Innovation History

The claim under test is that innovation follows recognisable patterns that history reveals. It is the premise of most innovation policy, of technology forecasting, and of the popular literature on how breakthroughs happen. This page takes it seriously enough to separate it into the four different claims it actually contains, because they have very different evidential standing and running them together is how the weakest one borrows credibility from the strongest.

Frontier The four claims, in ascending order of how badly they are supported. First, descriptive regularities in aggregate series — cost curves, S-curves, exponential improvement. Testable, and partly confirmed. Second, structural regularities in what invention is — recombination, general-purpose technologies, epistemic bases. Model-shaped, with some quantitative support. Third, inevitability, that a given invention would have arrived anyway. This is the strongest empirical leg and it is stronger than most readers expect. Fourth, predictive patterns usable in advance, that history tells you where to invest. This is where the project is weakest and where its critics land, and it is the version the framing is usually deployed to support.

Frontier What the page lands is the same lesson twice, thirty-three years apart, from opposite ends of the field's instrumentation. Griliches, surveying patent statistics for the Journal of Economic Literature in 1990, found that the widely discussed decline in US patents granted during the 1970s “is found to be an artifact of the budget stringencies at the Patent Office”. Petersen, Arroyave and Pammolli showed that the most cited innovation-slowdown finding of the 2020s drifts toward its conclusion mechanically as reference lists lengthen — which they did by a factor of about 2.6 over the period in question. Two headline historical patterns, two recording instruments, and in both cases the metric moved rather than the world.

Established Underneath both sits a coverage fact that constrains everything computed from patents. At the 1851 Great Exhibition — 17,062 exhibitors from 40 countries, the founding display of industrial modernity — only about 11% of British exhibits were patented. Roughly nine in ten exhibited innovations at that event left no trace in the series most claims about patterns of innovation are computed from. Any pattern found in patents is, first, a pattern in patenting.

This is a measurement-validity finding, not a survivorship argument. Failed Technologies owns survivorship bias and documentation asymmetry — the way a sample selected on outcomes changes which features look causal — and this page references that rather than restating it. Moser's world's-fair number is a different object: an independent catalogue against which the coverage of a specific instrument, the patent register, can be quantified.

Established The field's own senior survey of its main instrument is unusually candid, and two of its findings should be carried in substance wherever patent series are used. Griliches, Patent Statistics as Economic Indicators: A Survey (1990), reports that the 1970s decline in US patents granted “is found to be an artifact of the budget stringencies at the Patent Office”; and that the long-run downward trend in patents per R&D dollar should be read “not as an indication of diminishing returns but rather as a reflection of the changing meaning of such data over time”. His verdict is nevertheless positive: “in spite of many difficulties and reservations, patent data remain a unique resource for the study of technical change”. A visible, much-discussed downturn in the headline innovation series turned out to be an administrative fact about the granting agency, and the person who found it went on recommending the series.

Established Moser supplies the independent check that patent series cannot supply for themselves, using catalogues of nineteenth-century world's fair exhibits, which record innovations whether or not they were patented. At the 1851 Crystal Palace, 17,062 exhibitors from 40 countries and about 11% of British exhibits patented — by industry, under 5% of chemical exhibits, about 8% in food processing, 10% of scientific instruments and 20% in manufacturing machinery, with prize-winners at 15%. The later fairs are comparable in scale: 1876 Philadelphia with 10,863 exhibits from 35 countries, 1893 Chicago, and 1915 San Francisco with 30,000 exhibitors from 32 countries. Moser's own statement of the limit is the one to quote: patent data are an “imperfect, fallible measure”, and “by construction, patent data fail to capture these innovations”.

Established The same catalogues answer a policy question the patent series cannot: what patent laws actually do. Countries without patent systems produced comparable numbers of exhibits and prize-winners, concentrated in industries where secrecy worked — so patent laws changed the direction of innovation rather than its level. After the Netherlands abolished patents in 1869, the Dutch share of innovations in food processing rose from 11% to 37% between 1851 and 1876. And compulsory licensing of German-owned chemical patents under the 1917 Trading with the Enemy Act was followed by a 20% increase in US invention in the affected technologies, with German inventors increasing patenting by 28% and high-value patents by 17% controlling for quality. The strongest historical evidence on patent strength says it redirects invention; it does not say it raises the rate.

Established The strongest “pattern” result in the field is an invariance across 220 years, and its unit of analysis is an administrative code. Youn, Bettencourt, Strumsky and Lobo analysed US patents from 1790 to 2010 using the USPTO technology-code system, treating patents as carriers of combinable technologies. In their words, the combinatorial inventive process “exhibits an invariant rate of 'exploitation' (refinements of existing combinations of technologies) and 'exploration' (the development of new technological combinations)” — a dynamic that “contrasts sharply with the creation of new technological capabilities…which has significantly slowed down”. And despite that slowdown in new building blocks, “the generation of novel technological combinations engenders a practically infinite space of technological configurations”. Frontier It is also the result most exposed to the objection that the pattern belongs to the classification scheme rather than to invention, since the codes are the patent office's own and have themselves been revised.

Frontier Simultaneous discovery is the leg that carries the most weight, and the modern numbers are both smaller and more interesting than the classic claim. Ogburn and Thomas's 1922 paper is the canonical list, usually reported as nearly 150 examples of multiple independent discovery — the article is paywalled and was not obtained, so the count here is second-hand. Modern measurements, as compiled in a curated living literature review rather than a peer-reviewed synthesis: Hagstrom's 1974 survey of 1,947 academics found 63% had been scooped at least once and 1.2% anticipated on the current project, implying about a 2.5% annual probability; Hill and Stein, on Protein Data Bank structures from 1999 to 2017, found 2.9% of proteins involved multiple independent discovery; Painter and colleagues found 197 of 3,488 new technical terms (5.6%) appearing simultaneously in evolutionary medicine, likely an overstatement; Bikard's twin-discovery pairs in PubMed give 10,927 pairs against 29.3 million papers, about 0.03%, with a citation filter that understates; and Ganguli, Lin and Reynolds, on 1,329 US patent interference decisions against 6.3 million granted applications, about 0.02%.

Frontier The most striking number in this brief is a convergence between two methods that share nothing. Hagstrom's 1974 survey estimate of roughly 2.5% a year and Hill and Stein's 2020 structural-biology measurement of 2.9% agree across nearly five decades and completely different designs — a self-report survey of academics and a structural database of solved proteins. That is a convergence, not a replication, and it is reported through a literature review rather than by either author. Established And the same review carries the evidence cutting the other way: simultaneity is concentrated in high-impact work, with a strong correlation between predicted citation impact and the number of groups working on a protein; a 2–3% annual rate implies a 54–67% probability that nothing gets published on an abandoned idea within twenty years; and early patent disclosure reduced subsequent infringing applications by only 5–15%, implying most apparent reinvention happens independently of available information. The review's conclusion is that “there is a lot of redundancy in the most important ideas and inventions, but not in the details”.

Frontier On whether returns are falling, three literatures disagree about their own measurements and the honest reading is a split verdict. Bloom, Jones, Van Reenen and Webb report research productivity falling “at a substantial rate in virtually every place we look”: aggregate US from the 1930s to 2015 at −5.1% a year, with total factor productivity falling by a factor of 41 while research effort rose by a factor of 23; agricultural yields from 1969 to 2009 at −9.9% a year for corn on the narrow measure and −6.2% broad, soybeans −7.3%/−4.4%, wheat −6.1%/−3.3%; clinical trials at −5.1% for all cancers, −10.1% for breast cancer and −7.2% for heart disease; and firm-level declines of −11.1% a year in Compustat over three decades and −7.8% in the Census of Manufacturing from 1992 to 2012. Established Their most quoted figure is the Moore's Law one and it is routinely misread: the number of researchers needed to sustain the doubling rate is “more than 18 times larger than the number required in the early 1970s”, a decline of about −6.8% a year over roughly 43 years. It is a ratio of research effort. It is not a claim about researcher quality, about total output, or about any other field.

Established The proposed mechanism is measured separately and agrees. Benjamin Jones, on USPTO patents from 1963 to 1999 with 2.1 million analysed in detail and age data on about 55,000 inventors, finds age at first invention rising about 0.66 years per decade — roughly 27–28 in the mid-1970s to 30–31 by the late 1990s; team size rising from 1.73 to 2.33 inventors per patent between 1975 and 1999, up 35% and consistent across all 36 technology categories; and field-switching among solo inventors falling about 6% per decade. The mechanism is that knowledge accumulation imposes an educational burden, with three adaptations: longer training, narrower expertise, and reliance on teams. Established His own four caveats are carried rather than dropped: the age sample is only about 7.5% of US inventors; there is acknowledged selection bias through zip-code availability; the specialisation measure is unreliable for team inventors because codes attach to patents rather than to people; and age at first invention is an imperfect proxy for education completion.

Frontier The third literature is the famous one and it is the cautionary tale. Park, Leahey and Funk reported in Nature that papers and patents are becoming less disruptive over time — 45 million papers, 3.9 million patents, six datasets, six decades, scored on the CD index, with declines from 1945 to 2010 of 91.9% in the social sciences and 100% in the physical sciences, and from 1980 to 2010 of 78.7% (computers and communications) to 91.5% (drugs and medical) for patents, attributed to a narrowing in the use of previous knowledge. Petersen, Arroyave and Pammolli then showed the index has a bounded numerator and an unbounded denominator containing an extraneous citation term, so lengthening reference lists drive it toward zero mechanically: mean references per paper rose from about 9 in the 1960s to about 23 in the 2000s, a factor of 2.6; the citation network grows 5.1% a year, doubling every 13.6 years; the log-reference-count coefficient is −0.0253; and their corrected team-size coefficient is positive, reversing a much-quoted result. Established The bibliographic details of the critique could not be confirmed — the fetched page returned an internally inconsistent venue and year — so it is cited by URL rather than by a reconstructed citation.

Frontier How the three fit together, stated honestly rather than stacked. The first two agree and use different data. The third is contested at the level of its instrument. A page that stacks all three as converging evidence for “innovation is slowing” is overclaiming; a page that dismisses all three because one is contested is underclaiming. The defensible statement is that two independent output-per-input measures decline and the citation-network measure is currently unusable without correction. Scientific Revolutions uses the same dispute for a different purpose — as a warning about bibliometric proxies for conceptual change — and the two pages do not re-run each other's argument.

Established And the strongest pro-pattern result in the field is real, quantitative, and much narrower than it is usually taken to be. Farmer and Lafond modelled the costs of 53 technologies as a correlated geometric random walk with drift, derived a closed-form forecast-error distribution, and showed by hind-casting that forecast errors across different technologies and horizons collapse onto a universal distribution — permitting genuine probabilistic prediction and comparison between competing technologies, demonstrated on solar photovoltaics. Frontier Note precisely what it is: cost trajectories of technologies already in production. It is not a method for predicting which technologies will exist, and nothing in the result licenses the fourth of the four claims above.

Frontier Do general-purpose technologies explain growth eras, or are they identified after the fact? Bresnahan and Trajtenberg define them by “pervasiveness (they are used as inputs by many downstream sectors), inherent potential for technical improvements, and innovational complementarities', meaning that the productivity of R&D in downstream sectors increases as a consequence of innovation in the GPT”. Their model's negative result is usually dropped and is the interesting part: because the relationship runs through vertical externalities, “if the relationship between the GPT and its users is limited to arms-length market transactions, there will be 'too little, too late' innovation in both sectors”. Frontier They close by claiming the analysis “has testable implications in the context of R&D and productivity equations, that can in principle be estimated”. “In principle be estimated” is a promissory note, and thirty years on the identification of which technologies are GPTs remains largely retrospective. The structural argument for why arms-length markets under-supply complementary innovation is developed as a method case in Bell Labs and DARPA, and this page does not restate either.

Frontier Does sustained growth require a widening epistemic base? Mokyr's framework distinguishes propositional knowledge — “generalized principles such as natural laws and empirical observations”, knowledge that — from prescriptive knowledge, “techniques, prescriptions, and instructions, which reside in human memory, artifacts or storage devices”, knowledge how. The epistemic base is the propositional knowledge underwriting a technique: a technique with a narrow base can be stumbled on but not systematically improved. His Industrial Enlightenment thesis is that self-sustaining growth required a feedback loop between the two plus falling access costs — postal and transport improvements, cheaper print, standardisation through mathematics and classification — and that earlier growth episodes failed because the base stayed narrow in agriculture, transport, power and medicine. Established Mokyr shared the 2025 Sveriges Riksbank Prize in Economic Sciences, his individual citation being “for having identified the prerequisites for sustained growth through technological progress” — reported here from an economists' commentary column rather than from the Nobel Foundation's own page, which was not fetched. Frontier A prize is a fact about a committee. The framework has no decisive test, and the specialist reviewing it argues that Britain's institutions restricted knowledge access to elites and that democratised access requires “institutions deliberately designed to ensure a process of democratization” rather than merely falling costs.

Frontier Is the innovation-centred history the wrong history? Edgerton's The Shock of the Old argues for a history told “in terms of what everyday people have actually used — and continue to use — rather than just sophisticated inventions”: celebrated innovations such as the V-2 and Concorde were expensive failures while corrugated iron had worldwide impact, and the book questions “innovation-centered historical frameworks that overstate technology's transformative power”. Established Those are the publisher's words in its own marketing copy, and the book itself was not read for this pass — it is cited for the existence and shape of the use-centred critique and for nothing more. Frontier The charge is precise and this brief has to answer it: patterns extracted from the record of invention may be patterns in a category that does not track what technology actually did.

Frontier The retrospective-fit problem is common to the three structural accounts and is the most serious general objection. General-purpose technologies are identified after they have been pervasive. Epistemic bases are diagnosed after the technique matured. Recombination is a description that fits any invention after the fact. None of these is refuted by the observation; all are weakened by it as predictive claims, which is the only version of the claim that would tell anyone where to invest. The objection is standard and, in this form, this page's.

Speculative And the strongest form of the inevitability claim is not supported by the numbers that are supposed to support it. Measured simultaneity rates run from 0.02% (patent interferences) to 5.6% (an overstated keyword measure), cluster around 2–3% a year in the two best estimates, and concentrate in high-impact work. Important things get discovered twice; details do not, and the arithmetic from a 2–3% annual rate gives a 54–67% chance that nothing gets published on an abandoned idea within twenty years. A claim that inventions are inevitable is therefore true for a small, identifiable, high-impact subset and false for the bulk of the record.

Handwave The claim that long-run patent series measure innovation is where this literature does its work by assertion. It is implicit wherever the series is used without correction, it is contradicted by the single best independent check available, and the field's own senior survey documents one famous trend in that series as an artefact of the granting agency's budget. The assertion is load-bearing for a very large fraction of quantitative innovation history and it is almost never defended.

Frontier The binding bottleneck is that the modern period has no independent register. Moser's world's-fair catalogues work because a nineteenth-century exhibition recorded innovations whether or not they were patented, and no comparable instrument exists after about 1915. For everything since, the patent register is checked against the patent register. That is why a coverage figure of about 11% at the founding event of industrial modernity is so hard to update: nobody knows the modern equivalent, and nobody has built the catalogue that would establish it.

Frontier The second is that the units of analysis are administrative artefacts that get revised. The recombination invariance is computed on USPTO technology codes; the disruption index is computed on citation links; patent counts are computed on grants, which depend on examiner capacity. Each is a real record and none is the thing the field is trying to measure, and the two documented failures on this page are both cases of an administrative property being read as a fact about invention.

Established The third is arithmetic and generalises beyond this subject. Any index with a bounded numerator over an unbounded denominator that grows with the record will trend toward its floor as the record grows. The citation network doubles every 13.6 years and reference lists lengthened by a factor of about 2.6 between the 1960s and the 2000s. Under those conditions a secular decline is the null hypothesis, not the finding, and the correction is a stated behaviour under record growth rather than a larger sample.

Frontier The fourth is that output-per-input measures inherit every problem of their chosen output proxy. Bloom and colleagues measure research productivity against transistor density, crop yields, trial outcomes and firm revenue growth — each defensible, each contestable, and none of them “ideas”. The result is robust across proxies, which is the strongest thing that can be said for it, and it remains a measurement of chosen outputs per counted researcher.

Frontier And the fifth is access. The canonical simultaneity paper from 1922 is paywalled and was not obtained for this pass, so its famous count is carried second-hand. Mokyr's book was not read; the framework is attributed through a specialist's review. Edgerton's book was not read; the argument is attributed through publisher's marketing copy. Three of the load-bearing conceptual claims in this subject reach this page at one remove, and each is marked at the point of use rather than in a footnote.

Established Nothing on this map produces a result this brief waits on, and no typed depends-on edge is claimed. It waits on a corrected disruption series that is a computation rather than a research programme, and on an independent non-patent register of invention for the modern period, which is an institutional undertaking rather than a discovery. Both are recorded as typed requirements below.

Established The boundaries with four neighbours are stated so nothing gets argued twice. Failed Technologies owns survivorship bias and documentation asymmetry, and this page references rather than restates them — Moser's coverage number is measurement validity against an independent catalogue, not selection on outcomes. Bell Labs and DARPA own industrial and agency research as method cases. Scientific Institutions Through History owns journals, academies and laboratories as institutions; where the German laboratory appears as an innovation engine this page defers to it. Scientific Governance Models owns present-day funding allocation and peer review; this page stops at the historical measurement.

Frontier The one genuine research dependency runs the wrong way for the framing. Every future quantitative claim about innovation history will be computed on the same patent and citation infrastructure that produced the two documented artefacts on this page. Until an index in that family is validated against an independent record of invention rather than calibrated on the assumption that it tracks one, results from that infrastructure cannot settle whether the patterns are in the world or in the register.

Frontier The cheapest high-value work is a computation whose specification is already published. Petersen and colleagues did not merely allege a bias in the disruption index; they named the mechanism and quantified it. Recomputing the CD series across the same six datasets with the correction applied would resolve the most-cited quantitative claim in the field one way or the other, and it requires no new data. Until it is published, innovation history has a headline finding and a specific arithmetic reason to distrust it.

Frontier Second: build a modern equivalent of the world's-fair catalogue. Trade show exhibitor lists, standards-body submissions, product-safety registrations, open-source release records and regulatory approval databases all record innovations independently of whether they were patented. Linking any one of them to the patent register would give the first modern estimate of what fraction of innovation the patent series captures — a number that currently exists only for 1851 and only for Britain, and that every long-run patent study implicitly assumes has not changed.

Frontier Third: settle the recombination invariance against a second classification. The exploration/exploitation invariance across 220 years is computed on USPTO technology codes. The European and international patent classifications are independently maintained and independently revised. If the invariance survives a change of classification scheme it is a fact about invention; if it does not, it was a fact about the scheme, and the data to check are already public.

Frontier Fourth: replicate the simultaneity convergence deliberately. Hagstrom's 1974 survey design and Hill and Stein's structural-database design agree at 2.5% and 2.9% a year without ever having been run against each other. Re-running the survey on a modern sample, alongside a structural-database measurement in the same period and field, would convert a striking cross-method convergence into an actual replication. Speculative It is the one place in this brief where two very different instruments already agree, and nobody has tested whether the agreement is real.

Speculative Fifth: test Edgerton's charge quantitatively rather than rhetorically. A use-centred history says the technologies that mattered are not the ones the invention record foregrounds. That is checkable: rank technologies by installed base, by hours of use, or by units in service, and compare against rankings by patent count, citation impact or historiographical attention. A large rank divergence would make the case; a small one would undercut it. Nobody has computed it, and the data for the twentieth century largely exist.

Frontier The engineering requirement in this subject is instrumentation for a record, and the specification is legible from the two failures. A usable long-run innovation series needs three properties none of the current series has together: coverage that can be quantified against an independent catalogue, which exists only for 1851; invariance to administrative change, which patent grants do not have because examiner capacity moves them; and a stated behaviour under record growth, which the CD index did not have.

Frontier The linkage problem is the practical obstacle to the second and it is solvable. Building a non-patent register means matching entities across trade catalogues, standards submissions, regulatory approvals and firm records — a record-linkage task at a scale that is now routine in economic history for censuses and much less common for artefacts. The nineteenth-century world's-fair work is the proof that the method produces a number; the missing part is the modern corpus, not the technique.

Frontier The third engineering fact is that classification schemes are versioned and their versions are published. Patent classifications carry dated revisions, concordances and reclassification records. Any invariance claim computed on a classification can therefore be recomputed on that classification's own history, which is the cheapest available test of whether a pattern belongs to invention or to the filing cabinet, and it has not been done for the exploration/exploitation result.

Frontier And the fourth constrains the forecasting result rather than extending it. Farmer and Lafond's method needs a cost time series for a technology already in production, which is why it covers 53 of them and not more. Extending probabilistic forecasting to a wider set is a data-collection problem — consistent, deflated, unit-normalised cost series — and extending it to technologies that do not yet exist is not a data problem at all.

Within this map: Scientific Revolutions, which shares the Park and Petersen dispute and uses it for a different purpose — as a caution about bibliometric proxies for conceptual change rather than as evidence about innovation rates; Scientific Institutions Through History, which owns the journals, academies and laboratories this page's inventors worked in; Philosophy of Science, which owns the question of what the measurements are measurements of; and Innovation Ecosystems, which is the present-tense version of this subject.

Also within it: Scientific Funding Models, where the declining-research-productivity finding is used as an argument; Scientific Governance Models, which owns present-day allocation and peer review; AI-Driven Productivity, which inherits the burden-of-knowledge mechanism directly; and Future Taxation Models, where patent-box and R&D-credit design rests on the assumption that patent counts track innovation.

Outside this map: Failed Technologies, which owns survivorship bias and documentation asymmetry; Bell Labs and DARPA, which own industrial and agency research as method cases. And beyond the site: economic history, which supplies the world's-fair evidence; scientometrics, which supplies the indices and their failures; and the economics of ideas, which supplies the productivity series.

Frontier The most consequential institutional fact in this subject is that the measuring instrument is an administrative agency with a budget. A patent series records grants, grants depend on examination capacity, and examination capacity depends on appropriations. Griliches found the 1970s decline in US grants to be an artefact of Patent Office budget stringency. No comparable audit has been published for any later period, and the same coupling has operated continuously since.

Frontier The second is that the field publishes indices faster than it validates them. The CD index reached 45 million papers and the cover of a general-science journal before anyone checked its behaviour under the largest secular trend in its own input data. What is missing is not a better index but a convention: no innovation or disruption index should be published without a stated behaviour under reference-list and network growth. That convention would have caught this one.

Established The interested parties and the substitutions are named where they matter. Mokyr's framework reaches this page through a specialist's book review rather than through his book; the 2025 prize citation through an economists' commentary column rather than the Nobel Foundation's own page; Edgerton's argument through a publisher's marketing copy. Two economics papers are cited from author-hosted copies at Stanford, Kellogg and NYU Stern, and one from the NBER working-paper record, because the publisher domains carried a stale-content warning and were deliberately avoided. The simultaneity compilation is a curated living literature review and is flagged as one rather than as a peer-reviewed synthesis.

Frontier And the policy consequence of the coverage fact is institutional rather than academic. Patent boxes, R&D tax credits, technology-transfer targets and university rankings all use patent counts as an innovation measure. If about 11% of exhibits at the founding event of industrial modernity were patented, and if patent laws changed the direction of invention rather than its level, then a policy instrument keyed to patent counts is selecting for patentability rather than for innovation. That is a testable claim about incentives and this pass located no test of it.

Frontier The strongest historical evidence on patent strength is not the evidence the policy debate uses. Countries without patent systems produced comparable numbers of world's-fair exhibits and prize-winners, concentrating in industries where secrecy worked; the Netherlands abolished patents in 1869 and its share of food-processing innovations rose from 11% to 37% by 1876; and compulsory licensing of German chemical patents in 1917 was followed by a 20% increase in US invention in the affected technologies. None of that says patents are bad. It says the level of invention is less sensitive to patent law than the direction of invention is, which is a different and much more actionable finding than either side usually claims.

Established The distributional question inside Mokyr's framework was raised by a sympathetic reviewer and is the sharpest criticism of it. If sustained growth turns on falling access costs to knowledge, then who has access is the mechanism rather than a side effect — and the argument is that Britain's institutions restricted access to elites, so democratised access requires “institutions deliberately designed to ensure a process of democratization” rather than merely cheaper print and better post. Frontier That reframes a story about technology as a story about who is allowed to use it, and it is a live disagreement between two economic historians rather than a settled qualification.

Frontier The simultaneity evidence has an ethical edge that is usually stated backwards. Because redundancy concentrates in high-impact work and not in details, the cost of a priority dispute falls on the small number of cases where two groups converge, while the cost of abandonment falls on the large number where nobody else is looking. Clancy's arithmetic — a 54–67% chance that nothing gets published on an abandoned idea within twenty years — means most unfunded work is simply not done by anyone, which is an argument about funding breadth rather than about credit.

Frontier And measuring innovation badly has consequences beyond scholarship. Two of the largest quantitative claims about the health of the innovation system in the last forty years turned out to be substantially properties of recording apparatus, and both were used in policy argument before they were audited. The ethical requirement is not caution about publishing; it is that an index used in policy should carry a stated failure mode, and neither of these did.

Frontier The terminal position is a split verdict and it should be stated as one. Descriptive regularities are real and partly measured — cost trajectories of 53 technologies collapse onto a universal forecast-error distribution, and an exploration/exploitation invariance holds across 220 years of the US patent record. Inevitability is real for a small, high-impact subset at rates around 2–3% a year. Structural accounts of what invention is are influential, retrospectively identified, and untested as predictions. And the fourth claim — that history tells you where to invest — is not supported by anything on this page.

Frontier The most transferable finding is about instruments rather than about innovation. A field measured its own health with two long series, and each produced a famous historical trend that turned out to be a property of the recording apparatus: a granting agency's budget in 1990, reference-list growth in 2024. The pattern is not that these particular series were bad. It is that a civilisation-scale claim computed from a growing administrative record inherits the record's history, and the defence is an audit of the instrument rather than a larger sample.

Frontier If the declining-returns finding survives, the implication is about the composition of research effort rather than about limits. Two independent output-per-input measures decline consistently, and the mechanism proposed and separately measured is the burden of knowledge: longer training, narrower expertise, larger teams. Those are adaptations that work, and they have costs — a rising age at first invention and falling field-switching are the signature of a system that gets more from specialists and less from generalists. Whether that trade is favourable at civilisational scale is not measured by anything here.

Speculative And the coverage problem gets worse rather than better. The share of technical work embodied in software, process knowledge, trade secrets, datasets and models is rising, and none of it is captured by the patent register in the way a machine tool was. If the 1851 coverage figure of about 11% was low for an era of physical artefacts, the modern figure for the categories that matter most may be lower still — and nobody has measured it, because the instrument that would measure it does not exist.

Handwave The strongest version of the framing — that historical patterns can guide investment prospectively — does the work by assertion and is worth naming as such. The best forecasting result in the literature covers cost curves of technologies already in production; general-purpose technologies and epistemic bases are diagnosed after pervasiveness; recombination fits any invention after the fact. There is no result here that identifies a technology worth backing before it is already visible, and the literature routinely gets cited as though there were.

These horizons track what could be measured and when, not what will be invented:

• 10 yr: Frontier The disruption dispute resolves inside this window or not at all: the correction is specified, the datasets are the same, and the computation is cheap. Frontier A modern non-patent register is technically feasible now and institutionally unclaimed — the most likely route is a national statistical office linking regulatory approvals or standards submissions to patent records for a single sector. Speculative Expect the recombination invariance to remain unchecked against a second classification, because reproducing a twelve-year-old result on European codes is not a fundable project.
• 25 yr: Speculative If the declining-returns series continue at measured rates the aggregate US figure implies research effort roughly doubling again for the same output growth, and the interesting question becomes whether the burden-of-knowledge adaptations — longer training, larger teams — have a ceiling. Speculative A plausible split: forecasting improves substantially for cost trajectories of production technologies and not at all for which technologies arrive, because those are different problems and only one of them has data.
• 50 yr: Speculative The coverage question is the one that plausibly changes shape: if most technical work by then is embodied in artefacts the patent system does not describe, long-run innovation series either acquire a new register or become uninterpretable. Handwave Which of those happens is an institutional prediction with no evidence behind it.
• 100 / 250+ yr: Handwave Beyond useful forecasting. The one anchor at that horizon is that the instrument this field depends on is about 400 years old and has been audited seriously perhaps twice. Handwave Two audits is not a rate.

• Depends on Nothing on this map. No brief here produces a result this one waits on, and no typed depends-on edge is claimed. The constraints that bind are a computation nobody has published and a register nobody has built, both typed below.
• Requires (not on this map) A disruption series recomputed with the published correction for reference-list growth. The critique is not an allegation but a specification: a bounded numerator over an unbounded denominator carrying an extraneous citation term, against mean references per paper rising from about 9 in the 1960s to about 23 in the 2000s, a citation network growing 5.1% a year and doubling every 13.6 years, a log-reference coefficient of −0.0253, and a corrected team-size coefficient that reverses sign. Applying it across the same six datasets is arithmetic on public data, and until someone publishes the corrected series the most cited quantitative claim about innovation rates cannot be used or discarded. It is typed as a missing result rather than an institutional gap because no permission, budget or law is standing in the way. And an independent non-patent register of invention for the modern period: the only quantified check on what the patent series captures comes from nineteenth-century world's fair catalogues — about 11% of British exhibits patented in 1851, under 5% in chemicals — and no comparable instrument exists after about 1915, so every modern long-run patent study implicitly assumes a coverage rate nobody has measured. Building one means linking trade-show exhibitor lists, standards submissions, regulatory approvals or product registrations to the patent record at scale. The record-linkage technique is routine; what is missing is an institution willing to assemble and maintain the corpus, which is why this is typed as institutional rather than scientific.
• Enables Innovation Ecosystems, Scientific Funding Models and AI-Driven Productivity all rest on quantitative claims about innovation rates that are computed from the series audited here. No typed enabling edge is claimed, and the reason is a finding rather than an omission: on the evidence assembled here the two longest series each carry a documented artefact, so nothing should be built on their headline trends without the corrections named below.
• Adjacent Economic history, which supplies the world's-fair evidence; scientometrics, which supplies the indices and their failures; the economics of ideas, which supplies the productivity series; and on this site Failed Technologies, Bell Labs, DARPA and Scientific Revolutions.

Established “Patent counts measure innovation.” Roughly nine in ten exhibited innovations at the 1851 Great Exhibition were unpatented — about 11% of British exhibits overall, under 5% in chemicals, 8% in food processing, 10% for scientific instruments — and Moser's own statement is that patent data “by construction…fail to capture these innovations”. Handwave The assumption is nevertheless load-bearing for a large fraction of quantitative innovation history and is almost never defended.

Established “The 1970s decline in patenting shows innovation slowed.” The field's own senior survey found it “to be an artifact of the budget stringencies at the Patent Office”, and read the long-run fall in patents per R&D dollar as reflecting “the changing meaning of such data over time” rather than diminishing returns. This is the cleanest documented case in the field of a famous historical trend belonging to the recording agency.

Established “Stronger patent protection raises the rate of innovation.” The best historical evidence says it changes the direction. Countries without patent systems produced comparable numbers of exhibits and prize-winners, concentrated where secrecy worked; the Netherlands abolished patents in 1869 and its food-processing share rose from 11% to 37% between 1851 and 1876; and compulsory licensing under the 1917 Trading with the Enemy Act was followed by a 20% increase in affected US invention, with German inventors increasing patenting by 28%. Frontier The direction result is well documented; the level claim has no comparable support in either direction.

Frontier “Innovation is slowing, and three literatures prove it.” Two output-per-input measures decline consistently on different data. The third and most famous is under a specific, quantified methodological attack that also reverses its team-size result. Frontier The defensible statement is that two independent measures decline and the citation-network measure is currently unusable without correction — stacking all three is overclaiming and dismissing all three because one is contested is underclaiming.

Established “It takes 18 times as many researchers to do what one did in 1971.” The figure is real and means something narrower than it is used for: the number of researchers required to sustain Moore's Law's constant doubling rate is “more than 18 times larger than the number required in the early 1970s”, a decline of about −6.8% a year over roughly 43 years. It is a ratio of research effort in one industry. It is not a claim about researcher quality, about total output, or about any other field.

Frontier “Simultaneous discovery shows inventions are inevitable.” Measured rates run from 0.02% for patent interferences to 5.6% for an overstated keyword measure, cluster around 2–3% a year in the two best estimates, and concentrate in high-impact work. Established Details are not rediscovered, and the arithmetic from a 2–3% annual rate gives a 54–67% probability that nothing gets published on an abandoned idea within twenty years. Frontier The Hagstrom–Hill and Stein convergence at 2.5% and 2.9% is striking and is a cross-method agreement reported in a literature review, not a replication.

Frontier “General-purpose technologies and epistemic bases can be spotted early.” Both are diagnosed after pervasiveness. Bresnahan and Trajtenberg claim only that their analysis has testable implications that “can in principle be estimated”; Mokyr's epistemic base is identified once a technique has matured. Speculative Recombination has the same problem in stronger form: it is a description that fits every invention after the fact, which is why the invariance result matters and the framework does not.

Frontier “Technological forecasting works.” Farmer and Lafond genuinely forecast, and what they forecast is cost trajectories of 53 technologies already in production, with errors collapsing onto a universal distribution. Established Nothing in that result predicts which technologies will exist, and the gap between those two things is the whole of the fourth claim this page separated out in section 1.

Frontier “The recombination invariance is a fact about invention.” Its unit of analysis is the patent office's own technology-code system, which is an administrative artefact that has itself been revised. Speculative The test is cheap and public — recompute on an independently maintained classification — and it has not been run, which means the strongest pattern result in the field currently cannot be distinguished from a property of a filing scheme.

Frontier “A Nobel-recognised framework is established.” Mokyr shared the 2025 economics prize with a citation for identifying the prerequisites for sustained growth through technological progress, reported here from a commentary column because the Nobel Foundation's own page was not fetched. Established A prize is a fact about a committee. The propositional and prescriptive knowledge framework is influential, has no decisive test, and is criticised by the specialist who reviewed it on the ground that access requires deliberately designed institutions rather than merely falling costs.

Frontier And the framing itself: “history reveals patterns of innovation” assumes the record is a window. What history reveals first, in this field, is the shape of its own instruments — a granting agency's budget, a classification scheme's revisions, a citation network's growth rate. Frontier That does not make the patterns false. It means every pattern claim here has to survive a question about the register it was computed from, and the field's most careful practitioners are the ones who keep finding that some do not.