Every source the main article rests on, grouped by role, with what each one says and which way it cuts. Everything below was fetched and read during the research except where marked not obtained; those are collected at the end, because an argument resting on a source nobody on this side has read should be flagged rather than buried.
Verbatim quotations, page locators, parameter tables and independent recomputations live in the fourteen dossiers beside this article: research/haldane-primary.md, research/sixties-debate.md, research/kimura-neutral.md, research/modern-popgen.md, research/remine-corpus.md, research/remine-critiques.md, research/waiting-time-behe-durrett.md, research/waiting-time-sanford.md, research/waiting-time-critiques.md, research/hominin-empirics.md, research/measured-adaptation-rates.md, research/cuts-both-ways.md, research/verification-history.md and research/verification-numbers.md. Together they run to about 175,000 words. Library keys refer to sources/index.tsv.
A note on method. No quotation in this article or in any dossier was established through WebFetch, which answers a prompt through a small summarising model rather than returning the page. Every quotation was read out of bytes downloaded locally — PDFs converted with pdftotext -layout, HTML stripped locally, and in four cases (Haldane 1957, Haldane 1960, Kimura 1968, Kimura & Maruyama 1969) read directly off 350-400 dpi rendered page images, because the available scans have no usable text layer or drop characters.
Haldane's own papers
Haldane 1957, "The cost of natural selection," J. Genet. 55:511-524. The founding paper, read whole. Source of the cost derivation D = -2 ln(p₀) and its four other dominance cases, of the padded figure of 30, of the judgment I = 0.1 that produces the 300, of the measured I = 0.03 that contradicts it, of the tachytelic exception at 13 generations, and of the summary sentence "This accords with the observed slowness of evolution." Cuts both ways, and mostly against the popular creationist reading: it is the birth certificate of the argument and the clearest statement that its author did not think he had refuted evolution. Key haldane-1957.
Haldane 1960, "More precise expressions for the cost of natural selection," J. Genet. 57:351-360. Confirms 1957 — "The main results of Haldane (1957) are confirmed" — and gives the exact strong-selection correction K/(-ln(1-K)). Contains the unclosed loophole: a gene drifting up while still slightly deleterious would substitute "with no 'cost' at all, in fact a slight negative cost," demonstrated in a forthcoming paper that never appeared. Key haldane-1960.
Haldane 1937, "The effect of variation on fitness," Am. Nat. 71:337-349. The genetic-load paper, included because the load/cost distinction is Haldane's own: 1937 handles the static depression of fitness at equilibrium, 1957 the dynamic depression during a substitution. Its own framing of load is as "the price paid by a species for its capacity for further evolution." Key haldane-1937.
Haldane 1964, "A defence of beanbag genetics," Perspect. Biol. Med. 7:343-359. His last published word on the cost. Key haldane-1964.
Dronamraju (ed.) 1990, Selected Genetic Papers of J. B. S. Haldane, Garland — the 536-page facsimile volume from which the 1937 and 1964 texts were extracted and against which the 1957 pages were cross-checked, and the source of Crow's Foreword and its "reproductive excess" gloss. Key haldane-1990.
The 1960s and 1970s debate
Read in the order they appeared, asking of each: which assumption of Haldane's does this attack, and does the attack work? Full treatment in research/sixties-debate.md.
Van Valen 1963 (Am. Nat. 97:185-190) coined the phrase and proposed the first solution; Van Valen 1965 withdrew it, calling the inter-locus interference "probably inexorable." The single most damaging fact for the "it was solved" narrative. Brues 1964 (Evolution 18:379-383) argues the load is an accounting artifact, and concedes in passing that the real limiting factor would more likely be "paucity of useful mutations" — which hands the waiting-time argument its premise.
Sved, Reed & Bodmer 1967, King 1967 and Milkman 1967 (all Genetics 55) introduce threshold and truncation selection. Sved 1968 (Am. Nat. 102:283) and Maynard Smith 1968 (Nature 219:1114) carry the truncation argument over from segregational to substitutional load, which is the move that matters. Scope warning: most of the 1967 work is about how many polymorphisms a population can carry, not how fast genes can be replaced, and both the creationist literature and some mainstream replies conflate the two.
Kimura & Maruyama 1969 (Heredity 24:101) add finite population size and find the substitutional load larger for selected alleles starting rare — by an additive 2, about 8.7% in their worked example, not by a factor of two as an earlier draft of this project's research recorded. Ewens 1970 (TPB 1:129) is the standard refutation, targeting Haldane's optimal genotype. Felsenstein 1971 (Am. Nat. 105:1) and Nei 1971 (Genetics 68:169) independently re-derive the limit from reproductive excess instead — the derivation the modern literature rests on, and the one Nei summarises as "HALDANE's conclusion is essentially correct." Grant & Flake 1974 (PNAS 71:3863) survey the solutions and concede the rate limit survives. Wallace 1975 (Evolution 29:465) defines hard versus soft selection, but about segregational rather than substitutional load.
Neutral theory, and the denominator that broke the cost argument
Kimura 1968 (Nature 217:624) founds the neutral theory on Haldane's arithmetic. King & Jukes 1969 (Science 164:788) reach the same conclusion and disown the load argument in the same paragraph. Kimura 1983, The Neutral Theory of Molecular Evolution, spells out the intolerable-load number and answers Maynard Smith with an objection that is the waiting-time problem in 1983 dress. Ohta 1973 and Ohta 1992 — the nearly neutral theory; note that the 1992 review contains no discussion of the cost at all, which is itself a finding.
Nei 2005 (MBE 22:2318) and Nei 2013, Mutation-Driven Evolution, carry the decisive criticism: Kimura counted all four billion nucleotides as targets when the unit is genes. Nei's recomputation gives a factor of about two rather than a factor of a hundred thousand. Kern & Hahn 2018 (MBE 35:1366) put the same point in one sentence. Jensen et al. 2019 (Evolution 73:111) is the reply defending neutral theory, and belongs beside Kern & Hahn or neither should be cited.
The constrained-fraction literature — Siepel 2005, Meader 2010, Ponting & Hardison 2011, Rands 2014 (8.2% of the human genome constrained), ENCODE 2012 and the Graur 2013 and Doolittle 2013 rebuttals — is here because the number of human-chimpanzee differences that fall in functional sequence is the denominator of the whole dispute.
The modern re-derivation and the first measurement
Matheson, Exposito-Alonso & Masel 2025, Genetics 229:iyaf011. The most important modern paper. It re-derives the Nei-Felsenstein reproductive-excess limit, states that the argument "has not yet been fully addressed," measures Haldane's 10% ceiling for the first time (8.5% to 95% in Arabidopsis), concedes that "the empirical collapse of historical arguments for neutral theory has not yet led to a re-evaluation of related arguments by Haldane," and names creationist attention as a reason to settle it. Code deposited. Two preprint versions were also obtained; the 2021 v1 is blunter than the published paper in several places and the differences are documented.
Nunney 2003, "The cost of natural selection revisited," Ann. Zool. Fennici 40:185-194. Cited in the article (subsection "The one simulation built to test it"). Direction: complicates. Verified: full, quotations read per page from the PDF (pdf page p = journal page 184 + p). A hard-selection simulation of a density-regulated population in a changing environment with 1-7 loci. Cost C is the shortest interval between substitutions at all selected loci the population survives; C = C0(M) + nC1(M), with M = 2Ku new mutations per locus per generation. Against Haldane, costs are sublinear across loci (p. 189-190); for Haldane, the speed-up needs M ≥ 1, i.e. a census above 10⁵ at u = 5 × 10⁻⁶ (p. 193), and below M ≈ 0.5 cost and extinction risk rise steeply. His K = 10,000 example ("about every 300 generations") is not a per-substitution rate: on his Table 1 (M = 0.1: 150.2 + 44.4n) it matches three or four loci each substituting in that interval, one substitution per 82-94 generations, with a marginal 44 generations per extra locus. An earlier version of this entry said the example "gives exactly Haldane's 300 generations"; that was wrong per substitution. Weak in citation terms: 19 citations in 23 years, almost all in conservation biology, and the 2025 Masel treatment does not cite him at all. The URL in most bibliographies is dead; the live one is annzool.net/shared/PDF/anz40/anz40-185.pdf.
Hickey & Golding 2019, "Sex solves Haldane's dilemma," Genome 62:761-768, with its 2020 Correction (a typesetting fix, obtained) and its published source code, which shows the model renormalises fitness each generation and truncates the population — i.e. it is a soft-selection model of exactly the kind Matheson et al. criticise.
Also Charlesworth 2013, Kondrashov 1995, Lesecque, Keightley & Eyre-Walker 2012, Agrawal & Whitlock 2012 and Galeota-Sprung, Sniegowski & Ewens 2020 on the load side. The genetic-entropy article already carries most of this material and it is cross-linked rather than restated.
The creationist corpus
ReMine 1993, The Biotic Message — the Haldane chapter and appendix, obtained through archive.org full-text search with the leaf-to-folio map verified. Source of the 1,667 figure at page 217 and of the rhetorical question the argument actually rests on. ReMine 2005, J. Creation 19(1):113-125, recovered as the original PDF from the Wayback Machine. ReMine 2006, CRSQ 43(2):111-120, the publisher's PDF — the paper in which he argues Haldane's cost figure should be higher, which cuts against the attribution of 1,667 to Haldane. Eight archived pages from his dead website, including his verbatim reproduction of Felsenstein's referee report and Crow's public reply.
Sanford, Brewer, Smith & Baumgardner 2015, Theor. Biol. Med. Model. 12:18 — the waiting-time flagship, with its full parameter set audited against the Mendel's Accountant Fortran source code, which is public under GPL-3.0 and was cloned and read. The recovered supplementary "Instructions for waiting time experiments" and the Mendel user manual, both retrieved by a Wayback CDX sweep of the dead mendelsaccountant.info, are the only public source for the defaults the paper silently imports.
Rupe & Sanford 2013, Proc. 7th ICC article 32, "Haldane's Ratchet" — the paper that actually makes the Haldane argument from Sanford's group, and whose headline deleterious result was never simulated. Sanford et al. 2018, Proc. 8th ICC 200-216, the Adam-and-Eve designed-diversity paper containing the sentence that decides the symmetry question. Sanford, Genetic Entropy, 4th edition — note that the copy on the reading shelf is the 2005 second edition and page cites are edition-dependent.
Hössjer, Bechly & Gauger 2018 (Springer Proc. Math. Stat. 271:245-313) and 2021 (J. Theor. Biol. 524:110657) — the mathematically serious ID treatment. Hössjer is a professor of mathematical statistics at Stockholm. Note that the brief for this research asked for a "Hössjer, Gauger & Sanford" paper; no such paper exists, and Sanford appears on none of Hössjer's work.
Behe 2007, The Edge of Evolution; Behe & Snoke 2004 and Lynch 2005 in Protein Science; Axe 2010 and Gauger & Axe 2011 in BIO-Complexity.
Popular creationist restatements: Batten 2005 (CMI), Sarfati 2025, Tay 2019, CreationWiki. These are recorded because they are what most readers meet, and because ReMine disowns the framing several of them use.
Young-earth sources on the other side of the ledger: Wise 1994, Wood 2003 and 2013, Jeanson 2015, Jeanson & Lisle 2016, Lightner 2012, Chaffey 2019, Carter 2016, Ramgren 2025.
Basener & Sanford 2018, "The fundamental theorem of natural selection with mutations," J. Math. Biol. 76:1589-1622 (sources/papers/basener-2018_*). Considered and not used. It revises Fisher's fundamental theorem to add a mutational-effects term and argues that mean fitness can decline under realistic mutation. That is the genetic-entropy argument about harmful mutations, not Haldane's cost of substitution or the waiting-time problem, so it belongs to the Genetic Entropy article, which this one cross-links rather than restating. Verified: abstract and text on disk.
The critics
Musgrave 2007, "Haldane's non-dilemma," The Panda's Thumb, with its full 66-comment thread — the standard reply, its arithmetic, and the two mainstream commenters who identified its method error within hours. Musgrave 1999, 2002 and 2008 as well; the 2002 demolition of ReMine's weasel argument is correct and decisive and the article says so.
Felsenstein in four venues across ten years, including a long direct exchange with the owner of this project at theskepticalzone.com. Source of the concession that the 1970s literature "was left unresolved, a mess," of his classification of ReMine's cost as legitimate in type, and of his identification of standing variation as the escape that scales.
TalkOrigins CB121 and the longer essay behind it — recorded with the warning that CB121 attributes to ReMine an argument he disowns and prints the incoherent figure "4.8 × 10⁷ genes." Robert Williams' four-part site, the most detailed non-professional reply written, whose author admitted in public that he had not read the book he was critiquing. Stern Cardinale 2021. Wilf & Ewens 2010 (PNAS 107:22454) and its only published critique, Ewert, Dembski, Gauger & Marks 2012 (BIO-Complexity 2012(4)).
The empirical base
Generations and divergence: Yoo et al. 2025 (Nature 641:401, complete ape genomes, 5.5-6.3 Myr); Langergraber et al. 2012 (PNAS 109:15716, generation times measured from parentage in eight wild chimpanzee communities); Moorjani et al. 2016; Wang et al. 2023 (Sci. Adv. 9:eabm7047, 26.9 years); Fenner 2005; the Chimpanzee Sequencing and Analysis Consortium 2005 (Nature 437:69) for the 1.23% and 35 million figures; Makova et al. 2024; Demuth et al. 2006 for gene-family turnover.
How many were adaptive: Hernandez et al. 2011 (Science 331:920) — the single most important paper here, and the one most often cited for the opposite of what it says; Uricchio, Petrov & Enard 2019 (Nat. Ecol. Evol. 3:977) — the only paper giving both α and the raw count; Boyko et al. 2008; Eyre-Walker & Keightley 2009; Galtier 2016; Zhen et al. 2021; Enard et al. 2016; Torgerson et al. 2009 for the regulatory fractions. Hernandez et al. 2011 supplement (sources/papers/hernandez-2011_classic-sweeps-rare-supplement.*, fetched 2026-10-27) gives the selection coefficients behind the main text's power claim (strong 1%, weak 0.1%) and an analytic 99%-power threshold of 4.5% and 10% of amino-acid substitutions respectively; cited in the article as [^hernandez-2011:som]. It shows Uricchio's weak class (αW = 0.097 at 2Ns = 10) sits under what Hernandez could detect. Direction: supports the creationist side on the narrow point that "classic sweeps were rare" does not show adaptive substitutions were cheap. Verified: full (text of the .docx). Human accelerated regions: Pollard et al. 2006, Prabhakar et al. 2006. Human-specific genes: Dennis et al. 2012 (SRGAP2C), Florio et al. 2015 (ARHGAP11B), Fiddes et al. 2018 (NOTCH2NL).
Drosophila calibration: Sella et al. 2009 (PLoS Genet. 5:e1000495) — the one source converting α into per-generation rates, and the source of "an adaptive amino acid substitution every 200-400 generations"; Smith & Eyre-Walker 2002; Andolfatto 2005. Fly mutation rates from Keightley et al. 2009, Schrider et al. 2013 and Haag-Liautard et al. 2007, which disagree among themselves by a factor of 2.4.
Measured sweeps: Bersaglieri et al. 2004 and Tishkoff et al. 2007 (lactase); Yi et al. 2010 and Huerta-Sánchez et al. 2014 (EPAS1, and note it was introgressed from Denisovans rather than newly mutated); Hedrick 2011 (malaria alleles); Mathieson et al. 2015 (ancient DNA); Cook et al. 2012 and 2013 (peppered moth); Colosimo et al. 2005 and Barrett et al. 2008 (stickleback Eda); Grant & Grant 2002 (finches); Hawks et al. 2007 for the outlier claim.
Experimental evolution: Blount et al. 2008 and 2012 (citrate); Tenaillon et al. 2016 (50,000 generations); Good et al. 2017 (60,000 generations) — the source of the calculation showing drift cannot account for the LTEE's observed frequency changes; Meyer et al. 2012 (phage lambda); Weinreich et al. 2006; Bridgham et al. 2009; Summers et al. 2014 (chloroquine).
Theory of valley crossing: Iwasa, Michor & Nowak 2004; Weissman et al. 2009 and 2010; Hermisson & Pennings 2005 (cited in the article since 2026-10-27: p. 2339, a new mutation destined for fixation is "at least as likely" as fixation from standing variation within the time a standing allele needs to fix, when Θ < 1; used, as this site's reading, to cap the standing-variation discount for single-site changes near twofold; direction complicates; verified full); Messer & Petrov 2013; Jensen 2014; Pritchard, Pickrell & Coop 2010; Berg & Coop 2014. Binding-site turnover: Dermitzakis & Clark 2002; Moses et al. 2006; Schmidt et al. 2010; Villar et al. 2015 — the measured form of the "many targets" answer.
Corrections made during this research
Recorded because a research file that contradicts a later verification is worse than no file.
- Kimura & Maruyama's "larger by about two" is additive, not multiplicative. An early draft read it as a factor of two. Read off the page image, their result is L(p) = 2[1 + loge(1/p)] against Haldane's 2 loge(1/p), and their worked example is 25.0 against 23.0 — 8.7% larger. Corrected in
research/sixties-debate.md. - The human-to-fly substitution ratio is 59-93×, not 40×. The original figure used an unsourced fly mutation rate. Corrected in
research/hominin-empirics.mdagainst three primaries. - The Hickey & Golding parameters quoted from the preprint were superseded by the published paper (p₀ 0.01 → 0.05, s 0.05 → 0.02). Corrected in
research/modern-popgen.md. - The claim that the cost of selection "dropped out of the textbooks" is too strong as first stated. Hartl & Clark and Felsenstein's own textbook do omit it entirely, but Futuyma carried it in 1986 and dropped it by 1998, and Ridley 1996 and Hamilton 2009 still teach it. The defensible claim is that the disappearance is datable and partial, not universal.
- The catalogue formerly carried two keys for Sanford et al. 2015. Merged; the redundant text extraction was deleted and the XML kept under the canonical stem.
- Weak selection is not cheap in Haldane's units. The article and a note under Uricchio said a weakly beneficial allele (2Ns of order 10) "costs far less than Haldane's D." Kimura & Maruyama 1969 (p. 112, and the haploid Table 1 on p. 109, read off the page images) show the load of a fixation from one new copy is slightly larger than Haldane's whenever 2Nes is well above 1, and falls toward zero only when |2Nes| ≪ 1. Corrected in the article on 2026-10-27; see
research/strengthen-2026-10-27.md. - Nunney's 300 generations is an interval for several loci at once, not a per-substitution rate. Corrected in the Nunney entry above.
Not obtained
Flagged so no one mistakes secondhand knowledge for firsthand.
- Van Valen 1963 full text — the paper that coined "Haldane's dilemma." Title and volume are confirmed against three databases, but the coinage sentence itself is verified only through Google Books publisher text of a later book. The article should not quote Van Valen's own words until the paper is in hand. Highest-priority acquisition.
- Dodson 1962, "Note on the Cost of Natural Selection," Am. Nat. 96:123-126 — the only item in Van Valen's reference list that could have supplied the name, and the last loose thread on the coinage question.
- Brues 1969, Science 164(3884):1130-1136 — full text unobtained; the standfirst and complete 26-item reference list were recovered. Note that the citation given in most secondary sources is wrong on the issue number, and that "The genetic load: its biological and conceptual aspects" is Bruce Wallace's 1970 book, not a Brues paper.
- Haldane 1961, "Natural selection in man," Progress in Medical Genetics 1:27-37 — archive.org holds only the 1976 New Series volume; HathiTrust returns 403.
- Haldane's two 1961 "in press" papers — established as never published.
- H. Allen Orr 2019, J. Theor. Biol. 476:1-4, "Measures of the efficiency of natural selection during gene substitution," which explicitly compares its measures to Haldane's cost. Elsevier, closed. Worth having: Orr is one of the most senior living theoretical evolutionary geneticists and this is recent.
- Stark 2026, Twin Research and Human Genetics, on Haldane's cost and Feller's critique. Abstract only.
- Wood 2002, Origins 54:5-34, the original AGEing paper. The 2003 ICC paper restates the model in full, so the loss is primacy rather than content.
- Jeanson's Replacing Darwin (2017) and Traced (2021) — the technical claims are fully available in his ARJ papers, which were obtained.
- Agrawal & Whitlock 2012 full text; Wallace 1975; the Hössjer, Bechly & Gauger 2018 Springer chapter (closed access, 69 pp.).
- A clean copy of the Journal of Genetics 55:511-524 printing. Everything in this article rests on two scans of the same reprint plus a facsimile. The Blackwell scan drops characters and should not be used for formulae.