Adam, Eve, and Human Genetic Diversity

"None of the studies above set out to explicitly test the hypothesis that humans could have passed through a single-couple bottleneck." — Richard Buggs, Professor of Evolutionary Genomics, Queen Mary University of London, 2017¹

"For humans alive today, the median TMR4A is approximately 500,000 years ago" — Joshua Swamidass, 2022²

"This model is not only consistent with what is observed, it would appear to be very hard to falsify." — John Sanford, Robert Carter and colleagues on their own model, 2018³

Summary

Could every living human descend from one couple who lived a few thousand years ago?  The genetics is usually reported as a settled no.  It is not settled.  The reasons are not the ones creationists usually give.

What the standard case gets wrong.

1. Three of the four arguments in Adam and the Genome do not test the hypothesis. Linkage disequilibrium, PSMC population curves, and incomplete lineage sorting with chimpanzees and gorillas each measure something real. None of them measures whether the human line passed through a one-generation crash to two people. The person who says so most bluntly is not a creationist: Joshua Swamidass, a computational biologist at Washington University, writes that "LD, ILS, and Ne are irrelevant lines of evidence, that do not rule out a single-couple bottlenck [sic] in any relevant time scale".²
2. The fourth argument, allele counts per locus, is false at the level most people picture it. A single base can carry at most four alleles. A couple can carry four. Measured: of 57,190 polymorphic sites in 896 nuclear genes, 113 had three alleles, and the paper reports no site with four.⁴ That is 99.8% of sites with exactly two. A couple could have carried every per-site allele count in the human genome.
3. Venema's strongest argument carried no citation. Richard Buggs asked twice for a reference. He was pointed to an online HLA primer with no bottleneck model in it, and said so in print.⁵
4. Venema misreported the linkage-disequilibrium paper he cited. Tenesa and colleagues open their discussion with "Overall, the estimates of Ne appear to be much lower than the usually quoted value of 10,000 (Takahata 1993)."⁶ Adam and the Genome reports the paper as confirming 10,000. It gets there by adding an African estimate of about 7,000 to a non-African estimate of about 3,000.⁷ The non-African populations descend from the African one, so their variation is a subset of it. Adding a subset to its superset is not a valid sum.
5. BioLogos deleted the strongest version of its own claim and did not say so. The 2014 article "Adam, Eve and Human Population Genetics" said "There is in fact no point in our evolutionary history over the last 18 million years or more where our lineage was reduced below 10,000 individuals." The version now online says "As far as anyone can tell, the genetic data in fact rules out a single couple for at least half a million years."⁸ That is a retreat by a factor of 36 in the time depth of the claim. The whole SNP-counting section went with it.

What survives, and it is serious.

6. HLA-DRB1 holds more deeply different lineages than four founding chromosomes can supply. Counted here from the IPD-IMGT/HLA allele list, release 3.65.0: 4,023 named DRB1 alleles, falling into 13 first-field lineages.⁹ The 4,023 is not an argument. The immunogeneticists who measured it found that over 90% of DRB1 alleles arose after the split.¹⁰ The 13 is not quite an argument either. The same group's 2007 full-length study assigns four of those lineages, 08, 11, 13 and 14, to a radiation from one proto-*03 ancestor within the last 5 million years.¹¹ That leaves nine as a ceiling on the pre-split count. A couple carries four. Measured directly, with no clock and no chimpanzee, the introns of 680 complete DRB1 gene sequences fall into five to nine families more than 1% apart in 16 of 18 stretches of 500 sites, even when each family may be a patchwork of the others.¹² Six thousand years of mutation separates two lineages by about 0.0005%.
7. The median time back to four surviving lineages at a human locus is 495,000 years. This statistic, TMR4A, is the only published analysis that asks the four-allele question directly and genome-wide. Swamidass computes a median of 495,000 years. Half the genome is older than that.² Venema is quoted by Swamidass as calling it "a better way of looking at this overall".² The same author's later average across three datasets is lower, 358,000 years.¹³ For the ten lineages Noah's family could carry, the median is about 180,000 years.² Robert Carter of Creation Ministries International has criticised the method in print, and one of his points is partly right. No creationist has computed a TMR4A or offered an alternative number.
8. No fitted single-couple model lands more recently than 500,000 years. Ola Hössjer, Professor of Mathematical Statistics at Stockholm University, and Ann Gauger fitted a single-couple model to real African allele frequencies and to real linkage-disequilibrium decay curves. It fits at about 500,000 years ago with pre-loaded diversity.¹⁴ A more recent date is a conjecture about three model extensions they did not try. The only fit at a young-earth date is Sanford and colleagues' 2018 simulation: 200 generations at a census of 1,000, scored by a mean squared error that its authors say "is not a significance test."³
9. The creationist answer to the four-allele ceiling comes in two versions, and only one of them is science. Hössjer keeps the ceiling. Two people carry four chromosome copies, and the amount of created difference between them is a fitted parameter.¹⁵ Sanford, Carter and colleagues drop it. If each of Eve's egg cells was separately created, "there is almost no limit to the number of variant alleles and linkage blocks that could have existed in Eve's ovaries."³ They then write of that model, "This model is not only consistent with what is observed, it would appear to be very hard to falsify."³ That self-assessment overlooks the Flood. However Eve's eggs were made, Noah's family carried at most ten copies of each chromosome, so after the Flood every locus must trace back to at most ten families. That makes even this version testable, and it meets a median time back to ten lineages of about 180,000 years.²
10. The young-earth model's real problem is not allele counts. It is the number of deeply different families. Computed for this article: start from eight people and grow 5% a generation. In 207 generations, about 195 new mutations arise in any given 2 kb window across the whole population. New alleles are easy. Deep divergence is not: 6,000 years of mutation separates two lineages by about 0.0005% of sites. Created heterozygosity supplies deep divergence among four founding chromosomes at no cost, and that is the model's real strength. It cannot supply a fifth deep family. DRB1 and TMR4A both count deep families, and both find more than four.

The bottom line.  The claim that human genetic diversity disproves a single couple with heliocentric certainty is over-stated, and its own authors have walked it back.  The claim that a single couple 6,000 years ago is genetically unproblematic is also false.  Two measurements stand in the way:  DRB1's lineage count and TMR4A.  No published calculation on the young-earth side has met either.

The four arguments, and what each one tests

Dennis Venema's 2017 book Adam and the Genome put this question in front of a general audience.  Chapter 3 gives four lines of evidence that the human line never fell below about 10,000 individuals.  The book opens with the Tasmanian devil as a picture of a two-individual founding:  "The implications are clear: Tasmanian devils experienced a severe bottleneck in the distant past, but humans did not."⁷

Three fail, each in a different way.  One survives, at the right level.

Argument 2: linkage disequilibrium

Alleles at nearby sites travel together.  Recombination breaks that association up, at a rate that depends on how many ancestors there were.  Albert Tenesa and colleagues measured that decay in HapMap data.  Their answer was smaller than the number Venema reports.  Their abstract:  "For CEU, JPT, and HCB, the estimate of effective population size, adjusted for SNP ascertainment bias, was ∼3100, whereas the estimate for the YRI was ∼7500".⁶  Their discussion opens by saying these are lower than the usual 10,000.⁶

Three things are wrong with using this as a disproof of a single couple.

The addition is invalid.  Venema adds 7,000 and 3,000 to reach 10,000.  Buggs, an evolutionary geneticist and not a creationist, wrote:  "It seems to me that the paper gives no warrant for Venema's addition of effective population sizes above and below the Sahara".¹

The method's own error bar is large.  Swapping one recombination-rate estimator for another moved every estimate by an average of −27%, with a range from +33% to −45%, giving a mean Ne of 1901.⁶  A quantity that swings across a 78-point range on one methodological choice cannot support a claim of near-certainty.  There is also a circularity.  The recombination rates themselves are estimated from linkage disequilibrium.¹

The window excludes the event.  The method reaches back 10,000 to 200,000 years and no further.  By construction it assumes a population near equilibrium.  A one-generation crash inside a fast-growing population is the one case it does not model.

The strongest reply to this argument is not a creationist objection.  It is a fitted model.  Hössjer and Gauger simulated a single founding couple forward in time.  They compared the resulting r² and D′ decay curves against real African data.  Their Figure 10 caption:  "The two very different scenarios give similar results, both in good agreement with data."¹⁴  Linkage disequilibrium does not rule out a single couple.  It never did.

Argument 3: PSMC population curves

Heng Li and Richard Durbin's method reads local coalescence times off the density of heterozygous sites along one diploid genome.  It turns them into a population-size curve.  It is a real measurement.  By the authors' own statement it is blind to the kind of event in question.

Two sentences from their paper decide this.  On resolution:  "More recently than 20kya or more anciently than 3Mya few recombination events are left in the present sequence, which reduces the power of PSMC".¹⁶  On sharp events:  "the instantaneous reduction from 12,000 to 1,200 at 100kya in the simulation was spread over the preceding 50,000 years in the PSMC reconstruction."¹⁶

A tenfold crash gets smeared across fifty millennia.  A one-generation crash to two individuals is far sharper and shorter than the case they tested.  And the method has no power more recently than 20,000 years ago.  It says nothing about 4,500 years ago in either direction.

Jack Elliott-Higgins and Swamidass tested this directly.  They simulated bottlenecks of 2, 4 and 8 individuals at various dates and asked whether the standard methods could see them.  "No method we tested accurately detected bottlenecks that last just a few generations."¹³  A bottleneck of two at 75,000 years ago was invisible to MSMC2.  Their conclusion:  "even a single couple bottleneck more ancient than 450 thousand years ago cannot be ruled out by estimates of Ne alone."¹³

That preprint has limits.  It is not peer reviewed.  The second author is a party to the dispute.  Their own Figure 7 reports that recent bottlenecks are more detectable than ancient ones.  It is a good citation for "these methods are blind to brief crashes."  It is not a citation for the claim that a Flood bottleneck would be undetectable.

Argument 4: incomplete lineage sorting

Where gene trees for human, chimpanzee and gorilla disagree with the species tree, the disagreement rate estimates the size of the population that existed before the split.  Venema's 2014 version put it at about 62,000 individuals from a 30% discordance rate.⁸

That is a real measurement of a real quantity.  It is not a measurement of the human lineage after the split.  Buggs's one-sentence answer has never been rebutted:  "the occurrence of ILS does not require maintenance of large populations sizes in every lineage after speciation and so does not exclude a bottleneck in the exclusively human lineage".¹

One form of the argument does constrain the human line:  trans-species polymorphism, where a human allele is more closely related to a chimpanzee allele than to another human allele.  Such a lineage must have survived in the human line since before the split.  A crash to two people would have to preserve it.  Ellen Leffler and colleagues scanned the genomes for this and found it rare:  "we identified 125 regions in which the same haplotypes are segregating in the two species, all but two of which are non-coding.  In six cases, there is evidence for an ancestral polymorphism that persisted to the present in humans and chimpanzees."¹⁷  Swamidass reads the same body of work as permitting "a minimum population limit of 5 to 10 individuals (not thousands) going back 6 million years."²

Five to ten is not four.  But it is nowhere near ten thousand, and it is the mainstream side that says so.

Swamidass then discounts his own number:  "However, HLA genes also show clear evidence of convergent evolution. … For this reason, TSV-based estimates of minimum population size are lower confidence and disputed."²  Convergent evolution means two species can reach similar sequences without inheriting them from a common ancestor.  If that happens at the MHC, some apparent trans-species sharing is not shared history.  The DRB1 argument later in this article is then weaker than a raw lineage count suggests.  That cuts for the young-earth side, and it comes from the mainstream.

The deletion

In June 2021 BioLogos removed "Adam, Eve and Human Population Genetics".  It posted a new article at a different URL, back-dated to the original 2014 date.  Peaceful Science republished the old text under Fair Use with the changes marked.  The diff is mechanical and checkable.  The framing around it, by a party to the dispute, is not.  What follows is the diff.

Deleted:  "The smallest population our lineage was reduced to in the last 15 million years or more, then, was to the bottleneck of 10,000 individuals once our lineage parted ways with the chimpanzee lineage."⁸

Deleted and replaced:  "There is in fact no point in our evolutionary history over the last 18 million years or more where our lineage was reduced below 10,000 individuals" became "As far as anyone can tell, the genetic data in fact rules out a single couple for at least half a million years."⁸

Deleted entirely, the SNP-counting section, including:  "SNP diversity in humans is far too large to result from one ancestral couple at any time in the last 200,000 years – we descend from a population."⁸

Eighteen million years to half a million years is a factor of 36.  The revised claim is defensible.  The original was not.  The revision was made silently.  The argument a skeptic is most likely to have met is the deleted one.

Where the allele count still bites: HLA-DRB1

Argument 1 is the one with force.  It has to be stated at the right level, or it is wrong.

The wrong level.  "There are thousands of HLA alleles and a couple can carry four."  Ann Gauger states the objection in this form and calls it "the strongest case from population genetics against two first parents".¹⁸  Lars Johan Erkell, a biologist at the University of Gothenburg, puts it to Hössjer the same way:  "in this model there can be a maximum of four gene variants — alleles — of a certain gene.  However, there are gene families with thousands of alleles.  Where do they come from?"¹⁵

The immunogeneticists themselves answered that version in 1998.  Francisco Ayala had dated DRB1 alleles from exon 2, the antigen-binding region.  That region is under strong balancing selection and is reshuffled by gene conversion.  His 1995 Science paper counted 32 human DRB1 lineages already present 6 million years ago and concluded that "no fewer than 16 individuals could have lived at any given time over that long span."¹⁹  His PNAS paper the year before had counted 21 at the same date, from 58 alleles rather than 59.  Neither paper mentions the other's figure.²⁰  That 1994 paper is where the number people quote comes from:  "human ancestral populations could never have been smaller than two or three thousand individuals at any time over the last several million years."²⁰  That figure is simulation output, not a lineage count.  His own lineage arithmetic gives a minimum of 29 people.²⁰  Tomas Bergström and colleagues dated the introns instead.  Introns are under no such selection and are not reshuffled.  Their result:  "Alleles within such a lineage, however, appear to be, on average, 250,000 years old, implying that the vast majority (greater than 90%) of the more than 135 contemporary human DRB1 alleles have been generated after the separation of Homo and Pan."¹⁰

That is not a creationist claim.  Gaby Doxiadis and colleagues confirmed it independently, in the open literature, in Molecular Biology and Evolution:  "Subsequent intron analyses, indeed, showed that the majority of contemporary HLA-DRB1 alleles have been generated during the past 250,000 years, whereas many DRB lineages and loci predate the divergence of the hominoid species".²¹

So the raw allele count is not evidence about founders.  Note who this cuts against.  Adam and the Genome appeared in 2017 and still leans on an intuition Bergström had undercut nineteen years earlier.

The right level.  Count lineages, not alleles.  Computed for this article from the official IPD-IMGT/HLA allele list, release 3.65.0 dated 14 July 2026:

Use DRB1's 13.  Treat the class I numbers as rough.  At DRB1 and DQB1 the first field of the allele name tracks allelic lineage.  That is what Bergström and Doxiadis mean by the word.  At HLA-A, -B and -C the first field originally tracked serological antigen groups.  Those are not guaranteed to be single branches of a tree.  DPB1 was excluded because its first field is a running number, not a lineage label.

Thirteen is not thirteen pre-split lineages.  The honest number is lower.  The same Uppsala group that dated the introns went back with full-length sequences in 2007.  Jenny von Salomé, Ulf Gyllensten and Tomas Bergström report that an ancestral "proto HLA-DRB103 lineage" split into the human-specific lineages 08, 11, 13 and *14 within the last 5 million years.¹¹  So four of the thirteen first-field lineages are a radiation from one ancestor, not four independent ancient lineages.  That leaves nine.

Nine is a ceiling, not a measurement.  Bergström's abstract says "most" of the lineages diverged before the Homo/Pan separation, not all.  Von Salomé's window of "the last 5 million years" straddles the 4 to 7 million-year split date rather than sitting cleanly after it.  The true pre-split count is at most nine.  Nobody has published it.

At most nine lineages against a founding ceiling of four.  That is the argument.  It is smaller than the thirteen a raw count gives, and it is still more than four.

The count can be measured directly, with no clock and no chimpanzee.  A founding couple carries four copies of chromosome 6.  If every living human descends from them a few thousand years ago, then in any short stretch of DRB1 intron, today's sequences must fall into at most four families.  Members of one family can differ only by the mutations a few thousand years add, about 0.0005% of sites.  Recombination and gene conversion can splice the four families into patchworks.  They cannot make a fifth family.

This was computed for this article from the official IPD-IMGT/HLA genomic alignment, release 3.66.0 of 14 October 2026, which holds 680 complete DRB1 gene sequences.²²  The 9,042 intron sites that nearly all sequences share were cut into 18 windows of 500 sites.  In each window, sequences differing at five sites or fewer were grouped together, so separate groups differ at more than 1% of sites.  The test then asked how few groups are needed to build all the others as patchworks, allowing two crossovers and two stray mutations in every window.¹²

Sixteen of the eighteen windows need more than four.  A typical window holds six families:  03, 08, 11, 12, 13 and 14 together;  15 and 16;  04;  07 and 09;  01;  and 10.  Across all 9,042 shared sites, 04, 07 and 09 differ from every other lineage at about 10 to 11% of sites, and 01 differs from 10 at 7%.  Even two of the closest lineages, 11 and 13, differ at 0.22%, over four hundred times what 6,000 years of mutation supplies.

Three limits apply.  The alignment lists named alleles, not how common each one is.  The thresholds are a choice, but tightening them only raises the counts:  at two differences per window, every window needs five to nine.  And the test does not model long gene-conversion tracts copied in from the neighbouring DRB genes that the founding chromosomes would also carry.  That is the one escape left inside the four-chromosome model, and it would have to rewrite intron sequence along most of the gene.

So the four-chromosome model fails at DRB1, and the designed-gametes model does not.  Under designed gametes, described below, the binding limit is the Flood, where Noah's family carried at most ten copies.  The worst DRB1 window needs nine.  DRB1 cannot separate that model from the data.  TMR10A, in the next section, is the test that can.

The creationist side has named the test and has not run it.  Robert Carter, John Sanford and Nathaniel Jeanson answer the allele count, not the lineage count.  Hössjer and Gauger do address the MHC, and precisely:  "In order to test whether or not the first pair hypothesis is correct, at any position along the HLA complex, all lineages should coalesce to at most four ancestral lineages", and they say "previous research indicates perhaps as few as four ancestral lineages" at DRB1.¹⁴  They also flag, against themselves, that "there is evidence that HLA genes are subject to convergent evolution".  So the test is stated.  Before this article, nobody had run it.  Their four rests on Titus-Trachtenberg 1994, Bergström 1998, von Salomé 2007 and Gauger's own 2012 book.  None of the three papers states four.  The number traces to the book.

One number often repeated on the creationist side could not be verified.  Gauger writes that Bergström's intron study "concluded that only seven versions of the gene existed in the ancestral population" of humans and chimpanzees.¹⁸  Bergström et al. 1998's full text is closed at the publisher and at every open-access route.  Only its abstract could be obtained, and the abstract does not contain the number.  The 1996 Science technical comment that Gauger, Doxiadis and Swamidass all cite is also closed, and PubMed carries no abstract for it.  Anyone repeating "seven ancestral DRB1 alleles" should say it is unverified.  Seven is in any case more than four.

One claim in the same chapter checks out, with named exceptions.  Gauger writes that "It is clear that the intron sequences group according to species, whereas exon 2 sequences show no species-dependent relationships."¹⁸  She cites Doxiadis and colleagues 2008.  Its abstract says "most of the primate DRB alleles investigated represent relatively young entities, possessing species-unique sequences", and that this "seems to contradict the current view that the highly similar peptide binding motifs" encoded by exon 2 "predate primate speciation".²³  Her sentence is a fair summary of the paper she cites.  The same group's 2012 study names the exceptions:  "apart from the pseudogenes, only DRB5 genes are shared by OWM and HOM".²¹  So the rule holds for DRB1, with the pseudogenes and DRB5 outside it.

TMR4A, the argument with no answer

Effective population size is a harmonic mean over thousands of generations.  It cannot reveal a single bad generation.  Swamidass says so:  "Ne is, instead, a harmonic average over a very large time window".²  So he built a statistic that asks the question directly.

For each haplotype block in the genome, look at the gene tree and ask how long ago the four most recent surviving lineages coalesce.  Call it the time to the most recent four ancestors, TMR4A.  If all humans descend from one couple, then at every locus all present variation must trace back into at most four founding copies at the moment of the bottleneck.  So the bottleneck must be older than the TMR4A at nearly every locus, not at the average one alone.

The measured median across the autosome is 495,000 years.  Figure 3:  "a single-couple bottleneck is ruled out before 495,000 years ago … The black line identifies the median TMR4A, at 495,000 years ago."²  Median means half the genome is older than that.

Venema accepted the framework.  Swamidass quotes him from a BioLogos forum thread rather than from anything Venema published:  "TMR4A is a better way of looking at this overall — it more directly addresses what we're really interested in."²  Stephen Schaffner independently ran allele-frequency-spectrum simulations and got a compatible answer.  Buggs summarised, "A bottleneck had to be at least 500,000 years ago before anything approaching a good fit was simulated."²⁴  Buggs's closing judgement on the whole exchange is the honest one:

To my mind, the question has now moved on from "Is an 'Adam and Eve' bottleneck inconsistent with human genetic diversity data?" to "At what timescale could an 'Adam and Eve' bottleneck be consistent with human genetic diversity data?"²⁴

This argument has two weaknesses.  First, TMR4A has never been refereed.  It lives in a Peaceful Science Print and a Zenodo deposit, published by a party to the dispute in his own venue.  The single most load-bearing number against a recent single couple has not been through peer review.  Swamidass himself reports that he could not reproduce Schaffner's separate result:  "I was unable to replicate Schaffner's results using simulation software commonly used in the field."²  Second, TMR4A is a molecular-clock statistic.  It converts branch lengths into years using a mutation rate and a generation time.  A young-earth model that rejects the clock is not obliged to accept the conversion.

That second point is weaker than it looks.  Rejecting the clock does not make the branch lengths go away.  The lineages are still separated by the observed number of substitutions.  Something has to have put them there.  That is the arithmetic in the next section.

A creationist has answered in print, and part of the answer is right.  Robert Carter of Creation Ministries International replied in 2021.  He describes the method as working "Given a constant mutation rate, a constant recombination rate, a constant generation time, and human population of at least 10,000 individuals", notes that Swamidass's ground rules excluded a specially created couple, and concludes, "They were not testing the biblical model in any sense."²⁵  Swamidass answers that "No simulations of population history are used, in any way, to infer the ARG and estimate the median TMR4A date," and that the young-earth criticism "indicates misunderstanding of the analysis, not any errors."²

Carter is right about the population assumption, though not in the way he puts it.  No population history is simulated.  But ARGweaver, the program behind the 495,000-year figure, dates its trees against a prior that assumes one constant population size, "N=11,534" for the human genomes.²⁶  Swamidass's own coauthored test of these methods measured what that does after a recent bottleneck:  "ARG inference algorithms, however, systematically overestimated the coalescence times in simulations with recent bottlenecks (Figure 7A), though ARGweaver is least affected."¹³  Read from their Figure 7, a two-person bottleneck simulated 10,000 years ago comes out about twice as old with ARGweaver, and older still with the two newer methods.  Swamidass's 2022 review cites the same preprint as "empirically validating TMR4A" and does not mention the bias.²

Carter is wrong about created diversity, and the size of the bias matters.  TMR4A counts back to four lineages, not one, precisely so that Adam and Eve may carry four different copies of every chromosome.  Only the designed-gametes version escapes it, and the Flood brings that version back into range through TMR10A.  A twofold overestimate does not close a gap of 60 to 80 times.

The same preprint also gives a lower number than the one usually quoted.  Averaged across three datasets and methods, it reports "TMR3A and TMR4A at 422 kya and 358 kya, respectively".¹³  And a young-earth model faces a second count.  Noah's family could carry at most ten lineages, and "the median TMR10A is approximately 180,000 years ago."²

In young-earth terms, these numbers set the size of the escape a faster clock would need.  The trees are already dated with a per-generation mutation rate of 1.26 × 10⁻⁸, close to the 1.25 × 10⁻⁸ this article uses.²⁶  To bring the median TMR4A down to 6,000 years, mutation would have to have run about 60 to 80 times faster, on average, across all of human history and the whole genome.  To bring TMR10A down to a Flood 4,500 years ago, about 40 times faster.²⁷  Carter lists candidate causes, among them radiation "perhaps associated with a burst that occurred during the Flood" and what he calls Patriarchal Drive, but in that reply gives no number for any of them.²⁵

Designed heterozygosity: two versions, one falsifiable

The creationist answer to the four-allele ceiling is that Adam and Eve were not created identical.  If their four copies of each chromosome differed from each other, a couple starts with real diversity rather than none.  Hössjer states the mechanism cleanly:  created founder diversity means "All differences between his reference chromosome and the other three founder chromosomes can be thought of as a very large number of mutations, all of which occurred in one generation."²⁸

This idea splits into two very different models.  They should never be cited as one.

Version A keeps the ceiling.  Hössjer's reply to Erkell is explicit:  "in the first founder generation we get four genetically different variants of each non-sex chromosome.  The degree of genetic dissimilarity between these variants, i.e. the degree of created variation, is a parameter that can be varied in the model."¹⁵  A parameter that can be varied can be fitted, and a fit can fail.  Hössjer stakes falsifiability on haplotype blocks.  Every human chromosome should be a mosaic of four founder blocks.

Sanford's Designed Alleles Model does the same thing:  starting frequencies restricted to 25%, 50% or 75%, "since there were only four sets of chromosomes in Eden".³  That is a real model with a real ceiling.

Version B abandons the ceiling.  Sanford's Designed Gametes Model reasons that Eve was created rather than born.  Her egg cells never formed by ordinary meiosis, so each could be separately designed.  "Therefore, there is almost no limit to the number of variant alleles and linkage blocks that could have existed in Eve's ovaries."³  Carter says the same in Journal of Creation:  "there is really no limit to the amount of diversity that could have been front-loaded into Adam and Eve."²⁹

Once that move is made, no allele count at any locus can ever count against the model.  Thirteen DRB1 lineages, thirty, three hundred:  all are accommodated in advance.  The authors say so:  "This model is not only consistent with what is observed, it would appear to be very hard to falsify."³

Take that sentence seriously.  A model that cannot fail is not an answer to an observation.  A skeptic handed the paper as a refutation will find that sentence, discount everything else on the page, and be right to.

The Flood puts a ceiling back.  Designed gametes free Adam and Eve's children, not Noah's.  Noah, his wife and the three wives of his sons carried at most ten copies of each chromosome, so after the Flood every locus must trace back to at most ten families.  Swamidass makes the same point and supplies the test:  "the median TMR10A is approximately 180,000 years ago."²  So the designed-gametes model is testable after all, against a ceiling of ten rather than four.  It passes DRB1, whose worst window needs nine families.  It meets TMR10A, 40 times older than a Flood 4,500 years ago on the same clock as TMR4A.

What the fitted models actually found

Hössjer and Gauger 2019 is the strongest technical work on this side.  Its numbers are often misreported, so here they are exactly.  They fitted two scenarios:  a couple 2 million years ago with zero heterozygosity, and "A single couple 25,000 generations ago (about 500kya) having primordial heterozygosity of 0.012, grows rapidly to a population of 16,000 people, then holds steady."¹⁴

Primordial heterozygosity of 0.012 means about 1.2% of sites differ between the founders' chromosome copies.  Measured human nucleotide diversity is about 0.1%.  The founders are about twelve times more diverse than any living human.  That is what "designed heterozygosity" has to mean as a number.  It is a coherent thing to claim.

They did not fit anything more recent.  Their own words:  "It is very likely possible to extend our first couple origin model of humanity in order to find more complex scenarios in which the data is compatible with a more recent origin, but we deliberately circumvented fine tuning of the model to data."¹⁴  The 100,000-year claim is a conjecture about three model extensions they list and did not try.  Hössjer repeated the position to Erkell.  The good fit holds "if Adam and Eve lived several hundred thousand years ago".  He expects a larger model can probably lower the date.¹⁵

A young-earth article that cites Hössjer and Gauger as support for a 6,000-year Adam misrepresents them.  A skeptic can check that in ninety seconds.

Three technical problems with the creationist simulations

Sanford's simulations run at a census of about 1,000 individuals.  The result depends on drift spreading the initial spikes at 25%, 50% and 75% out into the observed mass of low-frequency variants.  Drift at N = 1,000 is far stronger than drift in a real post-Flood population growing into the millions.  The paper flags the caveat and does not answer it.  The comparison is between normalised curve shapes.  Its only fit statistic is a mean squared error, which the authors say "is not a significance test", and there is no confidence interval or null distribution.³  The table is still instructive, and lower is better.  The young-earth model that keeps the four-allele ceiling scores 1.16 × 10⁻³.  The designed-gametes model scores 0.84.  A couple drawn from an old population 200 generations before the end scores 0.70, the best fit in the paper.  All three beat the authors' deep-time model at 5.75, which also ran at a census of 1,000.

The haplotype-block prediction has no citation at its load-bearing number.  The number turns out to be close.  Hössjer writes "Each block comes in a few variants, four for many parts of the genome" and gives no reference for it.²⁸  Stacey Gabriel and colleagues measured it in 2002 across 51 autosomal regions:  "an average of 5.0 common haplotypes" in Yoruban and African-American samples, "4.2 common haplotypes" in European samples, and "3.5" in Asian samples.³⁰  So four is a fair reading of the middle of the range.  It is also where the model gets into trouble.  The African average is above four.  Pooling populations raises it further, from "3.1 – 4.9 haplotypes with a frequency >5%" within any one population to 5.3 in the union.³⁰  Four founder chromosomes cannot supply 5.3 common haplotypes per block.  The gap is small and real.  It has to be closed by mutation, gene conversion or recombination inside blocks.  Nobody has published the distribution, which is the number that would decide it.

The block structure is not evidence of founders anyway.  Jeffrey Wall and Jonathan Pritchard tested the block model directly:  "real data show a partial fit to the haplotype block model;  some regions conform quite well, whereas others do not."³¹  Simulating recombination hotspots reproduced the observed patterns without any founder interpretation.  Blocks are weak evidence in either direction.

Jeanson and Lisle's created-heterozygosity paper does make a numerical haplotype-block prediction, and deserves credit for it.  Created alleles plus 6,000 years of recombination and gene conversion "plausibly explains the number of haplotype blocks present in the human population today".³²  But the number of blocks is not the quantity in dispute.  Recombination and gene conversion shuffle existing variation into new combinations.  They make new blocks.  They do not make two blocks look 495,000 years apart.

What 6,000 years of mutation can and cannot build

Two computations were run for this article, using the audited population-genetics modules already in this project.  Both scripts sit beside this page under research/.

Computation 1:  a founding pair keeps most of its heterozygosity and loses most of its alleles.  Running the Nei, Maruyama and Chakraborty 1975 setup forward, the minimum heterozygosity after a founding pair is 64.5% of the original at a growth rate of r = 1.0, 50.0% at r = 0.5, and 5.9% at r = 0.1.  So a fast-growing pair keeps about two-thirds of the heterozygosity of the population it came from.  That is the number the creationist side quotes.

Now the other half of the same calculation.  Take a source as diverse as the most polymorphic HLA loci, with θ = 40 and heterozygosity 0.976.  Under Ewens's sampling formula, a sample of four gene copies from that source holds an expected 3.86 distinct alleles out of the source's 489.  Across the range θ = 0.16 to 10, which spans ordinary loci up to DRB1's heterozygosity of about 0.91, a pair carries 1.3 to 3.5 alleles from a source holding 3.6 to 136.  A pair's expected allele count never reaches 4 and can never exceed it.

Both halves must always be printed together.  Quoting the heterozygosity number alone makes a bottleneck look costless.  Quoting the allele number alone makes it look fatal.  Both sides of this debate have quoted one and not the other.

Computation 2:  new alleles are cheap, divergence is not.  Take eight founders and 207 generations of 29 years each.  That is 6,000 years, which is generous to the young-earth model.  Add growth of 5% a generation, capped at 8 billion people.  About 7.8 million gene copies are born.  In any given 2 kb window, about 195 new mutations arise somewhere in the whole population.  A young-earth model has no difficulty producing many new, rare alleles.  This is why the SNP-level version of the allele-count argument fails.  Jeanson and Lisle said so in 2016:  "Since only four DNA base-pairs exist, virtually every possible genomic position allele … could have been present at the time of the Flood, if the individuals were heterozygous."³²

Now the divergence.  At a human mutation rate of 1.25 × 10⁻⁸ per site per generation and 29 years per generation, two lineages separated for a given time are expected to differ at this fraction of sites:

In a 2 kb DRB1 intron, that is 0.010 expected differences across 6,000 years, against 10.3 across 6 million.  Six thousand years of mutation delivers one one-thousandth of the divergence between lineages split at the human-chimpanzee divergence.

This is the whole problem in one line.  The young-earth model can make plenty of alleles.  It cannot make alleles that are dozens of substitutions apart.  It cannot make deeply divergent alleles that are also common.  Both exist and have to come from somewhere.  Created diversity is the young-earth source, and in Hössjer's four-chromosome version it is constrained.  It can supply at most four deeply different families at any stretch of DNA that recombination has not broken up.  That is a prediction an observation can check, and DRB1 and TMR4A are the checks.  The designed-gametes version raises the limit to ten at the Flood, and TMR10A checks that.

What Swamidass's own answer costs

Swamidass has a model that puts Adam and Eve 6,000 years ago and contradicts no genetics.  It is often cited loosely, so here it is exactly.

Genealogical ancestry is not genetic ancestry.  Go back far enough and everyone shares genealogical ancestors.  Most of those ancestors contribute no DNA to any given descendant.  So:  "Adam and Eve, ancestors of us all, could have been de novo created in the Middle East.  As long as their offspring eventually interbred with people outside the Garden, it is possible as recently as six thousand years ago."³³

Read the condition.  The model buys a recent Adam by giving up sole genetic ancestry.  There were other people.  The species' genetic diversity comes from them.  Adam and Eve are, in his phrase, "genetic ghosts, who leave us no identifiable DNA."³³

That is a real option and some Christians take it.  It is not the young-earth position.  A model in which every human descends genetically from one couple, and again from eight people at the Flood, gets nothing from it.

Objections and responses

"Human effective population size has never been below 10,000." Effective size is a harmonic mean across thousands of generations.  One bad generation barely moves it.  Every party who has looked closely concedes the point, including Buggs and Swamidass.  BioLogos deleted its 2014 version of the claim in 2021.  The objection should be retired.

"There are thousands of HLA alleles and a couple carries four." Wrong level.  Bergström's intron dating found that more than 90% of the 135-plus DRB1 alleles known in 1998 arose after the human-chimpanzee split, by gene conversion and point mutation.  Doxiadis confirmed it in 2012 in the open literature.  The count of 4,023 is not a founder count.  Restate the objection as thirteen lineages against four and it becomes the strongest thing on the skeptic's side.

"The Tasmanian devil shows what a two-individual founding looks like." Devil populations were small for a long time.  That is what destroys heterozygosity.  A single generation at two, followed by fast growth, keeps about two-thirds of heterozygosity.  That was computed above and first shown by Masatoshi Nei, Takeo Maruyama and Ranajit Chakraborty in 1975³⁴.  The right prediction for a brief crash is normal heterozygosity with fewer alleles, a far subtler signature than the devil's.

"Hössjer and Gauger proved a single couple is possible." This is a creationist objection to this article's verdict.  It fails.  Their fitted scenarios are 2 million and 500,000 years ago.  The recent date is a conjecture about three extensions they declined to try.  Citing them for a young Adam misreports a paper that is easy to check.

"You quote Buggs against Venema and stop before his conclusion." The article does not stop there.  Buggs dismantled the way three of the four arguments were used.  Then in April 2018 he reported the analyses that put the boundary at 500,000 years.  Quoting his 2017 criticisms without his 2018 conclusion is quote-mining.  He says so himself.

"Created diversity answers the lineage count." Only in the designed-gametes version, and only at DRB1.  Hössjer's version keeps the four-allele ceiling.  He and Gauger name the right test, that every position in the HLA complex should coalesce to at most four lineages.  They assert that prior research already indicates four.  It does not.  The number comes from Gauger's own 2012 book, not from the three papers they cite beside it, Titus-Trachtenberg 1994, Bergström 1998 and von Salomé 2007.  The intron count computed for this article finds five to nine families more than 1% apart in 16 of 18 windows, even when each may be a patchwork of the others.  Hössjer's version fails there.  Sanford's designed-gametes version absorbs any number at Adam but only ten at the Flood.  DRB1's worst window needs nine, so that version survives DRB1 and meets TMR10A instead.

"TMR4A is unrefereed, so it can be set aside." It is unrefereed, and this article says so.  But unrefereed is not wrong.  The response so far is Robert Carter's 2021 objection to its assumptions, not a competing calculation.  The right move is to recompute it from public 1000 Genomes data.  The tools are free.  Nobody has done it.

"If the molecular clock is wrong, the 495,000-year figure collapses." It would move.  It would not collapse.  The clock converts observed substitutions into years.  The substitutions are observed whether or not the conversion is right.  A young-earth model still owes an account of how alleles separated by dozens of differences came to exist, at high frequency, from at most four founding copies.  Rejecting the clock moves the problem.  It does not remove it.  On the per-generation rate the trees were dated with, the clock would have to have run about 60 to 80 times faster, on average and genome-wide, to bring TMR4A down to 6,000 years, and 40 times faster to bring TMR10A down to a Flood 4,500 years ago.²⁷

"Carter showed eight Ark passengers carry most of today's diversity." He showed a real and useful thing.  Measured on HapMap phase 3 data, "1.97 to 4.75 individuals would capture 55% to 77% of the allelic diversity".²⁹  That kills the Tasmanian-devil intuition.  But note what the statistic is.  It counts what fraction of catalogued SNP alleles a small sample of people carries.  It says nothing about whether any single locus holds more than four deep lineages.  That is the question in dispute.  The two must not be run together.  Details of the Flood bottleneck itself, including minimum viable population and inbreeding, are in Population Bottlenecks.

"African genomes look older than everyone else's." This is a real problem, and Hössjer names it himself.  The recent unique-origin model "faces some challenges, in particular to explain why African DNA looks older than non-African DNA."²⁸  He offers candidate explanations without a quantitative test.  Sanford's "demographic stirring" is the same idea and is also untested.  Nobody has simulated a young, structured post-Babel population that reproduces the African and non-African contrast in nucleotide diversity, rare-variant fraction and linkage decay at the same time.

"Nobody has refuted Hössjer and Gauger." No refereed rebuttal exists.  OpenAlex lists seven citing works, and none of them is one.  The nearest thing is Swamidass 2022.  It calls allele-frequency-spectrum inference "low-confidence and disputed" in general, and in the same breath says Hössjer and Gauger's normalised spectrum is "more aligned with the field" than the version the mainstream side published.²  The only substantive response found anywhere is Erkell's blog series with Hössjer's reply.  That is not a journal exchange.  This cuts both ways.  The model has not been refuted.  It also has not been engaged.  Erkell notes that its journal "is not included in the scientific database SCOPUS".¹⁵  A model published there will not be engaged by default.

Issues and unanswered questions

These computations would settle the question.  None has been run in the form needed.  Each is cheap, uses public data, and would decide against one side or the other.

1. A full DRB1 intron analysis, extending the count run here. This article's window count is a first pass at the test Hössjer and Gauger proposed. It finds five to nine families more than 1% apart in 16 of 18 windows. Three steps would make it decisive. Build maximum-likelihood trees for each window rather than distance clusters. Add the neighbouring DRB genes to the alignment, so that long gene-conversion tracts copied from them are tested rather than assumed away. Weight alleles by their measured population frequencies. Bergström found that gene conversion does not appear to extend into the adjacent intron sequences, which already narrows that escape route. If more than four families survive all three steps, the four-chromosome model is dead at DRB1, and only a model with ten deep families at the Flood remains.
2. An independent TMR4A, reported as a distribution. Recompute it on 1000 Genomes phased haplotypes with an independent ancestral-recombination-graph method. Publish the whole distribution, plus the fraction of loci whose TMR4A exceeds any candidate young-earth date. This is the highest-value piece of original work available on this question.
3. An identical-by-descent test of a bottleneck of eight. Segment-based methods can see recent severe bottlenecks. They recovered the Ashkenazi bottleneck at 25 to 32 generations. Simulate eight founders at about 155 generations followed by realistic growth. Ask whether that signature would be detectable in 1000 Genomes-scale data. If it would be, and it is not there, a method with resolution at the right depth refutes the Flood bottleneck. If it would not be, a large mainstream objection disappears.
4. How many common haplotypes does a block actually contain, block by block? Gabriel and colleagues published the averages in 2002: 3.5 in Asian samples, 4.2 in European, 5.0 in Yoruban and African-American, and 5.3 for the union across populations.³⁰ Nobody has published the distribution and its maximum. Recompute it from 1000 Genomes phased haplotypes: for every block, the number of distinct haplotypes covering 90% of chromosomes. Recombination among four founders cannot produce five or more. A substantial tail of blocks with five or more falsifies the created-diversity model at those blocks. A distribution that stops at four vindicates it.
5. Sanford's simulations at realistic population sizes. Re-run the Designed Alleles Model with growth from eight into the millions rather than a fixed census of 1,000. Report a goodness-of-fit statistic instead of a visual comparison of normalised curves.

Sources that could not be obtained, and what each blocks.  The full texts of Bergström et al. 1998 and von Salomé et al. 2007 are closed at every route tried.  Only their abstracts could be read, from NCBI E-utilities.  So Gauger's "seven ancestral versions" cannot be checked against the paper it is attributed to.  The 1996 Science technical comment on Ayala is closed and has no abstract.  Doxiadis et al. 2008, which Gauger cites for the intron trees, could be read only as an abstract.  Adam and the Genome has no scan on archive.org and none in Open Library.  Its chapter-3 quotations here were confirmed word for word by Google Books search-inside on the volume itself.  That tool returns no page numbers for a reflowable ebook, so none are given.

Verdict

The arguments most often used to rule out a single couple as the sole genetic ancestors of all living humans do not rule it out.  Linkage disequilibrium, population-size curves and incomplete lineage sorting do not test it.  The mainstream authors who use them have conceded as much in print or deleted the claim.  Raw allele counts do not test it either, because most alleles at hypervariable loci are recent.

Two measurements rule it out at a few thousand years, and no published calculation on the young-earth side has met either.  HLA-DRB1's introns, read from 680 complete gene sequences, fall into five to nine families more than 1% apart in 16 of 18 stretches tested, against a founding ceiling of four.  That measurement uses no clock and no chimpanzee.  The median time back to four surviving lineages across the genome is 358,000 to 495,000 years, depending on the data and method, and the median time back to Noah's ten is about 180,000.  Both could be wrong, and this article says where each is soft.  The DRB1 count is this article's own and has not been checked by anyone else.  Swamidass himself reports that HLA genes show convergent evolution, which would blur any count built from the protein-coding exons, though not one built from introns.  TMR4A has never been refereed, and the methods behind it date recent bottlenecks too old, by about twofold for the program behind the 495,000.  Robert Carter has objected to its assumptions in print.  Neither measurement has been challenged with a calculation.

The gap is not about counting alleles.  It is about counting deep families.  Six thousand years of mutation separates two lineages by about 0.0005% of sites.  DRB1's lineages differ in their introns by 0.2% to 11%.  Created diversity can supply four deep families at no cost, and Hössjer's version stops there, which makes it testable.  It fails the test at DRB1.  The designed-gametes version raises the limit to ten at the Flood, passes DRB1, and meets TMR10A, 40 times older than the Flood on the same clock.

What an honest young-earth position looks like here.  Say that the standard disproof is over-claimed, and show the deletion.  Say that the SNP-level version of the allele argument is false, and show the 99.8%.  Then say plainly that DRB1 and TMR4A are unanswered.  Hössjer and Gauger named the DRB1 test and sourced their answer to a book rather than a measurement.  Run on the published sequences, the four-chromosome version fails it.  The best model on this side fits at 500,000 years, not at 6,000.  Closing that gap needs a mechanism nobody has yet specified or measured.  Anything stronger than that is not supported by the data on the table.

Companion pages

• The literature. Every source consulted, catalogued with what it measured, whether it was read in full, and which way it cuts.
• Quarantine. The four quotations and numbers that did not survive checking, including one this article got wrong about a creationist source and one that a second access route resolved.
• Population Bottlenecks. Minimum viable population, inbreeding, purging, and Robert Carter's Flood-bottleneck papers.

Sources