Technology Trees of Civilization

“Technological development has a dependency structure that can be mapped” is at least four claims wearing one sentence, and conflating them is what makes a technology tree feel obviously true. First: some technologies are physically impossible without others — no integrated circuit without a way to grow single-crystal silicon. Nobody disputes that. Second: those prerequisites form a stable, discoverable graph whose edges are properties of the technologies rather than of a time and place. Third: that graph is the binding structure, so knowing it tells you what to build next. Fourth: the graph can be built by a working group and used to plan. The first is trivially true; the second has weaker evidence than intuition suggests; the third is where the measurement goes badly; the fourth has been attempted at enormous scale for forty years and essentially never evaluated against outcomes.

A disclosure, in the first section rather than in a footnote, because a reader who learns it later will read everything before it as setup. The Canadian Institute for Exotic Science publishes a technology tree. It has a typed schema, a ship threshold fixed in writing before adjudication began, and it has failed that threshold — decisively, not narrowly. The Institute is therefore an interested party in this subject in the most direct way available. It appears below as one case among six, alongside NASA's technology taxonomy, fifteen national hydrogen roadmaps, the semiconductor roadmap, four commercial game tech trees and the patent-network measurement of technology space. Its negative result about itself carries more weight than its positive claims would, by exactly the interested-party rule this site applies to everyone else — and that is the only reason it is admissible at all. If the argument here still stands with the Institute's case removed, the brief has done its job.

Established Begin with the strongest evidence that structure is real, because it is genuinely strong and it must not be shortchanged. Hidalgo and Hausmann model products as requiring subsets of unobserved capabilities, so that a country can make a product only if it holds all of them — their own analogy is that “a child is able to produce a Lego model if the child's bucket contains all of the necessary Lego pieces.” Formalised as a country-capability matrix crossed with a product-capability matrix, tested across more than a hundred countries and three product classifications. Established The synthesis of that research programme is the principle of relatedness, stated by fifteen authors across regional science, economic geography and network science as “an empirical principle describing the probability that a region enters (or exits) an economic activity as a function of the number of related activities present in that location.” Pooling roughly thirteen studies at country, regional, metropolitan and firm scale, they report that “the probability of entering an activity rises between eight-fold and twenty-fold when we move from an unrelated activity to a related activity.” An eight- to twenty-fold effect replicated across four scales by multiple independent groups is a serious result and it is conceded here in full.

Frontier But note exactly what it is a result about, because the transfer that everyone makes is the one the evidence does not license. Relatedness is symmetric similarity of required capabilities. Two activities are related when they “require similar knowledge or inputs.” The finding says that a place making X is far more likely to start making something resembling X. It does not say X must precede Y, and it supplies no mechanism for orienting an edge between two related activities. A technology tree is a directed acyclic graph. The principle of relatedness is an undirected proximity network. They are different objects, and the eight-to-twenty-fold result does not transfer to the first.

Established When somebody measures the structure directly, at the largest scale anyone has, the answer is “structure yes, tree no.” Alstott, Triulzi, Yan and Luo built technology-space networks from 3,911,050 US utility patents granted between 1976 and 2010, covering 2,756,508 inventors and 247,913 firms, across three classification systems at 430, 121 and 629 classes. The result is sparse: for most proximity measures, only “about 20% of the links were stronger than chance”, so most technology pairs have no significant relationship at all. It has community structure identified by weighted modularity — clustered, not hierarchical. And after normalisation the different proximity measures converge toward a single dimension of technological proximity, described as things being “particularly related” rather than prerequisite-ordered. Inventors were more likely to work across two domains when the pair was closely related; firms' activity was only modestly associated with the structure. Established Four million patents return a sparse clustered non-hierarchical network with one dominant proximity dimension. That is the most direct empirical answer available, and it is not a tree.

Established The intuition that it should be a tree has a datable origin and it is not a scientific one. Tuur Ghys traces the artefact to Francis Tresham's 1980 board game Civilization and its 1991 video-game descendant, then studies four commercial implementations — Age of Empires (1997), Empire Earth (2001), Rise of Nations (2003) and Civilization IV (2005) — through 160 hours of gameplay analysis plus interviews with lead designers. The measured result: only 30% of technologies overlap between the games, and of the overlapping technologies just 5 of 35 share similar effects. Established Four commercially successful trees, built by professionals to represent the same history, agree on less than a third of their nodes and on one in seven of the effects of the nodes they share. If the dependency structure of technology were discoverable, four independent attempts to discover it would not diverge that far. Frontier Ghys's specific criticisms are worth carrying because they are structural rather than aesthetic: invented prerequisites — China developed writing without an alphabet and Civ IV's structure forbids that path; he asks why mysticism must precede robotics — circular determinism taken from the manuals themselves, as in Age of Empires' claim that cultures “had to master the preceding technology”; domain exclusion, with textiles and domestic innovation simply absent and state-level technologies privileged over consumer ones; and a single western track where all civilisations share one tree.

Established The largest institutional technology map in existence has no edges. NASA's space technology roadmaps went through two National Academies reviews. In 2012 the committee and six panels examined 14 draft roadmaps covering 14 technology areas, identified 320 level-3 technologies which the panels revised to 295, designated 83 high priority, and narrowed to 16 highest-priority for the next five years. The 2015 revision covered 15 technology areas in a four-level taxonomy containing 340 level-3 technologies, of which 42 were newly added and 5 of those made high priority — grappling, remote interaction, terrain-relative sensing, autonomous targeting, and thermal protection system modelling. Established And the 2016 review chapter on future independent reviews states that it “does not address dependencies between technologies”, prioritising individual technologies through a lead / collaborate / watch / park framework and explicitly “not discuss[ing] how technologies may enable or constrain each other's development.” Frontier 340 nodes, four levels of hierarchy, a statutory mission, two independent reviews — and essentially no encoded edges, with the reviewers not raising the omission. That is evidence about the difficulty of the task rather than about NASA, and it is the closest thing to a control case this subject has. The uptake evidence is separately thin: a NASA inspector-general audit of December 2015 found that “deficiencies in NASA's management processes and controls may limit its efforts to effectively manage its portfolio of space technology investments,” including a delayed revision of the investment plan the roadmaps were supposed to feed.

Established The most careful evaluation of a roadmap family ever published concludes that roadmaps are not projections at all. Will McDowall read fifteen national and regional hydrogen roadmapping initiatives — Australia, Canada, China, the EU, Germany, Iceland, India, Japan, the UK and the US, dated 2002 to 2009 — and reports that “roadmaps should not be understood as projections or forecasts. Rather, roadmaps conflate and combine three different ways of understanding the future: expectations (what is thought likely to happen), desires (what is hoped will happen) and promises (what will be made to happen).” Established He names the roadmapper's dilemma: “a confident, prescriptive roadmap developed on the basis of a consensus of a subset of relevant (and powerful) actors will have most influence. Yet on the other hand, this is likely to reflect incumbent interests.” He observes that “in all the roadmaps, the future is much like the present in terms of consumer behaviour, cultural practices” despite each depicting a radical infrastructure transition, and that “most of the roadmaps reviewed appear to be one-off exercises, rather than ongoing management processes.” Established And he explicitly declines to score outcomes, on the grounds that attribution is impractical. That refusal is itself a finding: the field's most careful case study of a roadmap family judged outcome measurement infeasible. Frontier The single sentence about expectations, desires and promises is the most useful thing in this brief. A dependency edge in any published technology tree is one of those three, and no schema in the record distinguishes them — including this Institute's.

Established The one roadmap that unambiguously worked was terminated by the people who ran it. The Semiconductor Industry Association created a national roadmap in 1994, international from 1999, which Chris Mack describes as institutionalising an “industry standard Moore's Law” and supplying “common terminology to suppliers and customers” — while stating flatly that “Moore's Law is not a law, it is an act of will.” The last edition was published in 2016. The reason given in the trade press is precise: with only a few firms still developing leading-edge fabrication in each of logic and memory, “each leading-edge company has a secret internal roadmap and little motivation to compare directions.” Two successors were announced, one for devices and systems in May 2016 and one for heterogeneous integration. Frontier The best case for roadmapping is also the case against generalising it. It worked because many mutually dependent firms all needed the same equipment on the same schedule and none could build the toolchain alone. When consolidation removed that condition, the roadmap stopped. It was a coordination technology whose value depended on an industrial structure, not a discovery about how semiconductors depend on each other. Technology Forecasting uses the same termination to make a point about forecasts; here the point is about what kind of object a roadmap is.

Established And the practice around all of this is enormous, which should be said plainly before anything critical. Industrial roadmapping begins at Motorola in the 1970s; the academic centre of gravity is Cambridge, where a single 2004 framework paper carries 871 citations on Crossref's count and a 2015 overview reports a survey of public-domain roadmaps identifying more than 2,000 examples. A 2020 review maps the field into seven distinct schools — two at Cambridge, plus Seoul, Portland, Bangkok, Beijing and Moscow. Frontier Against that: a 2022 systematic review of industry-level roadmapping applications found 23 journal articles across two decades, reported that stakeholder involvement effort is low, proposed design improvements, and did not report effectiveness evidence, because there is little to report. Established Forty years, seven schools, two thousand published roadmaps, 871 citations on one paper, and no outcome evaluation with a comparison group was located. Mark the interested party: the Cambridge centre trains and consults on roadmapping, and its own promotional overview concedes that “the process of developing roadmaps is more important than the roadmaps themselves”, offering as its evidence of outcome a single packaging-company case with a “reinvigorated innovation strategy” and no quantitative metric or comparison group. When the method's most prominent advocates locate its value in the workshop rather than in the artefact, the artefact is not doing the work the framing attributes to it. This is the same shape as the finding Public Policy Foresight reaches about government foresight, arrived at independently in a different literature.

Frontier Can a directed prerequisite edge be recovered from data at all? This is the central open question and nothing in the record settles it. Proximity is measurable at four-million-patent scale. Direction is not measured anywhere. Speculative The obvious candidate — order technology pairs by first-use or first-patent date and call the earlier one the prerequisite — conflates precedence with necessity, and Ghys's example is the counterexample: writing preceded the alphabet in China and never required it. No published method takes an undirected technology-proximity network and returns a validated set of oriented prerequisite edges, and until one exists the tree remains an interpretation laid over a network rather than a measurement of one.

Speculative The strongest theoretical argument for prerequisite structure is a simulation, and it should be labelled as one. Arthur and Polak built logic circuits from primitive elements in a computer model in which previously discovered circuits become building blocks for later ones, and the qualitative claim — that technologies are combinations of earlier technologies, so intermediate stepping stones are necessary even when useless in themselves — is the best case anyone has made for why dependency should exist in principle. Speculative The full text was not obtained for this brief and it is cited by existence only. It is a simulation result rather than an observation of the historical record, and the gap between those two is the whole question.

Frontier Is the semiconductor condition general? The one roadmap that worked required many mutually dependent firms, a common technical schedule, and no actor able to build the toolchain alone. Nobody has tested whether those conditions are necessary, sufficient, or merely present in the one success everyone cites. Speculative If they are necessary, roadmapping is a coordination instrument for fragmented industries and its application to unfragmented ones is a category error. If they are not, nothing explains why the hydrogen roadmaps — fifteen of them, across a decade, in nine countries — produced expectations, desires and promises rather than a schedule anyone met.

Frontier Will anyone ever score a roadmap's milestones? The obstacle McDowall named is attribution, not data. Fifteen hydrogen roadmaps with dated targets are in the public domain and the outturns are known. Established Fourteen years after a careful researcher declared outcome scoring impractical, no one has published an attempt and reported what makes it impractical. That is a gap of effort, not of method.

Frontier Does dependency mapping work only on bounded systems? The design structure matrix is the honest, working version of the tech tree: it represents the elements of a system and their interactions to expose the system's architecture, and it is documented in use at Boeing and John Deere across software, aircraft engines and facilities design. Frontier It works because it is bounded — one product, one organisation, one process, every element enumerable and every interaction checkable by someone who owns both ends. The gap between mapping a jet engine and mapping civilisation is not a difference of degree; it is the difference between a system with an owner and a system without one. Whether anything in between is tractable is untested.

Handwave And the strongest unmade argument, named here precisely because nobody has made it. The category-error position holds that technological “dependency” is a retrospective description rather than a forward constraint: any given capability has many possible substrates, history picked one, and the resulting narrative is read back as necessity. Handwave It is consistent with Ghys's China-and-the-alphabet case and with the sparse non-hierarchical patent network, and no located author argues it as a thesis. It is flagged handwave and listed as unmade, because an argument nobody has made is not evidence.

Established The measurable bottleneck is that the dependencies which actually bind frontier technology are mostly not technological, and this shows up in an independent place before it shows up in this Institute's corpus. Lawrence Berkeley National Laboratory's 2026 interconnection-queue analysis reports 2,061 GW of generation and storage capacity actively seeking grid connection in the United States — solar 773 GW, storage 749 GW, wind 220 GW, natural gas 253 GW. Median duration from request to signed interconnection agreement was well over three years in 2025, and time to commercial operation exceeds five years. Of the capacity requested between 2000 and 2020, 13% had come online by the end of 2025 and 75% had withdrawn — with withdrawal rates above 40% even after an interconnection agreement is signed. Established Two thousand gigawatts of technology that already exists, already has a buyer and already has finance, is waiting on a queue. No technological prerequisite is missing anywhere in that number. The binding constraint is a queue-management process and a transmission-planning institution. A technology tree of the energy transition, however well drawn, contains none of the information needed to explain a 13% completion rate.

Frontier The second bottleneck is in the schema rather than in the world. McDowall's three categories — expectations, desires, promises — are not distinguishable in any dependency notation in use. A published edge that says A depends on B may record a physical necessity, an engineering convention, a funding sequence, or an aspiration held by whoever ran the workshop. Established No roadmapping schema located, including this Institute's, marks which of those an edge is, and no amount of adjudication effort can recover the distinction after the fact.

Frontier The third is that the people who have thought hardest about the formal properties of technology trees are optimising them for something else. The small formal literature on depth, branching, redundancy and dead ends is in games research, including an automated approach to measuring the structural and quantitative properties of trees so that analyses of the results “may help to produce significantly better technology trees.” Frontier Better there means more playable. That is the correct standing note and it is not a joke: the vocabulary the rest of us borrow was tuned for a different objective function. Only bibliographic records and abstracts were obtainable for those papers and nothing beyond them is attributed here.

Established This is where the Institute's own case belongs, disclosed, and it is reported here because it is a negative result about its own programme. The same numbers used to support a positive claim would not be admissible. Established On 2026-08-26, before any adjudication began, the Institute fixed a three-clause ship rule for its technology tree in writing, with no partial ship and no discretionary override: briefs carrying a typed brief-to-brief edge must reach 60% of briefs adjudicated; briefs never asked the dependency question must stay at or below 5%; and the longest simple dependency path must reach at least 4 edges. The rule states its own purpose: “the rule is fixed before the adjudication starts, so the adjudicator cannot tune the outcome without knowingly breaking a written rule. That is a weaker guarantee than independence and a stronger one than good intentions.” And on the obvious escape hatch: “'We adjudicated and came up two points short' is not a reason.”

Established The threshold was not missed by two points. It was unreachable by authoring before the authoring began. At 161 briefs the coverage clause read 72 / 161 = 44.7%. Reaching 60% required about 61.5 further briefs, every one of them carrying a dependency, and only 45 slots existed — a best possible case of 117 / 206 = 56.8%, still under the bar with every remaining brief perfect. The rate actually observed once 70 briefs had been adjudicated was 8 of 70, or 11.4%, which projected the ratio downward, to about 37%. Established That projection has now been tested and it held. Against the live corpus of 295 briefs the measurement returns 113 typed depends-on edges, 6 typed enables edges, and 76 / 295 = 25.8% of briefs carrying a typed edge, with a longest simple dependency path of 5 edges. 36 further briefs were written and the count of briefs carrying a typed edge rose by three — a net marginal rate of 3 / 36 = 8.3%, below even the figure the projection used. The ratio fell from 44.7% to 25.8% against a predicted 37%. The prediction was made in advance and landed within about one point.

Established Two of the three clauses passed and the failing one is now failing against a clean denominator. The never-asked clause is at 0 of 197, or 0.0%, down from 70 — every brief in the corpus has been asked the dependency question, so the ratio can no longer be moved by shrinking its denominator. The longest-path clause passed at 5 edges against a bar of 4. Coverage is the only failing clause and there is no denominator argument left to make about it. Frontier The explicit-decline marker has been used twice, on 1.0% of the corpus; the rule anticipated that zero declines would be “a finding about the template”, and two is barely better, so the finding stands. Established And every brief written moves the tree further from its own bar. That is arithmetic rather than rhetoric: at an 8.3% marginal rate against a 38.1% average, each new brief adds one to the denominator and 0.083 to the numerator.

Established Meanwhile the layer that was never part of the threshold grew, and it is now several times larger than the one that was. The same measurement returns 1003 typed non-brief constraint edges naming 999 distinct constraints, carried by 251 of 295 briefs, or 85.1%, against 119 brief-to-brief edges in total — a ratio of about three to one. By kind: institutional or regulatory 212, market or capital 51, industrial capacity 48, a scientific result nobody is producing 22, supply chain 14. Established Institutional constraints alone outnumber every technology-to-technology edge in the corpus combined by nearly two to one. 64.5% of briefs record a non-brief constraint; 38.1% record a dependency on another brief. The map the project set out to build is the sparser of the two things it found, and the trajectory is one-directional: an earlier release recorded 285 constraint edges across 108 briefs, and the Institute's own method page records a still earlier state in which Research Bottleneck Analysis notes that “the topic index declines to publish a technology tree … because the underlying briefs assert only 24 directed dependency edges across the whole corpus.”

Frontier The two layers are not summable and the measuring tool refuses to sum them, on stated grounds worth quoting because they are the whole argument in a sentence: seven briefs carry both kinds, and “the two measure different things — the map, and the world. There is no combined percentage because there is nothing a combined percentage would be about.” Established The finding, stated once: an organisation set out to map how its subjects depend on each other, pre-registered a threshold, adjudicated every one of 197 briefs, and found that the dependencies are mostly not on the map. Frontier subjects wait on institutions, on capital, on industrial capacity, on missing physics and on supply chains far more than they wait on each other. The threshold failed not because the adjudication was lazy — the never-asked clause is at zero — but because the thing being measured is less common than the schema assumed.

Frontier Generalise that exactly as far as it goes and no further. One corpus, one adjudicator, one taxonomy, 197 briefs, one subject area — frontier and speculative technology, which is precisely the region where institutional and scientific constraints would be expected to dominate. It is a data point, not a law. It is corroborated at two independent points — the interconnection queue, and NASA's decision not to encode edges at all — and it is in tension with the eight-to-twenty-fold relatedness effect, which is a much larger and much better-identified result about a different question. Speculative Whether 60% brief-to-brief coverage was ever the right bar for a frontier corpus is an open question this brief raises and does not settle, because the Institute cannot settle it about itself. The threshold's own authors wrote down in advance that a metric “measuring something other than what it names” would be sufficient reason to change it, and that clause is now live.

Established The cheapest valuable experiment is to score a roadmap, and the corpus is already assembled. Fifteen hydrogen roadmaps with dated milestones sit in the public domain, their target years have passed, and the outturns are known. McDowall declared outcome scoring impractical on attribution grounds fourteen years ago and nobody has tested that judgement. Somebody should try, and publish either the scorecard or a precise account of what makes it impossible. Both results are useful and the second is more interesting.

Frontier Second, and this is the one that would settle the subject: try to orient the edges. Take the normalised patent proximity network, add first-grant dates, and test whether any orientation rule recovers prerequisite relationships that domain experts endorse out of sample. Speculative The likely outcome is that precedence and necessity come apart — that is what Ghys's counterexamples predict — but nobody has run it, and a clean negative would be the first direct evidence that directed prerequisite structure is not recoverable from the data we have.

Frontier Third: replicate the four-game comparison at larger n. Ghys measured 30% node overlap and 5 of 35 shared effects on four titles. Dozens of strategy games with published tech trees exist, machine-readable. A twenty-tree comparison would establish whether 30% is the stable level of agreement among independent attempts to represent the same history, or an artefact of four cases.

Established Fourth: mark the edges by type. Any dependency schema could require each edge to be labelled physical necessity, engineering convention, funding sequence or aspiration — McDowall's three categories plus the trivially true one. Frontier No published schema does this, the labelling cost is one field per edge, and until somebody pays it, no dependency graph can be audited for the failure mode its own literature has identified.

Frontier Fifth: a comparison group for roadmapping. Forty years of practice and no study with one. Firms and agencies adopt roadmapping at datable moments, and adoption is uneven across otherwise similar organisations. The design is ordinary and the data are in annual reports and patent records; the absence is of will, not of method.

Established Dependency modelling has a working engineering form and its defining property is boundedness. A design structure matrix represents the elements of a system and their interactions, exposing the architecture, and it is documented across product architecture, organisation design and process at Boeing, John Deere and in software, aircraft engines and facilities design. Frontier It works where every element is enumerable, every interaction is checkable, and one organisation owns both ends of every edge. Scaling it requires the enumeration and the ownership, and civilisation-scale technology maps have neither.

Frontier What a usable dependency schema would have to carry is now fairly clear from the failures. Edge typing, so that necessity and aspiration are distinguishable. A recorded adjudication state, so that “asked and empty” is distinguishable from “never asked” — the distinction that turned this Institute's coverage figure from ambiguous into meaningful. A vocabulary for constraints that are not nodes in the graph, since those turn out to be the majority. And a fixed threshold, written before the count, since a graph metric with a movable bar measures the adjudicator. Established All four are cheap. None of the six cases in this brief has all four, and the one that has three of them failed on the fourth.

Frontier The coordination artefact is the other engineering object here, and it has a different requirements list. The semiconductor roadmap needed a common technical schedule, many mutually dependent firms, and no single actor able to build the toolchain alone; the International Energy Agency's own guide to producing roadmaps offers no method for representing dependencies between technologies and treats evaluation as “development of indicators to help track progress against roadmap milestones.” Frontier Mark the interested party: that agency is a producer of roadmaps publishing guidance on how to produce roadmaps, and its collection lists ten spanning roughly 2014 to 2023. Established Cross-cut that against the finding in Technology Forecasting: over the same period, 2,905 integrated-assessment-model projections of solar cost decline averaged 2.6% a year, none above 6%, against a 15% outturn. The institutions producing energy technology roadmaps were, over the same decade, producing systematically and unidirectionally wrong forecasts for the flagship technology inside them.

The boundaries with three neighbours are tight and are stated rather than implied. Technology Forecasting owns rate and date forecasting — S-curves, experience curves, the tracked forecast record; this brief owns dependency structure and sequencing. The two share exactly two sources, on Moore's law and on the semiconductor roadmap's termination, and use each for a different purpose. Research Bottleneck Analysis is this Institute's own four-constraint taxonomy — physics, engineering, cost, institutional — and a prior self-assessment cited here rather than re-derived. Megaproject Governance owns cost and schedule overrun.

Within this map, also: Civilization Timelines, which asks the same question of historical periods that this brief asks of technologies and reaches a structurally identical answer; Innovation Ecosystems and Scientific Funding Models, where the portfolio-selection question a tech tree is supposed to answer is actually decided; Public Policy Foresight, whose evaluation gap this brief re-encounters in a different literature; Civilizational Planning, which owns the plans a roadmap would feed; Institutional Design, where the constraint type this brief measures as dominant is stated generally; Energy Corridors and Energy Storage Revolutions, which sit directly behind the interconnection queue; and Existential Risk Governance, where sequencing claims about dangerous capabilities do real work.

Outside it: economic complexity and evolutionary economic geography, which supply the relatedness result; scientometrics, which supplies the patent-network measurement; systems engineering, which supplies the design structure matrix; game studies, which supplies the only careful comparative analysis of tech trees as artefacts; and science and technology policy, which supplies the roadmapping literature and its evaluation gap.

Established The agency with the strongest reputation for producing frontier technology does not use a dependency tree. DARPA's selection instrument is the Heilmeier catechism, eight questions formulated by a director in the mid-1970s: what are you trying to do, in no jargon; how is it done today and what are the limits; what is new in your approach and why will it succeed; who cares and what difference will it make; what are the risks; how much will it cost; how long will it take; and what are the mid-term and final exams for success. None of the eight asks about prerequisites or foundational technologies. Established The broader model has been characterised as three elements — organisational flexibility at the administrative level, significant authority for program directors to design programs and select projects, and active project management — aimed at “mission-oriented research on nascent S-curves within an inefficient innovation system.” That characterisation presents no outcome data; it is a framework paper and is cited as one.

Frontier The one careful outcome evaluation of an agency built on that model is mixed and the mixed part is the important part. Goldstein, Doblinger, Baker and Díaz Anadón tracked 25 cleantech startups funded by ARPA-E in 2010 over ten to fifteen years against three comparison groups — rejected applicants, recipients of other government programmes, and comparable cleantech startups. They find a strong innovation advantage in patenting over all comparison groups, and better post-award business success than rejected applicants, but no significant difference from other cleantech startups on business outcomes, concluding that the agency “was not able to fully address the 'valley of death' for cleantech startups within 10–15 yr after founding.” An author correction was published in September 2020 and travels with the paper. Established The question a tech tree exists to answer — what should we build next — is answered in practice by asking eight questions about one project, none of which is about the graph, and the best-evaluated instance of that approach shows a real innovation effect and no measurable effect on the outcome that matters.

Frontier The institutional requirement that would change this subject is a scored roadmap, and it is nobody's job. Roadmaps are produced by agencies, trade bodies and consultancies whose interest lies in the roadmap being used rather than in its milestones being checked. Established Both institutional constraints recorded in section 13 are of that kind — things a funder, a standards body or a system operator could choose to supply and has not. Established And interested parties are dense here and run in every direction: the Cambridge centre sells roadmapping training, the energy agency produces the roadmaps it publishes guidance about, the national laboratory reports on the queue it also studies, and this Institute publishes a technology tree while assessing technology trees. The last is the reason only its negative result appears above.

Established The tech tree carries a theory of history and the theory is one academics reject. Ghys's transmission finding is the part of his paper that belongs in a serious brief: tech trees cultivate the belief that “social progress is driven by technological innovation, which in turn follows an 'inevitable' course” — a belief “considered a fallacy in the academic community” yet “still widely diffused in popular culture.” Frontier The structural criticisms are the mechanism: invented prerequisites presented as history, whole domains excluded, state technologies privileged over consumer ones, and a single western track imposed where all civilisations share one tree. Speculative What is not established is that players come away believing false things about history. Ghys documents the content and its academic rejection; he did not measure beliefs, and no such study was located. The honest statement is that the artefact encodes determinism, not that it produces determinists.

Established A roadmap is a claim on other people's investment, and its authorship is therefore a distributional question. McDowall's dilemma states it exactly: the roadmap with most influence is the confidently prescriptive one built on consensus among a powerful subset of actors, and that is the one most likely to reflect incumbent interests. Frontier His observation that in all fifteen hydrogen roadmaps “the future is much like the present in terms of consumer behaviour, cultural practices” despite each depicting a radical infrastructure transition is the same point from the other side: the technology changes and the social arrangements do not, which is a choice about whose future is being drawn.

Established And an organisation assessing an instrument it publishes has one obligation, which is asymmetry. This Institute's negative measurement about its own tree is admitted here; a positive measurement would not have been. Frontier The rule is not a courtesy. It is the same standard applied to the Cambridge centre's promotional overview, to the energy agency's guidance, and to a vendor's forecasting shape, and applying it to oneself is the only thing that makes applying it to others defensible.

Established There is a dependency structure. It is measurable. It is not a tree, and it is not the thing that binds. At the largest scale anyone has measured it — 3.9 million patents over 35 years — technology space is a sparse, clustered, non-hierarchical network in which about a fifth of links exceed chance and proximity collapses onto roughly one dimension. Real structure, wrong shape. Established The strongest positive result in the field — the eight- to twenty-fold rise in the probability of entering a related activity — is about similarity of required capabilities, not about ordering. It tells you a place making X will probably start making things like X. It does not tell you X must come first, and it supplies no method for orienting an edge.

Established And the record of attempts is consistent across six unrelated cases. Four professionally built trees for the same history agree on 30% of their nodes and on 5 of 35 shared effects. NASA, with a statutory mission, 340 catalogued technologies and two National Academies reviews, encoded no dependency edges at all and its reviewers did not raise it. Fifteen national hydrogen roadmaps turned out on inspection to fuse expectations, desires and promises rather than to project. The semiconductor roadmap — the one unambiguous success — was ended by its own participants when the industrial structure that made coordination valuable disappeared. Two thousand gigawatts of built-and-financed technology sits in a queue that no technology map describes. And an organisation that pre-registered a threshold and adjudicated every brief in its corpus returned 38.1% brief-to-brief coverage against a 60% bar it could not have reached by authoring, with 347 non-technological constraint edges against 118 technological ones.

Frontier The synthesis is that the map is real and the map is not the territory the framing assumes. Technological development has structure — proximity, relatedness, combinatorial building-up — and that structure is genuinely measurable at scale. What it does not have is a prerequisite graph that is both discoverable and binding. Where dependency mapping works, it works on bounded systems with an owner. Where it is attempted on unbounded ones it produces a taxonomy with no edges, or a coordination artefact that dies with its industry, or a document that fuses hope with forecast, or — if you measure honestly and publish the failure — a graph whose edges mostly point out of the graph.

Frontier The practical consequence for anyone planning at civilisational scale is a change of question. Asking what has to be invented before X is usually the wrong question, because the answer is short and already known. Asking what has to be permitted, financed, built and supplied before X is the question with the long answer, and it is the one no tech tree is shaped to hold. Established Technology Forecasting arrives at the same place from the other end: the technologies that got forecast right were the ones somebody organised to deliver. Two briefs, two methods, one conclusion, and neither set out to reach it.

These horizons track measurements, institutions and methods rather than technologies, because the brief's own finding is that a technology date is the wrong unit:

• 10 yr: Frontier The patent-network measurement of technology space will get larger and better normalised, because the data are growing and the method is published; whether anyone attempts to orient its edges is the open question and there is no sign of it yet. Speculative Expect at least one systematic attempt to score a roadmap family against outturn, because the hydrogen corpus is now old enough that every milestone has resolved. Frontier The semiconductor roadmap's two successors either acquire the participation the original had or they do not, which is observable rather than forecastable. Established And this Institute's own coverage ratio will continue to fall as long as briefs are written at an 8.3% marginal rate, which is arithmetic.
• 25 yr: Speculative If edge typing is adopted anywhere — a schema that distinguishes physical necessity from funding sequence from aspiration — the first auditable dependency graphs become possible about a decade later, and the question of whether prerequisite structure is real becomes answerable rather than arguable. Speculative If it is not adopted, the field in 2050 looks like the field in 2000: many roadmaps, seven schools, no comparison group. Handwave Which of those happens depends on whether a funder ever requires it, and predicting that is predicting a committee.
• 50 yr: Speculative The deeper possibility is that the interesting structure is not in the technologies at all but in the institutions, in which case the mapping effort migrates — from dependency graphs of capabilities to graphs of permits, queues, standards and financing, which is what the constraint measurements in this brief keep returning. Speculative That would be a genuinely different instrument, and nothing currently under construction resembles it. Handwave The alternative, that a directed prerequisite structure is eventually recovered from data and turns out to be predictive, cannot be excluded and has no evidence behind it.
• 100 / 250+ yr: Handwave Beyond anything this brief supports. The oldest object here is forty years of roadmapping practice and the longest measured series is 35 years of patent records. Handwave The only durable observation is negative and structural: every attempt so far to draw a prerequisite graph of an unbounded technological system has produced either no edges, contested edges, or edges pointing out of the graph, and there is no reason yet to expect a century of further attempts to change that.

• Depends on Nothing on this map. No brief here produces a result this one waits on, and recording a dependency edge in the brief that is about the temptation to record dependency edges would be the exact failure it measures. No typed depends-on edge is claimed, and the reason is the brief's own finding: the constraints that bind are not other briefs.
• Requires (not on this map) One missing scientific result and two institutional facts, none of them produced by any brief on this map. The scientific one is specific and it is the question this brief exists to ask: proximity between technologies is measurable at four-million-patent scale, and direction is not measured anywhere. No published method takes an undirected technology-proximity network and returns oriented prerequisite edges that domain experts endorse out of sample; the obvious candidate, ordering pairs by first-use date, conflates precedence with necessity, and the counterexamples are well known — writing preceded the alphabet in China and never required it. Until that method exists, every published technology tree is an interpretation laid over a network rather than a measurement of one, and this is a discovery nobody is currently attempting. The institutional ones are cheaper and just as absent. Fifteen national hydrogen roadmaps with dated milestones have been sitting in the public domain for over a decade with every target year now resolved, and no one has scored them; the researcher who read all fifteen judged outcome scoring impractical on attribution grounds in 2012 and nobody has tested that judgement or published what makes it impossible. And the binding constraint on the technologies this map is largely about is a queue: 2,061 GW of generation and storage seeking US grid interconnection, over three years to a signed agreement, more than five to commercial operation, 13% of capacity requested between 2000 and 2020 online by the end of 2025 and 75% withdrawn, with withdrawal above 40% even after an agreement is signed. Neither institutional requirement is a research problem; both are things a funder, a standards body or a system operator could choose to supply and has not.
• Enables In principle a validated method for orienting dependency edges would change how every brief on this map records section 13. No typed enabling edge is claimed, because no such method exists and an edge pointing at a capability nobody has is not a dependency, it is a wish.
• Adjacent Economic complexity and evolutionary economic geography; scientometrics and patent-network analysis; systems engineering, for the design structure matrix; game studies, for the only careful comparative work on tech trees as artefacts; science and technology policy, for the roadmapping literature; and within this map Technology Forecasting, Innovation Ecosystems and Institutional Design.

Handwave “A technology tree is how technology actually works.” The artefact is a game-design import with a datable origin in a 1980 board game and its 1991 video-game descendant. Established Four professionally built versions of it, made to represent the same history, agree on 30% of their nodes and on 5 of 35 shared effects. Established And the direct measurement of technology structure at 3.9 million patents returns a sparse clustered non-hierarchical network, not a directed acyclic graph. There is structure. The tree is a rendering convention.

Established “Roadmaps are forecasts.” The most careful evaluation of a roadmap family in the literature says the opposite in its own words: roadmaps “conflate and combine three different ways of understanding the future: expectations… desires… and promises.” Frontier A published dependency edge is one of those three and no schema in use marks which. Reading a roadmap as a projection is reading a negotiated commitment as a measurement.

Frontier “The semiconductor roadmap proves roadmapping works.” It proves that roadmapping worked once, in an industry of many mutually dependent firms who all needed the same equipment on the same schedule and none of whom could build the toolchain alone. Established When consolidation removed that condition the participants ended it, on the stated grounds that each remaining leading-edge firm now has a secret internal roadmap and little motivation to compare directions. Speculative Whether the condition is necessary for any roadmap to work is untested, and the fifteen hydrogen roadmaps are the obvious case where it was absent.

Frontier “Economic complexity research proves technologies have prerequisites.” It proves something large and different: an eight- to twenty-fold rise in the probability of entering a related activity, where related means requiring similar knowledge or inputs. Established That is symmetric proximity. A prerequisite is directed. The relatedness literature contains no mechanism for orienting an edge, and transferring its effect size to a tech tree is transferring a result about one object to a different object.

Established “NASA has a technology tree.” It has a four-level taxonomy of 340 level-3 technologies, twice independently reviewed, and the 2016 review states that it “does not address dependencies between technologies” and does not discuss how technologies may enable or constrain each other. Frontier The world's largest institutional technology map is a catalogue. The correct inference is not that dependencies do not exist; it is that the largest institutional attempt chose not to encode them and its independent reviewers did not ask why.

Frontier “This Institute's tree failed, therefore dependency mapping is impossible.” No. One corpus, one adjudicator, one taxonomy, 197 briefs, in exactly the region of technology space where institutional and scientific constraints would be expected to dominate. Speculative It is corroborated at two independent points and it is in tension with a much larger and better-identified result. It is a data point, and it is admitted here only because it is negative and self-reported — the same numbers used to support a positive claim would carry far less weight, which is the rule this site applies to everyone and is applying to itself.

Speculative “A 60% brief-to-brief coverage bar was obviously the wrong bar.” Possibly. The threshold's authors wrote down in advance that a metric “measuring something other than what it names” would be sufficient reason to change it, and after adjudication the constraint layer the bar ignored turned out to be three times larger than the layer it measured. Frontier But the argument that a bar was wrong is much more persuasive before the number comes in than after, and this brief deliberately does not settle it, because the party with the interest cannot.

Frontier “Roadmapping improves outcomes.” Forty years, more than two thousand public-domain roadmaps, seven research schools, 871 citations on one framework paper — and a systematic review of the industry-level literature found 23 articles and reported no effectiveness evidence. Established The field's most prominent practitioner, in a promotional document, says the process matters more than the artefact and offers one case study with no metric and no comparison group. That is a claim about workshops, and it may well be true; it is not a claim about dependency graphs.

Handwave And the framing itself: “what has to be invented first?” is usually the wrong question. Where a hard physical prerequisite exists it is generally short, known, and uncontroversial. The long answer — what has to be permitted, financed, built, supplied and staffed — is where the time goes, and it is measurable: 2,061 GW in a queue, 13% completed, 75% withdrawn; and in one adjudicated corpus, 435 institutional constraints against 119 technology-to-technology edges. Established The dependencies are mostly not on the map. That is the assessment.