea squirts firmly within Chordata. An adult without a notochord is a different thing from a lineage that never had one. Meanwhile, the adult develops a large branchial basket and an entire filter-feeding system.

The second route is getting smaller. Extreme miniaturization repeatedly reduces the number of skeletal elements. Comparative work on frogs has found that tiny-bodied lineages are more likely to lose or fuse certain skull bones and bones of the fingers and toes.

The mechanism need not be “selection removes unused organs.” Limited space for development, shifts in developmental timing, the retention of juvenile features, functional reintegration, and mechanical constraints can all prevent bones from forming or cause them to grow together. When an animal becomes smaller, its parts are not simply taken away one by one. The whole assembly process is run again in a smaller space.

The third route is extreme environments. The tube-dwelling worms at deep-sea hydrothermal vents have neither a functional mouth nor a gut as adults. They are fed by symbiotic bacteria inside them. The bacteria use chemical energy from sulfides to synthesize organic matter; the worms provide accommodation and raw materials.

Yet digestive structures may form temporarily in the larvae, then be lost or remodeled in the adult. Again, the same pattern: the larva accounts for what the adult no longer reveals.

Deep-sea fish show reduction, not an across-the-board decline. The genomes of trench-dwelling snailfish show losses or degeneration of vision-related genes. Some hadal species also have smaller olfactory-receptor families than their shallow-water relatives, while components involved in pressure adaptation, membrane homeostasis, and other aspects of chemical sensing may expand.

The fourth route is the absence of predators. Flight loss on islands is among the most frequent structural losses in animal macroevolution. It occurs in more than half of all bird orders and has evolved independently many times. Exactly how many depends on how one counts: estimates range from a hundred or so independent origins to more than a thousand island lineages evolving toward flightlessness.

A global analysis in 2020 produced an even more interesting result. Include extinct island birds, and flightlessness was historically far more common than today’s few dozen flightless species suggest. We see so few flightless birds today chiefly because humans have already wiped out most of them.

Mechanistically, reduced predation pressure is one of the most consistent variables. Phylogenetic comparisons of island birds that can still fly show that the fewer kinds of predators an island has, the more pronounced the morphological shifts “toward flightlessness,” such as smaller pectoral muscles and relatively longer hindlimbs.

The evidence from armor is harder still. Freshwater populations of three-spined sticklebacks have repeatedly reduced their bony plates and spines. Experiments show that environments where certain aquatic insects act as predators directly favor genotypes with shorter spines and less armor. Net loss of defensive structures can be selected for, not merely allowed to happen.

The fifth route is a return to water. Across different duckweed genera, a sequence emerges: progressively fewer cortical layers, greatly simplified root tissues, and finally the complete loss of roots themselves. Yet they must still manage the complexities of life at the water’s surface and reproduce rapidly by cloning.

In their independent returns to the ocean, seagrasses have repeatedly lost genes associated with terrestrial conditions, including some functions involved in stomata, volatile signaling, ethylene, and terpene metabolism. At the same time, they have acquired a whole new repertoire for ion homeostasis in seawater and pollination beneath the sea. Five routes, five entirely different causes, all producing net loss.

V The Ruler Called “Complexity”

Start with a worm. Suppose it originally had a hundred body segments and later had twenty. The number of parts has fallen. But suppose those twenty segments have differentiated into entirely distinct functional units: some handling feeding, some locomotion, some reproduction. Differentiation among the parts has increased.

The same animal, the same change. One ruler says it has become simpler. Another says it has become more complex. Neither has measured incorrectly.

That is the issue for this section: “complexity” is not a quantity waiting to be measured. It is a family of quantities from which we have to choose. There are at least nine measures we can actually put to use.

The most intuitive is the number of cell types: count how many reproducibly distinguishable kinds of cells can be identified by morphology, histology, or gene-expression state. Humans have a little over two hundred; sponges only ten or so. This ruler asks whether the variety of differentiated types has increased or decreased. It measures something distinct from gene count, behavioral complexity, and the complexity of ecological networks.

Incidentally, this ruler has become genuinely useful only in recent years. As techniques for measuring gene expression one cell at a time matured, “How many kinds of cells are there?” ceased to be simply a question of what the eye could distinguish and became a statistical question.

Next comes the pair most easily lumped together: the number of morphological parts, meaning countable units such as bones, limb segments, appendage types, teeth, and body segments; and differentiation among parts, meaning whether the same number of parts has become more specialized. The worm at the beginning shows how they can move in opposite directions.

The remaining measures address other levels. Gene count and the number of gene families measure the size of the gene repertoire, which has only a weak relationship with morphological complexity: duplication, alternative splicing, and regulation all alter the mapping between them. Regulatory-network size measures the scale of organization at the regulatory level.

The number of developmental states measures the length of a developmental trajectory and how many branches it takes. The number of hierarchical levels measures how many layers of parts nested within wholes there are. Information-theoretic measures calculate how much information about the environment is stored in a sequence. All nine rulers are real. They do not add up to one overall score.

This is not a casual list. Separating “complexity” into several nonequivalent kinds is itself a methodological claim: different structures and processes must be quantified separately. We cannot combine them into a single score first and then go looking for a trend.

Only a few of the nine have been measured systematically on a macroevolutionary scale. Counting developmental states is particularly difficult: how states are defined depends heavily on the resolution of observation. A new generation of microscopes can change the count. The information-theoretic measure presents the reverse case. In digital life it can be calculated directly and defined rigorously, but what it measures is not morphology at all.

One example shows especially well what this family of rulers can do. Researchers measured the differentiation of serial appendages in sixty-six crustacean taxa: in plain terms, how many different kinds of legs had developed from a row of legs that originally looked much the same. They detected widespread increases in complexity across several lineages, which the authors regarded as a rare large-scale example compatible with a “driven trend.”

In crustacean legs, complexity really is increasing. This is one of the few positive results in the essay, and its scope is precisely the scope within which it was measured: particular taxa, a particular ruler.

The general rule from Section II therefore takes a specific form here. When you read that “something has become more complex,” the first question is not “Is that true?” It is “Which ruler did they use?”

VI Passive Trends, and What the Null Hypothesis Should Be

Now for the most abstract section of this essay, and the most useful. In 1994, researchers separated the observation that “average complexity rises over time” into two fundamentally different processes.

A driven trend: through most of the state space, a persistent force or bias in transitions within lineages favors increasing complexity. Even far from the lower bound, increases systematically outnumber decreases.

A passive trend: change within a lineage can be approximately directionless, or even locally biased toward decreases. But its starting point lies close to a lower bound that cannot be crossed.

The space below is cut off, while there is no equivalent restriction above. Diffusion consequently occupies more of the upper space, and the mean and maximum of the population rise over time. The two processes can produce curves that look exactly alike. That is the difficulty.

The mechanism of a passive trend can be stated quite concretely. Imagine a set of lineages starting near the lowest permissible complexity. Each change is itself directionless: upward and downward moves have equal probability. But on the downward side there is a wall. Below a certain level, an organism would no longer be an organism.

Downward mutations are turned back, while upward ones face no equivalent restriction. Even if each individual change is directionless, the spreading variance can make the right tail, the maximum, and even the mean rise over time.

A wall and a random walk are enough to make a curve look as though it has a direction. This is no metaphor. It is a model that can be written down and tested with Monte Carlo simulations. There are explicit statistical ways to distinguish the two processes.

The crucial question is not “Did the mean rise?” but “Did the lower bound move too?” A driven trend predicts that the minimum will rise over time. A passive-diffusion model allows the minimum to remain close to a fixed lower bound for a long time while the mean and maximum rise.

The second test pairs ancestors and descendants far from the boundary. Exclude the lineages close enough to the lower bound that geometry prevents them from decreasing. Only if descendants still show significantly more increases in complexity than decreases does the result support a driven trend. If increases and decreases are roughly balanced, or decreases are even more common, while the overall distribution continues to expand toward higher values, the result fits the passive model better.

The third test is the skewness of distributions within clades. Starting from an ancestor near the lower bound, unbiased diffusion produces a right-skewed distribution stretched toward higher values. Within local subtrees far from that bound, the direction of skew can again help distinguish a general driving force from a boundary effect. What do the results actually show?

The 1996 review of metazoans concluded that, for metazoans as a whole, convincing evidence of a long-term increase was very limited. The clearer signals were concentrated chiefly in the early Phanerozoic and in particular dimensions of complexity.

To date, there is no widely accepted quantitative finding that “complexity generally increases over time across all groups of life and all measures of complexity.” Narrower claims hold: in certain taxa, during certain time windows, measured with a particular ruler, driven increases really do occur. Other data show only passive expansion. Many measures show no consistent trend at all. This brings us to the section’s real yield.

The null hypothesis should not be “complexity remains constant.” A more rigorous null is directionless random change, plus the known boundaries, plus branching and extinction. This null is harder to overturn, which is precisely its value: it first subtracts “the shape produced by the wall.”

If observations consistently depart from this model—an upward bias persists far from the boundary, for example, or the minimum rises—then cross-lineage driving mechanisms deserve testing. Sampling biases and omitted conditions also need checking. Rejecting one null model does not establish an inherent upward force in life.

While we are here, let us put down the correct reading of “the minimum stays constant while the maximum rises,” because it is so often read as something else.

Its most immediate meaning is that the attainable range of complexity has widened. Evolution has occupied regions of high complexity that did not previously exist, while the lowest end can persist for a very long time.

Simple forms have not been displaced by complex ones. Microbes remain widespread and ecologically indispensable. But plants dominate estimated global biomass measured as carbon; biomass cannot serve as evidence that the simplest forms constitute the largest share.

This distinction is no technical detail. It determines whether the very same rising curve should be read as “evolution tends to become more complex” or “evolution is taking a random walk beside a wall.”

VII Can What Was Lost Grow Back?

Begin with something that almost should not have happened. Stick insects are the insects that disguise themselves as twigs. A reconstruction of their phylogeny in 2003 concluded that their ancestors first lost their wings, and at least several later lineages grew flight-capable wings again.

Lose an entire set of organs, then grow it back long afterward: by the usual account, that should not be possible at all.

The usual account is called Dollo’s law, often quoted as “once a complex structure is lost, it can never be regained.” When the law was proposed in the late nineteenth century, its wording was much milder: an organism does not return to a former ancestral state in exactly the same way. “It does not retrace the same path” and “it can never go back” are claims of very different strength.

Another famous example comes from frogs. Many living frogs have upper-jaw teeth; almost all lack lower-jaw teeth. Yet an analysis proposed that lower-jaw teeth were lost in the common ancestor of frogs, and one lineage grew them back roughly 200 million years later. How long is 200 million years? Longer than the whole span from the appearance of dinosaurs to their exit.

Neither case is settled, and both could go wrong in the same way: ancestral-state reconstruction can itself manufacture false “regains.” An incorrect root state, unequal diversification rates, or an inappropriate transition model can reconstruct an independent gain as “loss followed by recovery.” Any claim of an exception to Dollo’s law must therefore run both kinds of model, one allowing regain and one forbidding it.

If we narrow the scale to genes, is there anything we can measure? Yes, and it gives us one of the most useful numbers in this essay.

An estimate published in 1994 suggested that an inactivated gene, or a developmental program that had been switched off, could retain a substantial chance of being restarted by a reversal mutation for roughly 500,000 to six million years. As neutral mutations accumulate, however, the probability that every necessary site remains intact falls rapidly. That is a model-dependent estimate involving mutation rates and functional constraints, not a universal expiry date for developmental programs.

“Irreversible,” then, is a matter of probability, not an absolute. It has a rough shelf life.

When a phenotype has only recently disappeared, many of the genes that build it may still be under selection because of pleiotropy. Or perhaps only a regulatory switch has been turned off. Reopening it at that stage is not difficult to imagine. The transplant experiment in Section III shows exactly this situation: the cavefish no longer has eyes, yet put a surface-form lens back into its eye cup and the eye can still grow.

After several million to tens of millions of years, however, if many dedicated structural genes have become pseudogenes or been deleted and developmental interactions have also broken down, recovery along exactly the same route becomes drastically more difficult.

As for “which is more common across the whole tree of life, loss or gain,” there is currently no single trustworthy ratio. Here is a neat instance of the rule from Section II. Impose the condition that “each gene family can be gained only once, and can only be lost thereafter,” a method called Dollo parsimony, and the reconstructed ancestral gene repertoire becomes extraordinarily large. The result is an extreme picture such as “tens of thousands of losses against fewer than a thousand gains.”

That ratio is first of all a product of the rule, not a biological conclusion. Use a model that allows regain, and the very same data can yield a completely different ratio.

Now return to the unfinished question from Section I: why are morphologically simple lineages so often put at the base of the tree?

Myxozoans provide the most complete historical case, overturning roughly a century of taxonomic tradition. The name “Mesozoa” is itself a fossil. The simple structures of orthonectids and dicyemids were once interpreted as primitive states intermediate between protists and animals. The idea is written into the name: “middle animals.”

An analysis in 2018 that controlled for long-branch attraction nested orthonectids within Annelida, the phylum of earthworms and ragworms. They still retain a microvillar cuticle, repeated circular muscles, and a dorsal ganglion, some of the few morphological clues left for placing a simplified lineage. Dicyemids consist of only a few dozen cells. Molecular data place them within Spiralia, while their genomes show marked reductions in homeobox transcription factors, a class of transmembrane signaling receptors, and gene families associated with nerves and muscles.

The correct description is “a bilaterian lineage underwent an immense simplification of its body plan,” not “a living fossil from before bilaterians appeared.”

That argument is still not over. It now sits at the very lowest fork of the animal tree.

The sponge-versus-ctenophore debate asks which branched off first from the rest of the animals: sponges or comb jellies. This is no dispute over names. It directly changes reconstructions of ancestral nerves, muscles, digestion, and cell types.

In 2023, evidence arrived in an unusual form. Comparisons of ancient gene linkages across animals and their unicellular relatives identified a set of chromosome-fusion and mixing patterns shared by sponges, cnidarians, and bilaterians, while ctenophores retained the ancestral state. This supported ctenophores as the sister group of all other animals. In principle, rare structural changes of this kind in the genome are less likely to produce coincidental resemblances than rapidly changing sequence sites.

This did not end the debate. Root placement depends on models, long branches, sampling, and gene selection. A 2025 paper supporting sponge-first was retracted in 2026 after analytical errors were identified. That corrects a particular analysis; the retraction alone cannot decide the entire controversy. The question remains unresolved in 2026.

What does this unresolved question change? If ctenophores branched off first, there are at least two explanations for nervous systems. The common ancestor may already have possessed one, with sponges and possibly placozoans subsequently undergoing extensive secondary loss. Alternatively, the nervous systems of ctenophores and of other animals arose independently. Position on the tree comes before conclusions about “who lost what.”

Placozoans themselves are a warning. Their bodies are flat, without conventional neurons, muscles, a gut, or obvious organs. Their simple morphology was precisely why early trees often drew them close to the root. Molecular and phylogenomic analyses, however, often place them after sponges and ctenophores.

More importantly, some conserved genes involved in establishing animal body axes still show organized expression in their simple bodies. The absence of obvious organs does not mean the absence of complex developmental regulation.

Two rules must therefore be set down together. First: morphological simplicity cannot be read directly as phylogenetic ancestrality. Second, in the opposite direction: not every instance of simplicity can be explained as loss.

The real product of this section is a usable checklist. To claim that a lineage underwent secondary simplification, at least one of four kinds of positive evidence is needed. The phylogeny nests it within an ancestral group that possessed the trait.

Homologous structures are retained during development. Structure-specific genes or pseudogenes remain. Or ancestral-state reconstruction strongly supports the ancestor’s having possessed the structure.

These four are exactly what the two cases in Section I relied on. Myxozoans use the first and third: their position on the phylogenetic tree, plus the two classes of nematocyst-specific genes. Rhizocephalans use the second: the adult reveals nothing, while the larva spells it out. Without any of the four, all we can say is “its present-day form is simple.”

VIII Applying the Criterion in Reverse, and the Remainder in the Null Hypothesis

We have reached the point of this essay. Episode X established a criterion for deciding whether a shift between levels had occurred. It consists of three questions. Have units that could originally replicate independently lost the capacity to reproduce on their own?

Has a new reproductive entity emerged as the unified bearer of fitness? Have conflicts among lower-level units been systematically suppressed? Only when all three conditions hold do we have a shift between levels.

This entire essay has been doing subtraction. A natural question follows: has any of that subtraction amounted to the reverse of this criterion?

Turn the three questions around. Have lower-level units regained the capacity for independent reproduction? Has a lower-level entity once again become the unified bearer of fitness? Has the suppression of conflict established at the higher level been systematically dismantled?

The result first, so we are not left hanging: of five candidates, only two score any points, and neither scores full marks. Now let us see where each one runs into trouble.

The first and second can be assessed together: eukaryotes that have lost mitochondria entirely, and the wholesale loss of an endosymbiont that was already integrated.

The loss of mitochondria is startling enough on its own. The textbook statement “eukaryotic cells all have mitochondria” has exceptions. A unicellular organism found in a rodent’s gut showed no trace of even mitochondrial remnants in its genome.

Neither case constitutes a reversal, for the same reason. Once the transition was complete, neither the mitochondrion nor the fully integrated symbiont was still a Darwinian individual capable of reproducing freely outside the host cell.

Deleting it outright does not restore its ancestors to a free-living existence. The host cell remains the bearer of reproduction and fitness. Nor has conflict suppression been lifted. The relevant lower-level unit has simply ceased to exist.

Verdict: zero out of three in both cases. The accurate description is “secondary loss of a highly integrated component.” Deleting an integrated partner is a different thing from separating two fused Darwinian individuals back into two independently reproducing entities.

The third candidate is loss of sexual reproduction, which is not rare in nature. Many dandelion lineages are apomictic: seeds develop without fertilization. Sexually reproducing dandelions exist too, however, so not every drifting roadside seed belongs in this category.

Giving up sex, however, does not reverse a hierarchy. No reproductive unit previously subordinated as a lower-level component thereby regains independent reproduction. The organism remains the bearer of fitness.

Verdict: zero out of three. The loss of sex is a reversal in a trait involved in an important evolutionary transition, not a reversal of a phase transition between levels of individuality.

The fourth candidate is a return from eusociality to solitary life. Here we begin to score points.

Sweat bees provide ready material. Within the same group, some species are eusocial and others solitary, and phylogenies show that the latter returned to solitude from the former.

Workers in many eusocial insects are not absolutely incapable of reproducing. Their reproduction is suppressed. When a lineage loses its worker caste and returns to solitary life, what happened may simply be “the worker caste disappeared,” rather than the same lower-level unit, having lost reproductive capacity, regaining it.

The second question can be answered “yes” in a solitary lineage. Yet individual insects were physiological individuals even in the eusocial ancestor. To demonstrate a reversal in the unit of selection, we must first show that the ancestral colony itself had become a higher-level Darwinian individual in the sense of a superorganism. The third condition should hold in theory, but it needs empirical evidence.

Verdict: roughly one or two out of three. Only a return to solitary life from a superorganism with obligate worker sterility in its ancestral organization and a clearly defined colony-level reproductive unit has a prospect of scoring three out of three.

The fifth candidate is secondary unicellularity. This is the closest category, and we can produce it in the laboratory: select yeast for a multicellular form that grows in clusters, then see whether it can go back.

If a clonal multicellular organism has genuinely completed a transition in individuality, its constituent cells can no longer complete their life cycles independently, and a descendant lineage restores unicellular reproduction, then the first condition holds. If the reproductive bottleneck shifts from a multicellular propagule back to a single cell, and the unit of selection really moves back toward the cellular level, the second holds too. The third is hardest. It requires evidence that the higher-level “ratchets” stabilizing multicellularity, such as a reproductive division of labor, bottlenecks, dependence on cell adhesion, or apoptosis, have been dismantled. Merely adding a unicellular stage to the life cycle is not enough.

Verdict: under strict criteria, usually one or two out of three. Some experimental systems could in principle score three out of three, but direct measurements of the unit of selection and conflict suppression are essential. Five candidates: three kept outside, two passing partway.

Nor is this result peculiar to the examples chosen by this series. By 2026, published work was explicitly discussing the same problem within a framework of “evolutionary transitions and reversals in individuality,” treating renewed internal conflict as one possible way of pulling a collective back from higher-level individuality. What it provides is a theoretical and modeling framework, not an established catalog from natural history. Too few genuine reversals have yet been assembled.

The great majority of net losses in this essay, including eyes, wings, digestive tracts, plastid genes, neural structures, and the entire body plans of parasites, alter traits or functions without changing the principal unit of selection. They are changes within the same level, however drastic they may be, even when the ancestor becomes unrecognizable. Why is genuine reversal so difficult? Because major transitions create their own ratchets.

A single-cell bottleneck is an accessible example. Many multicellular organisms start each generation from one cell, giving later cells a very recent shared genetic origin. That reduces internal variation and conflict; it does not guarantee their absence. Mutations can arise during development, and cells can still act against the interests of the whole.

Once that arrangement is established, the constituent cells become increasingly unable to leave the body and live on their own. Reproductive division of labor, mutual dependence, irreversible gene loss, and developmental control are all ratchets of the same kind.

They turn collective cooperation that was initially reversible into a higher-level individual that is difficult to take apart.

A genuine reversal therefore requires at least two opposing developments at once. Lower-level units must regain sufficient capacity for autonomous replication and survival in their environment. The higher-level structures of heredity, reproduction, and conflict suppression must fail, decoupling lower-level fitness from higher-level fitness once more. Merely deleting one higher-level component achieves neither.

This shows what a criterion is for: it must distinguish changes rather than admit every case. The three questions exclude three candidates and give two partial passes, keeping “phase transition” from becoming a synonym for “something changed.” Rejecting more cases does not itself make a criterion correct, however. The criterion too must face counterexamples and measurement.

Next, we need to clear away teleology, because this essay could all too easily be read as another version of it.

As Section II noted, parasites may account for roughly 40 percent of all animal species. Simplification is plainly not some rare condition of failure.

The stronger claim in the opposite direction, that “parasitism generally raises diversification rates,” does not hold either. Pair parasitic lineages with their free-living sister groups, and the overall conclusion is that parasites have no greater potential for speciation than their free-living relatives. The enormous number of living parasitic animal species comes chiefly from a few exceptionally successful radiations. Not every origin of parasitism sets off an explosion.

The myxozoan comparison is still useful. Myxozoans already have more than 2,180 described species, while their phylogenomic sister, Polypodium, is a tiny, essentially single-species lineage. This cannot prove that “simplification causes high diversification.” The two branches differ in age, opportunities for finding hosts, geographic dispersal, and research effort. It can refute only a stronger claim: that a lineage undergoing immense net loss must be an evolutionary dead end.

Simplification neither means failure nor guarantees a radiation. Outcomes are highly specific to the lineage. Finally, the addition this essay makes to the list of remainders.

The same pattern has appeared three times. The first was in Section VI: “complexity increases over time.” If the null hypothesis is “complexity remains constant,” directionless diffusion shaped solely by an impassable lower bound will be read as a driving force.

The second was in Section VII: “losses are dozens of times as numerous as gains.” If the null is Dollo parsimony, the ratio is first of all a product of the model. The third was also in Section VII: “this is an exception to Dollo’s law.” If the null does not include the possibility that an independent gain has been wrongly reconstructed as loss followed by recovery, a false regain will be taken for a real one.

The same thing all three times: the observed “direction” may reside not in the data, but in the unstated comparison model. Hence the fourteenth kind: the remainder is in the null hypothesis.

This must be recorded separately from Episode XIII’s “the remainder is in the explanation itself.” In that case, the explanatory framework brings its remainders with it. Change the framework, and you get a different set.

Here the remainder is not in the explanation. It is in the baseline against which the explanation is compared, and that baseline is usually not written down at all. One is in plain sight. The other is in the dark.

We can therefore establish a rule more durable than this essay’s particular subject: whenever a trend, a ratio, or an exception is reported, first ask what it is being compared with. If there is no answer, the conclusion is speaking about the unwritten model, not the world.

This essay has applied a phase-transition test in reverse, and the great majority of cases failed to qualify. It is the first time the criterion has been used in that direction.

Subtraction happens all the time, sometimes on an immense scale. Yet what it changes is almost always a trait, not a level. An animal reduced to a handful of cells is still a cnidarian.

A reduced body need not mean a smaller story. Next we follow the apes that once ranged across the Old World, and the uneven retreat that followed.