Cultural Evolution and Population Size

Frontier If the demographic hypothesis holds in the form its policy users assume — that the number of connected learners sets the ceiling on how much technique a population can hold — then group size is a control knob on cultural complexity, and section 7 says what would show it. The measured version of the prize is narrow and recent: in the largest experiment ever run on the question, 941 undergraduates at a single Australian university, six-person groups beat three-person groups on a virtual arrowhead task by 36.28 points of score (t = 2.09, p = 0.037). That is the demonstrated upside at the population margin, and it appeared in two of four conditions.

Established The honest magnitude is six against three, not a nation against a village. The largest group ever run in this literature is sixteen players, in the 2013 experiment that started it. The same hypothesis is applied to an island population isolated for roughly 10,000 years, to the appearance of complex technology across a continent, and to a national research workforce that grew 23-fold between the 1930s and 2015. Nobody has run an experiment at any population scale between sixteen and several thousand; the extrapolation the policy argument needs spans two to three orders of magnitude and is unmeasured across all of them.

Speculative What the 2026 experiment bought was a filter, not a headcount. The six-person advantage appeared when a participant saw the single best-scoring solution, or had a persistent on-screen record of every solution and its score. It was absent when participants simply saw all solutions in random order (p = 0.499 and p = 0.178 in the two experiments). If that transfers, the gain accrues to whoever owns the ranking and retrieval layer of a research system rather than to whoever employs the most researchers — a different asset, held by different institutions, and one section 12 returns to.

Handwave The condition is a number nobody reports. No field of science publishes the size of its effectively connected learning group, as distinct from its headcount or its co-authorship graph, and archaeology has no population proxy validated against an independent census. Until one exists the upside is bounded by a quantity that is asserted rather than measured, which is why section 5 is mostly about instruments and section 15 is mostly about claims that outran them.

Established Cumulative cultural evolution is the observation that human technique improves across generations beyond what any individual could invent alone; the demographic hypothesis is one proposed explanation for why it sometimes does not. The mechanism is a race: social learning is imperfect, so each learner recovers a degraded copy, and against that loss runs the chance that some learner recovers a version better than the model. The more demonstrators a learner can sample, the more often the best of the sample exceeds the average. Population size enters as the supply of demonstrators.

Established The canonical formal statement is Joseph Henrich’s 2004 paper in American Antiquity, and it is explicitly about Tasmania. Its abstract states that “over an approximately 8,000-year period, from the beginning of the Holocene until European explorers began arriving in the eighteenth century, the societies of Tasmania lost a series of valuable skills and technologies,” naming bone implements, cold-resistant garments, hafted instruments, nets, barbed fishing implements, spear-throwers and boomerangs. The model’s claim is that the “relatively sudden reduction in the effective population size (the size of the interacting pool of social learners)” caused by post-glacial sea-level rise, which cut Tasmania off from mainland Australia for ten millennia, could have started a cultural-evolutionary process of loss.

Frontier The phrase in parentheses is the whole difficulty. The interacting pool of social learners is not the census population, not the archaeologically visible population, and not the number of people alive in a region. A hypothesis stated in terms of a quantity nobody can measure in the past is not thereby false, but every archaeological test then substitutes a proxy, and the proxy carries the test.

Established At least four distinct quantities travel under the name, and the experiments measure the smallest. Census size; the effectively interacting pool; network degree; and the number of demonstrators a learner actually observes. Experiments manipulate the fourth and sometimes the third, archaeology estimates something closer to the first, and the models are written about the second. Section 4 treats that substitution as the field’s central open question.

Established Three evidence streams exist and they do not agree. Controlled experiments in which group size is assigned; correlational tests against ethnographic and archaeological toolkit inventories; and formal models deriving what population size should do under stated assumptions. Maxime Derex and Alex Mesoudi’s review in Trends in Cognitive Sciences summarises the first as “five experiments from four different research groups provide support for a positive effect of group size on cultural complexity” and the second as “Results with this approach have been mixed. Some studies support the hypothesis, but others do not.”

Frontier The hypothesis has a second, less-discussed limb: connectivity. Whether a population is one pool or several partly isolated pools changes the answer, and in some designs it changes the sign. Partial connectivity preserves between-group diversity while letting innovations cross; full connectivity homogenises. That limb makes a simple headcount policy argument suspect even if the demographic hypothesis is true.

Speculative A meta-science corpus owns this question because the policy inference runs the argument backwards. “More researchers, therefore more progress” borrows a model built to explain loss in small isolated populations to justify expanding large connected ones. Nothing in the model says the function keeps rising; several models in section 3 say it saturates, and at population sizes between ten and a hundred.

Established Start with the result the field is tethered to. Maxime Derex, Romain Beugin, Bernard Godelle and Michel Raymond ran a dual-task computer game in groups of 2, 4, 8 and 16 players, published in Nature in 2013. A simple task required drawing an arrowhead on a grid; a complex task required building a fishing net. Larger groups were less likely to lose the complex trait and more likely to improve both. The methods record the failure rate at the bottom: “Only 4 2-player groups (26.7%) conserved the complex task and were therefore excluded from the analysis,” implying fifteen two-player groups were run and eleven lost the fishing net entirely. The accessible page does not state the total number of volunteers, and none is asserted here.

Established Two further laboratory results are cited alongside it. Marius Kempe and Alex Mesoudi ran 80 participants in ten transmission chains on a 100-piece jigsaw, letting each generation see either one semi-completed puzzle from the previous generation or three. Mean pieces solved rose over generations in the three-puzzle condition and not in the one-puzzle condition, by Page’s trend test. Michael Muthukrishna and colleagues reported the same direction in 2014 using image-editing and knot-tying chains. That is the positive record: a handful of experiments, all at group sizes of sixteen or fewer.

Frontier The direct challenge came in Nature itself, and it was a reanalysis of Derex’s own data. Claes Andersson of Chalmers and Dwight Read of UCLA published a Brief Communications Arising in 2014 arguing that on reanalysis “larger groups actually perform worse than smaller groups,” consistent with anthropological evidence discounting the group-size hypothesis for non-food-producing societies. Derex and colleagues replied in the same issue. This brief could not fetch that reply — the publisher’s robots policy refused it — and so makes no claim about what it concedes or defends. The same two authors developed the objection at length in Current Anthropology in 2016, arguing that the demographic treadmill model has become entrenched beyond what its evidence supports and should sit inside a wider set of explanations rather than be privileged among them.

Established The clearest null is large, recent, and points the other way. Nicolas Fay, Naomi De Kleine, Bradley Walker and Christine Caldwell ran 543 participants in transmission chains building paper planes, scored by flight distance, learning from the planes of one, two or four models from the previous generation. Their summary: “there was an inverse relationship between population size and CCE: plane flight distance did not improve over the experimental generations in the 2-Model and 4-Model conditions, but did improve over generations in the 1-Model social-learning condition.” The paper is Increasing population size can inhibit cumulative cultural evolution, PNAS 116, 6726–6731, 2019.

Frontier The commission behind this brief attributed that inhibition result to Derex, Perreault and Boyd in PNAS in 2018. The attribution is wrong. The inhibition finding is Fay, De Kleine, Walker and Caldwell, 2019. Derex, Charles Perreault and Robert Boyd did publish in 2018, but in Philosophical Transactions of the Royal Society B 373, article 20170062, as Divide and conquer: intermediate levels of population fragmentation maximize cultural accumulation — a result about the shape of the connectivity curve, not about group size inhibiting accumulation. Both cut against a monotone reading, but they are different results with different mechanisms, and conflating them overstates the case against.

Established The largest experiment was published in February 2026 and reconciles the positive and negative results by making the effect conditional. Bradley Walker and Nicolas Fay ran two experiments at the University of Western Australia: 287 undergraduates recruited and 281 consenting in the first, 654 recruited and 644 consenting in the second, a combined N of 941. Participants worked a virtual arrowhead task individually or in three- or six-person groups, over ten trials with feedback. Six-person groups beat three-person groups only when shown the single highest-performing arrowhead (beta = 36.28, t = 2.09, p = 0.037); with all models shown in random order there was no effect (p = 0.499). In the second experiment the advantage appeared only with a persistent external record of solutions and scores (beta = 18.78, t = 2.13, p = 0.034) and not in the view-all condition (p = 0.178).

Frontier The authors’ framing is that group size helps when attention can be filtered or memory offloaded, and not otherwise. Their rationale for the two group sizes is cognitive: three-person groups present two observed solutions plus scores, near the roughly four-chunk working-memory limit, while six-person groups present five solutions plus scores, around ten chunks, exceeding it. On that reading the population term does work belonging to a cognitive constraint. The page carries no preregistration statement; data and analysis are posted openly.

Established Now the archaeology, where the commission’s anchor is correct and important. Krist Vaesen, Mark Collard, Richard Cosgrove and Wil Roebroeks published Population size does not explain past changes in cultural complexity in PNAS 113, E2241–E2247, in 2016. This brief could not fetch the paper: the publisher returns 403 and the PubMed Central mirror serves a CAPTCHA. Leiden University’s own announcement states that the demographic models support the relation only under “extremely implausible conditions,” that no archaeological evidence shows behavioural modernity around 40,000 years ago emerging during population growth, and that the Tasmanian record contradicts the models’ predictions.

Frontier The exchange that followed is part of the record. Joseph Henrich and a large group of co-authors replied in PNAS as Understanding cumulative cultural evolution; Vaesen and colleagues replied to that as Reply to Henrich et al.: The Tasmanian effect and other red herrings. Neither could be fetched and neither is characterised here beyond its title — which is itself enough to show that Tasmania is the disputed centre and that each side thinks the other is arguing beside the point.

Established Independent archaeological tests keep coming back null. Briggs Buchanan, Michael J. O’Brien and Mark Collard tested the population-size and environmental-risk hypotheses against projectile-point-type richness in prehistoric Texas, from initial occupation more than 13,000 years ago to about 400 years ago, using bivariate correlations and a generalised linear model. They found no relationship between temporal changes in point-type richness and changes in population size, with global temperature variation predicting richness instead.

Frontier The positive field evidence is thinner than its citation count suggests, and it reuses data. Michelle Kline and Robert Boyd’s 2010 analysis of Oceanic island populations and marine foraging toolkits is the most-cited supporting dataset; Adam Powell, Stephen Shennan and Mark Thomas’s 2009 Science paper on Late Pleistocene demography is the second. Neither could be fetched and neither is quoted here. What a fetched source does say is Derex and Mesoudi’s summary that results against real-world data “have been mixed,” and that comparison is complicated because some analyses rest on overlapping datasets.

Established The formal models are fragile to a definitional choice, and the demonstration is quantitative. Adrien Querbes, Krist Vaesen and Wybo Houkes built an NK model in which a trait is a binary string of N components with interaction density K, so complexity means Herbert Simon’s interaction-based notion rather than a count of parts. They swept population sizes of 10 to 10,000, N of 5 to 20, K of 0 to 4 and copy-error rates from 0.001 to 0.1 across 6,000 parameter combinations over 100 generations. Under weighted selection — copying in proportion to observed fitness rather than identifying the best model perfectly — the population advantage survives only at low complexity, and at high interaction density 10,000 individuals are not reliably better than 10.

Frontier A 2024 model puts a number on where the population term stops mattering, and the number is small. Alexandre Bluet, François Osiurak and Emanuelle Reynaud extended a micro-society model across population sizes from 0 to 100, running 200 simulations of 200 generations per condition. Technology quality rose logarithmically rather than linearly, and where the curve flattened depended on the innovation rate: roughly linear at an innovation probability of 0.001, diminishing returns around 100 individuals at 0.01, a plateau near 20 at 0.05 and near 10 at 0.10. Their conclusion is that “population size has a limited impact on cumulative technological culture and other demographic factors such as innovation rate could offer a viable alternative explanation for archeological records.”

Frontier Two further model results move the target rather than the estimate. Laurel Fogarty, Joe Yuichiro Wakano, Marcus Feldman and Kenichi Aoki showed that demographic trajectory matters independently of level: “a small but growing population could have a different number of cultural traits from a shrinking population with the same absolute number of individuals in some circumstances.” Kaleda Denton, Yoav Ram and Marcus Feldman extended Boyd and Richerson’s 1996 model with finite population size, frequency-dependent transmission and multiple skill levels, concluding that “multiple barriers to cultural accumulation may explain its rarity.”

Frontier Connectivity is the limb where the evidence is most interesting and least settled. Derex and Boyd reported in 2016 that partial connectivity increases accumulation within groups by letting culturally distinct subgroups keep diversity while exchanging innovations; Derex, Perreault and Boyd reported in 2018 that intermediate fragmentation maximises it. Neither paper could be fetched and both are reported only as Derex and Mesoudi’s review characterises them. An interior optimum, if real, means the population-size policy argument has no monotone form at all.

Established The one large natural experiment in a connected modern population points at convergence. Elena Miu, Ned Gulley, Kevin Laland and Luke Rendell analysed 19 online programming contests over 14 years: 47,967 entries, 483,173 unique lines of code, around 120 participants per contest with a range of 63 to 202. Only 3.8% of entries contained at least 50% novel lines; half of all entries had similarity above 0.9 to the current leader; successful small modifications outnumbered successful radical ones 16 to 1; the best solution scored 40 times better than the first, while code diversity fell. A population of about a hundred, fully connected and fully informed, converged rather than diversified.

Frontier The corresponding modern claim about scientific workforces is the weakest link, and it is measured in a different literature. Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb report that United States research effort “has risen by a factor of 23, an average growth rate of 4.3 percent per year” while “research productivity has fallen by an even larger amount, by a factor of 41 (or at an average growth rate of −5.1 percent per year).” The number of researchers required to double chip density is more than 18 times larger than in the early 1970s. In the large connected modern case headcount and output per head move in opposite directions, and this Institute treats that measurement in Research Productivity Decline.

Frontier The first open question is which quantity the word names, and it is not rhetorical. An experiment manipulating the number of demonstrators a learner observes tests a claim about attention and working memory. An archaeological test regressing toolkit richness on a radiocarbon-derived proxy tests a claim about census size. A model written in terms of an effectively interacting pool tests neither. Nothing establishes that the three are monotonically related in real populations, and the 2026 result is most naturally read as saying they are not.

Frontier The second is whether the laboratory effect is about population at all. The six-versus-three advantage appeared only with a best-solution filter or an external record. If the operative variable is how many options a learner can evaluate rather than how many exist, then “population size” here is a label on a cognitive bottleneck — which would explain the 2019 inversion without any appeal to demography.

Frontier The third is whether the archaeological proxies measure anything. Summed radiocarbon distributions, site counts and assemblage richness are all sensitive to excavation intensity and preservation, and richness in particular is sensitive to sample size in a way that can manufacture the predicted correlation from nothing. No proxy in use has been validated against an independent census.

Frontier The fourth is the shape of the connectivity curve. If accumulation is maximised at intermediate fragmentation, the policy-relevant quantity is a network statistic rather than a headcount, and the sign of adding people depends on where the system already sits. Nobody has measured where any real research field sits on that curve, or the curve itself at more than a handful of points.

Speculative The fifth is task depth, and the field has been least willing to pose it. Every laboratory task in the record — arrowheads, fishing nets, jigsaws, paper planes, knots, wheels — has a solution space a motivated individual can search in an hour. The hypothesis is a claim about traits too deep for an individual to reconstruct. A task whose optimum requires combining independently discovered partial solutions is where population size ought to matter most, and it is precisely the condition the experiments do not create.

Frontier The sixth is whether innovation rate substitutes for population size. The 2024 model makes the two interchangeable over part of the parameter space: when innovation is frequent the population term flattens at ten to twenty individuals. If so, an archaeological record showing constant toolkit complexity through a demographic collapse is evidence about innovation rates rather than about the hypothesis.

Speculative The seventh is the information-sharing dilemma underneath all of it. Alex Mesoudi, Ángel Jiménez, Keith Jensen and Lei Chang model cumulative culture as a cooperation problem: social learning is cheaper than invention, so learners free-ride and inventors hoard. Kin selection, reputation-based partner choice and cultural group selection can each resolve it in their models. If sharing is the binding constraint, adding people adds free-riders as fast as demonstrators.

Frontier The experimental answer to that, published in 2026, is a null. Jiménez, Chang, Jensen, Hui Jing Lu and Mesoudi ran 716 participants across four online studies in the United Kingdom and in China and Macau. Reputation-based partner choice was reliably above zero but small — 0.14 with a 95% interval of 0.06 to 0.23 in the UK sample, 0.29 from 0.19 to 0.40 in the Chinese sample — against a simulated threshold near 0.4, and participants shared anyway, including when reputation was removed. Their conclusion: “indiscriminate, group-based sharing is more important for facilitating cumulative cultural evolution than discriminate reputation-based sharing.”

Established The binding constraint is the task, not the participants. Every published experiment uses a solution space shallow enough that an individual working alone gets most of the way there, which is why individual-learning controls keep matching or beating social conditions — in the 2019 paper-plane study the control improved as fast as the best social condition. A task that is genuinely beyond one person within a session, scores automatically, and whose optimum requires combining separately discovered partial solutions does not exist as shared apparatus.

Frontier The second constraint is group size, and it is logistical rather than conceptual. Sixteen simultaneous participants is the ceiling in the record because sixteen people have to be in a laboratory at once. Online platforms lift it — the 2026 partner-choice studies ran 716 participants in groups of seven and eight across two countries — but nobody has used them to run a group of 64 on a cumulative task. The gap between the largest experimental population and the smallest population the hypothesis is applied to is at least two orders of magnitude.

Established The third is the archaeological proxy, and it is the hardest. Past effective population size is estimated from summed radiocarbon dates, site counts or coalescent inference; toolkit complexity from assemblage richness or technounit counts. Two noisy proxies regressed against each other, sharing a sensitivity to how much digging was done, can produce the predicted relation from sampling alone.

Frontier The fourth is that there is no shared instrument, so results do not stack. Derex used arrowheads and fishing nets; Kempe and Mesoudi a jigsaw; Muthukrishna image editing and knots; Fay paper planes; Walker and Fay a different virtual arrowhead. No two of the five share a task, a scoring rule or a generation structure, so a discrepancy between any two is uninterpretable.

Speculative The fifth is measurement of the modern analogue. No scientific field publishes the size of its effectively connected learning group. Co-authorship graphs, citation networks and headcounts are all available and none is the quantity the model needs. The nearest instruments — disruption indices, research-productivity ratios — are themselves contested, and their contest is a measurement dispute rather than a scientific one.

Established That contest is now documented, with numbers. Michael Park, Erin Leahey and Russell Funk reported in Nature in 2023 that disruptiveness fell across 25 million papers from 1945 to 2010 and 3.9 million patents from 1976 to 2010, with declines of 91.9% in the social sciences and 100% in the physical sciences on their index. Vincent Holst and colleagues replied in Nature in August 2026 that removing entries with zero references, which take the maximum disruption score by construction, cuts the measured decline by about 97% in two paper datasets — from −0.31 to −0.01 — and by 58% to 71% across five patent categories. Alexander Petersen and colleagues separately showed the index is biased by citation inflation, with mean reference counts rising from 9 in the 1960s to 23 in the 2000s.

Handwave The sixth is that nobody has to fix any of this, because nothing depends on it. No funding formula, immigration rule or research-workforce target is written as a function of a measured cultural-accumulation curve. An instrument with no decision attached attracts no instrument budget, and that, rather than technical difficulty, is the likeliest reason the experiment in section 7 has not been run.

Frontier The first dependency is a population proxy validated against something else. Every archaeological test rests on an estimate of past population never checked against an independent census of the same population at the same time. Periods with both archaeological signal and documentary census — colonial settlements, tax rolls, parish registers — supply the calibration set, and nobody has built it.

Frontier The second is an independent measure of trait complexity. The NK result is that the answer flips depending on whether complexity means a count of parts or a density of interactions among them, and that available empirical evidence does not let anyone choose between the definitions. Until complexity can be computed from an artefact rather than assigned to it, the models cannot be tested against the archaeological record at all.

Established The third is replication infrastructure, and the corpus already owns the argument. Five experiments from four groups, on five tasks, with one large inversion and one large conditional result, is a literature with no replication layer. Registered reports, preregistered analysis and shared task code are standard in adjacent fields and absent here; the 2026 experiment that sets the state of the art carries no preregistration statement on its published page. This Institute’s Forecasting Replication treats the general problem.

Speculative The fourth is a theory of what a learner samples. Derex and Mesoudi name the mismatch: most experiments let participants learn from several demonstrators at once, while the influential models assume a single source. Models and experiments are therefore not about the same process, and agreement between them would be coincidence rather than confirmation.

Frontier The fifth is a measured connectivity statistic for at least one real population. The interior-optimum result requires knowing where a population sits on a fragmentation axis. Hunter-gatherer camp networks, open-source contributor graphs and co-authorship networks are all measurable; none has been placed on the experiments’ axis.

Speculative The sixth is cross-cultural sampling. The laboratory record is overwhelmingly undergraduates in wealthy countries. The 2026 partner-choice study is unusual in running parallel United Kingdom and Chinese samples, and its partner-choice parameter differed between them — 0.14 against 0.29 — while the behavioural conclusion held. A hypothesis about the human species cannot be settled in one subject pool.

Frontier The decisive experiment is a pre-registered, large-N factorial design that crosses population size against network connectivity independently, on a task deep enough to require multi-step recombination. Every existing result confounds the two. Derex 2013 varied group size with everyone connected to everyone; Derex and Boyd 2016 and Derex, Perreault and Boyd 2018 varied connectivity at fixed size; Fay 2019 varied the number of demonstrators observed, which is neither. The factorial version is the only design that can say whether the population term survives once the network term is controlled, and it is the result this brief would most change its assessment on.

Established The design is specifiable now, in units. Five population levels — 3, 6, 12, 24 and 48 simultaneous participants — crossed with four connectivity structures: fully connected, partially connected subgroups with a stated migration rate, fragmented with rare contact, and a single-demonstrator chain as baseline. Ten to fifteen generations per chain. At the effect size the 2026 experiment measured, a population main effect needs hundreds of participants per cell and the interaction several times that; a total sample in the low thousands is the honest requirement, which online recruitment already supports.

Frontier The task is the part that has to be built, and nobody has built it. It must have a solution space whose optimum an individual cannot reach within a session, must score automatically, and must have at least two independent dimensions whose combination outperforms either alone, so that recombination across sub-populations is the only route to the top. The programming-contest data show why this matters: with about 120 fully connected participants, only 3.8% of entries carried at least half novel lines and diversity fell. A shallow space under full connectivity converges; whether a deep space behaves differently is the question, and there is no apparatus for it.

Established Two conditions from the 2026 experiment have to be carried forward as factors, not as design choices. Filtering, showing only the best solution, and external record, a persistent display of all solutions and scores, each turned a null population effect into a positive one. If they are not manipulated, the experiment measures whichever one the experimenters happened to build in — the likeliest explanation for the field’s inconsistent results, and testable directly.

Speculative Pre-registration would settle a second thing the field cannot. The literature contains one reanalysis of published data reaching the opposite conclusion, and no way to adjudicate it, because the original analysis plan was not fixed in advance. A registered plan would also settle what to do with groups that lose the complex trait entirely, which in the 2013 experiment was eleven of fifteen two-player groups, dropped from the analysis.

Frontier The archaeological side has its own decisive test and it is cheaper. Take a period and region with both an archaeological record and an independent documentary census, compute the standard population proxies blind to the census, and report the correlation. If summed radiocarbon distributions or site counts track a known population, the archaeological stream recovers its standing; if they do not, every existing test on both sides is uninformative and the nulls and the positives fall together.

Speculative The natural experiment for the modern claim is already running and nobody is scoring it. Open-source software, competitive programming and preprint communities give continuous records of population size, network structure, contribution and outcome quality at scales from tens to hundreds of thousands. Miu and colleagues showed the data can be analysed. Nobody has extended that to the size-versus-connectivity question, and that analysis needs no participants and no ethics approval — only access and an outcome measure.

Handwave The experiment that would settle the policy claim does not exist and probably cannot. Testing “more researchers, therefore more progress” causally would need research workforce size assigned exogenously across comparable fields; field-level funding shocks are the nearest analogue and they move money, people and topic selection together. Nobody has proposed a credible test, and the policy argument therefore rests on an analogy to an experiment run on six people.

Established None of this is blocked by instruments. The recruitment platforms exist, the group-size ceiling is a software problem, scoring is automatable, and the archaeological calibration needs archive work rather than fieldwork. Nobody has scheduled or funded the factorial experiment, and no group has announced an intention to run it. The obstacle is that no decision currently turns on the answer.

Frontier What would change this brief’s assessment, ranked. A pre-registered factorial result showing a population main effect surviving the connectivity control, on a recombination-deep task, would move the general hypothesis from contested to supported. The same design returning a null, or an interaction in which added population helps only at intermediate fragmentation, would retire the monotone version. Nothing in the modern research-policy literature would move this brief, because it measures a different system with contested instruments.

Established The apparatus is software, and the specification is short. A synchronous multi-participant platform that can hold 48 people in a session, assign them to an arbitrary contact graph, pass artefacts between generations, score automatically and log every observation a participant made before acting. Frameworks of this kind ran the 2026 partner-choice studies at 716 participants in groups of seven and eight. What does not exist is a shared cumulative-culture task running on one.

Frontier The task engineering is the hard part and it has a stated criterion. The solution space must be deep enough that the individual-learning control does not track the best social condition — the failure mode of the paper-plane study — and must reward combining partial solutions discovered independently. A rugged landscape with tunable interaction density is the obvious construction, and is exactly the structure Querbes, Vaesen and Houkes used in simulation; building it as a human-playable task would make model and experiment commensurable.

Established Instrumentation of the filter is now mandatory rather than optional. An experiment must record and manipulate what each participant could see and in what order: best-only, all-in-random-order, or a persistent scored record. Logging which solution a participant inspected, for how long and in what sequence turns an ambiguous population effect into a measurable attention effect, and makes the serial-position explanation testable rather than post hoc.

Speculative The archaeological side needs data engineering, not excavation. Joining radiocarbon databases, assemblage inventories and documentary census series for the same place and period is a records-linkage problem of the kind this corpus describes in Science of Science and Research Front Detection. The risk is the absence of a common spatial and chronological key, which is a schema problem and a familiar one.

Frontier Cost is small and that is itself informative. A factorial design at a few thousand online participants, including development of a new task, sits within a single mid-size research grant, and a subject population of 941 was assembled from one university’s course-credit pool in 2026. Any account of this field’s stagnation that blames resources is inconsistent with what the existing work cost.

Established Analysis engineering matters more here than instrument engineering. Mixed-effects models with participants nested in chains nested in groups, an explicit power analysis for the interaction rather than the main effect, and a stated rule for groups that lose the trait entirely. The last is not a technicality: dropping the eleven two-player groups that lost the fishing net changes what the 2013 result means.

Established The nearest technical neighbour is collective problem solving, and it has the opposite convention about connectivity. Network-science work on collective search reports that fully connected groups converge prematurely while sparse networks preserve exploration — the same interior-optimum intuition the fragmentation result expresses in cultural-evolution vocabulary. This Institute’s Collective Intelligence treats the design discipline; the two literatures cite each other rarely.

Frontier Cognitive science supplies the mechanism the 2026 result points at. Working memory capacity, serial-position effects and payoff-biased copying are the terms Walker and Fay reach for to explain why six-person groups need a filter that three-person groups do not. That places the explanation inside individual cognition, next to Distributed Cognition, and away from demography.

Established Technical reasoning research is an active rival explanation with its own numbers. Alexandre Bluet, François Osiurak, Nicolas Claidière and Emanuelle Reynaud modelled micro-societies under reverse-engineering, observation and communication conditions across 1,000 simulations of 1,000 generations, reporting mean technology quality of 213.11, 266.39 and 591.87 respectively, and concluding that technical reasoning rather than theory of mind drives cumulative technological culture. That programme treats cognition as driver and demography as modulator — the reverse of the demographic hypothesis’s ordering.

Frontier The causal-understanding result constrains what transmission has to carry. Derex, Jean-François Bonnefon, Robert Boyd and Alex Mesoudi showed that a physical artefact becomes progressively optimised across generations of social learners without explicit causal understanding, and that transmitting causal models between generations had no noticeable effect on the pace. If accumulation does not require understanding, what population size buys is variation and selection, and the quantity that matters is the number of independent trials rather than of minds.

Speculative Economic complexity is the adjacent field with the largest datasets and the same directionality problem. Relatedness and capability models predict what a place can make next from what it already makes, and this Institute’s Technology Trees of Civilization shows the resulting structure is a sparse clustered network with no direction. A demographic hypothesis about cultural accumulation and a relatedness hypothesis about economic diversification make claims about the same object from opposite ends, and neither can orient an edge.

Frontier Palaeogenetics now supplies population estimates here. Coalescent inference from genomes yields effective-population-size trajectories independent of site counts, but it estimates a genetic effective size — a different quantity again from the interacting pool of social learners, and the substitution has not been justified.

Established The evolutionary-theory neighbourhood is contested in its own right. Cultural inheritance is one of the mechanisms disputed under the heading of the extended evolutionary synthesis, which this corpus treats in Extended Evolutionary Synthesis. The framework within which the demographic hypothesis is stated is itself a live argument, though the results in section 3 do not depend on how it comes out.

Established The institutional fact that explains the state of this field is that no decision depends on the answer. No research council allocates by a measured cultural-accumulation curve; no immigration policy cites an interior connectivity optimum; no science-funding formula contains a population term derived from this literature. The hypothesis justifies positions arrived at otherwise, which is a rhetorical use, and rhetorical uses do not fund instruments.

Frontier The publication pattern is a second-order institutional problem. The positive results are in Nature and PNAS; the reanalysis contradicting the flagship result is a Brief Communications Arising in the same journal and is cited far less; the nulls are in specialist archaeology and psychology journals. A reader sampling by venue prestige recovers a settled hypothesis. A reader sampling by design quality recovers a contested one.

Established The field has no registered-report track and its flagship recent experiment carries no preregistration statement. That is not an accusation of bad practice — the 2026 study posts its data, analysis notebook and full statistical output openly, which is more than most — but a literature whose central claim has flipped sign twice in seven years is exactly where the analysis plan needs to precede the data.

Speculative Subject pools are an institutional constraint disguised as a methodological one. The group-size ceiling of sixteen is a property of undergraduate scheduling. Online platforms remove it, and the institutions funding experimental psychology have not reorganised around a 48-person synchronous session being purchasable.

Frontier The archaeology side has a data-governance problem rather than a funding one. Radiocarbon databases, assemblage inventories and regional chronologies sit with different national bodies under different licences and spatial keys. Building the calibration set section 7 asks for is mostly a permissions exercise, and permissions exercises have no natural owner.

Established Where this bites institutionally in the present is science policy, and the corpus has the numbers. Arguments for expanding research workforces, for researcher immigration and for concentrating or dispersing research capacity all reach for a population-and-ideas intuition. Research Productivity Decline records United States research effort rising 23-fold while research productivity fell by a factor of 41, and Innovation Ecosystems records that the innovation-policy instruments with the cleanest quasi-experimental evidence are the small ones. Neither is compatible with a monotone population-to-progress function used as a premise.

Speculative An institution wanting to act on this literature honestly has one available move. Fund the factorial experiment as a multi-laboratory registered report with a shared task. That converts five incommensurable experiments into one comparable series, and costs less than a single instrument in most other categories on this map.

Established The textbook case in this literature is a colonised people, and that is not incidental. The Tasmanian argument is a claim that the Aboriginal people of Tasmania lost technologies over 8,000 years of isolation. It is stated in Henrich’s 2004 paper as a model result about effective population size and contested on archaeological grounds by Vaesen, Collard, Cosgrove and Roebroeks. Whatever the resolution, the claim concerns the technological capacity of a specific living people’s ancestors, made largely by researchers outside that community, and it circulates far beyond the journals where its evidence is argued.

Frontier Assertions that a named population had simpler technology because it was small have a long history of being wrong and a longer history of being useful to somebody. The honest position is not to avoid the question but to hold the evidentiary standard in proportion to the claim’s social weight: a general law inferred from one island, one reconstructed toolkit and a model resting on an unmeasurable quantity is a weak inference carrying a heavy load.

Established The critique is archaeological, not rhetorical. The case against the Tasmanian reading is that the record does not show what the model predicts, and it comes from researchers who work that record, including a specialist in Tasmanian archaeology among the authors. It is not a claim that the question is improper to ask.

Speculative The modern application has its own distributional edge. “More people, more ideas” is deployed in arguments about immigration, fertility policy and the concentration of research capacity, converting a contested hypothesis into a premise. This corpus’s Fertility Decline records that the best-identified evidence on fertility policy concerns the timing of births rather than completed family size; an ideas-and-population argument layered on top is two weak inferences stacked.

Frontier There is a symmetrical error on the other side. Concluding from the archaeological nulls that demography is irrelevant is not what the critics claim and is not supported. The laboratory effect is real under stated conditions and connectivity effects are real. Overcorrecting into “population does not matter” replaces one unsupported general law with another.

Established Research ethics in the experimental stream are unremarkable and worth saying so. These are low-risk behavioural experiments with consented participants, reported in enough detail for a reader to check. The ethical weight in this subject sits on the archaeological and policy interpretations, not on the laboratory work.

Frontier If the demographic hypothesis is false as a general law, the most consequential casualty is a reassurance. The implicit promise in “more researchers, more progress” is that a wealthy civilisation can buy its way out of diminishing returns by training more scientists. The measured record in the one large connected case runs the other way: research effort up 23-fold and research productivity down by a factor of 41. Those numbers admit several explanations; none is a population term with a positive coefficient.

Speculative The depopulation argument inherits the same weakness in the opposite direction. Projections of global population decline are sometimes paired with a prediction of technological stagnation, and the pairing is made by way of this literature. If accumulation saturates at population sizes of tens to hundreds in the models — plateaus near 10 to 20 individuals at higher innovation rates in the 2024 micro-society model — then eight billion people and four billion are both far above any saturation point the models identify, and the hypothesis makes no prediction distinguishing them.

Frontier Connectivity, if the interior-optimum result holds, is the civilisational variable, and it has been moving fast in one direction. Global scientific communication has become more connected and more synchronous within living memory. The programming-contest data show what full connectivity does to a shallow search space at a population of about a hundred: half the entries above 0.9 similarity to the leader and diversity falling through the contest. Whether the same happens to a research field is not known and is the most policy-relevant unasked question here.

Speculative The loss side of the hypothesis survives the critique better than the growth side. That small, isolated, poorly connected groups can lose techniques nobody individually reconstructs is consistent with the models, the laboratory record and ordinary experience of abandoned industrial capabilities. It is a claim about redundancy and documentation rather than headcount, which makes it actionable in a way the growth claim is not — and this corpus treats the archival version in Civilizational Archives.

Handwave The strong civilisational version — that a population threshold gates the arrival of complex technology — is the least supported claim in this brief and the one that circulates most. It requires a monotone relation, a threshold and a measurable population quantity, and the literature supplies none of the three: the relation is conditional in the largest experiment, no threshold has been estimated from data, and the population quantity in the models is explicitly the unmeasurable one.

Frontier What survives is a design lesson about institutions rather than a law about populations. The conditions that made larger groups better in 2026 were a filter and an external record — properties of libraries, journals, search systems and review processes, which a civilisation builds deliberately. If the result generalises, the lever is institutional design and not the census.

These horizons track experiments, instruments and the fate of a policy argument rather than a technology, because the subject has no artefact to date:

• 10 yr: Frontier Online experiments at group sizes above 16 will be run, because the platforms exist and the 2026 studies show the recruitment works; whether anyone crosses size against connectivity in one design is open and nothing has been announced. Speculative Expect at least one direct replication attempt on the 2013 Nature result, which now carries a published reanalysis reaching the opposite conclusion and a 2019 inversion against it. Frontier The archaeological dispute will not resolve on current instruments; each Texas-style null constrains the proxy rather than the hypothesis. Established And the adjacent disruption-index dispute is already resolving toward measurement artefact, with a 2026 Nature reply putting about 97% of the measured decline in two paper datasets down to zero-reference entries.
• 25 yr: Speculative If a validated archaeological population proxy is built, every existing test on both sides becomes reinterpretable. Speculative If it is not, the archaeological stream stays where it has been since 2016: nulls that cannot distinguish a false hypothesis from a bad instrument. Frontier A shared experimental task would make the laboratory stream cumulative for the first time, and its natural sponsor is a multi-laboratory registered report. Speculative The likeliest settled result on this horizon is not about population at all but about attention: that the experimental effect is a filtering effect, which the 2026 conditions already isolate.
• 50 yr: Speculative On present evidence the monotone demographic hypothesis is more likely to be retired than confirmed, replaced by a conditional statement about connectivity, innovation rate and information handling in which population size is one term among several and not the leading one. Handwave Whether the policy argument follows the evidence is a separate question with a poor historical record: Research Productivity Decline and Innovation History both document measured trends that outlived their own refutations in public argument. Frontier The genuinely open item is whether anyone builds an experimental task deep enough to require recombination, because without one this literature stays a study of shallow search.
• 100 / 250+ yr: Handwave At this range the question stops being empirical and becomes one about what records survive. If global population declines substantially while connectivity stays high, the civilisation running that experiment will have data no laboratory can generate. Handwave The one claim this brief makes at this horizon is negative: no result currently in the literature licenses a prediction about the technological trajectory of a smaller, better-connected human population, and any brief offering one would be extrapolating an effect measured between three people and six.

• Depends on Nothing on this map blocks the experimental work, and that is worth saying plainly rather than manufacturing an edge. The nearest genuine dependency is measurement practice: the instruments this subject would borrow to test its modern application are the ones Science of Science reports as disputed, and the replication machinery it needs is the subject of Forecasting Replication. Neither is a result this brief waits on; both are capabilities it would use.
• Requires (not on this map) Three scientific results and two institutional facts, none produced by any brief here. The first is the experiment this brief exists to ask for: every published result confounds population size with network connectivity, no design has crossed them, and the sign of the population term under a controlled network structure is therefore unmeasured. The second is apparatus: every task in the record — arrowheads, jigsaws, paper planes, knots, wheels — has a solution space an individual can search in a session, which is why individual-learning controls keep matching social conditions, and a task whose optimum requires combining independently discovered partial solutions has not been built. The third is the archaeological instrument: no population proxy in use has been checked against an independent census of the same population, so the nulls and the positives are equally uninformative. The institutional two are cheap and absent: the field has no registered-report track and its flagship 2026 experiment carries no preregistration statement, while its group-size ceiling of sixteen is a property of undergraduate scheduling that online platforms already dissolve. None of the five waits on an invention; all five are things a funder or a journal could supply and has not.
• Enables A settled answer would give Research Productivity Decline and Scientific Funding Models a mechanism rather than a correlation for the relation between research workforce size and output, and would tell Collective Intelligence whether its connectivity results and this subject’s fragmentation results are the same finding in two vocabularies. No typed enabling edge is claimed, because the answer does not exist and an edge pointing at an unrun experiment is a wish.
• Adjacent Cultural evolution and evolutionary anthropology; archaeology of the Palaeolithic and of Holocene island populations; experimental economics; network science; and within this map Technology Trees of Civilization, Science of Science, Collective Intelligence, Distributed Cognition, Generational Turnover in Science and Extended Evolutionary Synthesis.

Established “Tasmania proves that small populations lose technology.” It is the textbook case and it is the contested one. Henrich’s 2004 paper sets out an 8,000-year record of lost skills and a model in which post-glacial isolation reduced the effective pool of social learners. Vaesen, Collard, Cosgrove and Roebroeks argue in PNAS that the archaeological record does not match the predictions, and Leiden University’s announcement of that work states the Tasmanian record contradicts them. Henrich and colleagues replied; Vaesen and colleagues replied again, under a title calling the Tasmanian effect a red herring. A reader who has met only the first of those four documents has met a quarter of the argument.

Frontier “The laboratory experiments settled it.” Five experiments from four groups, no two sharing a task, with one large inversion and one large conditional result against them. That is the field’s own count, from Derex and Mesoudi’s review. The 543-person paper-plane study found improvement only with a single model; the 941-participant 2026 study found the group-size advantage only under filtering or an external record; and the 2013 flagship has a published reanalysis of its own data concluding that larger groups did worse.

Speculative “More scientists therefore more progress.” This is the claim the brief was commissioned around and it is the weakest one in it. It takes a model built to explain trait loss in isolated populations of tens to hundreds and applies it to a connected research workforce that grew 23-fold; it assumes monotonicity the models do not have, at population scales the experiments have never reached, using a population quantity nobody measures. The measured record in the target system — research productivity down by a factor of 41 over the same period — is the opposite of what the borrowed model would predict if it applied.

Established “The critics say demography does not matter.” They do not. The critique is that the formal models support the relation only under implausible parameter choices, that the archaeological evidence does not show the predicted pattern, and that the definition of complexity decides the result. Querbes, Vaesen and Houkes state their own conclusion as a robustness finding: the demographic dependence is highly contingent rather than general. Contingent is not absent.

Frontier “Bigger networks are better for innovation.” The result closest to a direct test says there is an interior optimum. Derex and Boyd reported that partial connectivity increases accumulation; Derex, Perreault and Boyd reported that intermediate fragmentation maximises it. This brief could not fetch either paper and reports them as Derex and Mesoudi’s review characterises them. What was fetched points the same way: about 120 fully connected programming contestants with full visibility converged, only 3.8% of entries carrying half or more novel lines.

Speculative “Population size explains the Upper Palaeolithic transition.” The claim is prominent and the direct rebuttal is the anchor of this brief. The demographic reading of the appearance of complex technology, ornament and abstract art around 40,000 years ago is one of the two cases Vaesen and colleagues address, and the Leiden announcement of their work states there is no archaeological evidence that behavioural modernity emerged during a period of population growth. This brief could not read the primary paper and does not assert its internal statistics.

Speculative “Neanderthals were outcompeted because their populations were too small to sustain complex culture.” A live hypothesis, not a finding. Fogarty, Wakano, Feldman and Aoki frame the question exactly this way, and their contribution is that demographic trajectory can matter independently of level, so that “a small but growing population could have a different number of cultural traits from a shrinking population with the same absolute number of individuals in some circumstances.” That makes the comparison harder, not easier.

Frontier “Complexity is complexity.” The most underrated result in this literature says the definition decides the answer. Under a count-of-parts definition, larger populations sustain more complexity in the models. Under Herbert Simon’s interaction-density definition, with realistic weighted rather than perfect selection, a population of 10,000 is not reliably better than a population of 10. No available empirical evidence distinguishes the two definitions.

Handwave “There is a population threshold for cumulative culture.” No threshold has been estimated from data. Models produce saturation points that move with the innovation rate — near 100 individuals at an innovation probability of 0.01, near 20 at 0.05, near 10 at 0.10 in the 2024 micro-society model — which is a statement about a parameter nobody has measured in any real population, ancient or modern. A threshold quoted without its innovation rate is not a number.

Established “Population size at least predicts toolkit size in the ethnographic record.” The field’s own summary is that the record is mixed and partly overlapping. Derex and Mesoudi write that results from real-world data “have been mixed,” and that interpretation is complicated because some analyses rely on overlapping datasets. The independent test this brief could read — more than 13,000 years of projectile-point richness in Texas — found no relationship with population and environmental risk predicting richness instead.

Speculative “This is a settled corner of evolutionary anthropology that only contrarians dispute.” A Nature Brief Communications Arising in 2014, a PNAS paper in 2016, a reply and a reply to the reply, an inversion in PNAS in 2019 and a conditional replication in Scientific Reports in 2026 is not the citation trail of a settled question. The dispute is between the original authors and named specialists, and it is twelve years old.