Innovation Ecosystems

The framing under test is that innovation can be cultivated by policy. It is not wrong. Some innovation policies have credible causal evidence that they work, and the evidence is better than most people arguing about industrial policy realise. The problem is the ordering. Rank the instruments by identification quality and the list comes out almost exactly inverted against the attention they receive.

At the top are a regression discontinuity on a tax-credit eligibility threshold and a quasi-random rigidity inside a grant-review committee — boring administrative machinery, with elasticities and patent counts attached. In the middle are investment subsidies identified off pan-European state-aid rules, which buy jobs cheaply and buy no measured productivity at all. Near the bottom is government venture capital, whose effect is non-monotonic in dosage and turns negative at high government funding shares. At the very bottom, with no credible design anywhere in the fetched corpus, is building a cluster — which is what a minister announces.

The single finding that most complicates the framing is the founding evaluation of the archetypal programme. Lerner's study of the US Small Business Innovation Research programme found awardees grew significantly faster than matched firms over ten years, and that the positive effects were confined to firms in zip codes that already had substantial venture capital activity. The policy's effect is conditional on the thing the policy is meant to create. That is not a caveat; it is close to a circularity, and it recurs in two further independent studies.

Established State the stratification first, because everything below is a position on it. Ranked by the credibility of the design behind them: R&D tax credits — regression discontinuity on eligibility thresholds, large elasticity, works. Public research funding — quasi-random peer-review rigidities, measurable private spillovers, works. Investment subsidies — instrumented off exogenous state-aid rule changes, works for small firms with no productivity effect. Public R&D grants to small firms — discontinuity on applicant rankings, real effects on employment, revenue and wages. Government venture capital — instrumented international panel, non-monotonic and harmful at high dosage. Enterprise zones and place-based tax — very mixed. Cluster creation and ARPA-model replication — no credible design exists. Frontier The ranking is this brief's synthesis rather than a claim any single source makes; each row is a fetched study or a fetched survey's own assessment.

Established The one place identification is genuinely clean is R&D tax relief, and the design is directly testable. Dechezleprêtre, Einiö, Martin, Nguyen and Van Reenen exploit a change in asset-based size thresholds for UK R&D tax subsidy eligibility as a regression discontinuity. They find an R&D tax-price elasticity of about 2.6 — a 1% fall in the after-tax cost of R&D raising R&D spending by 2.6% — patenting up about 60% among affected small and medium firms with quality-adjusted measures also rising, no pre-policy manipulation of assets around the thresholds when tested directly, and positive spillovers onto the innovation of technologically related firms. Established The spillover result is what distinguishes real innovation from accounting relabelling, and it is the reason this sits at the top of the list. Frontier The caveats are real: the elasticity is estimated at a threshold, on small firms, in one country under one tax regime, and an elasticity of 2.6 on R&D spending is not an elasticity of 2.6 on innovation. The patent result is the bridge, and patents are a noisy proxy.

Established Public research funding produces private-sector patents at a measurable and modest rate. Azoulay, Graff Zivin, Li and Sampat exploit idiosyncratic rigidities in NIH peer-review funding rules to get exogenous variation in funding across research areas, and link specific grant expenditure to firm patents bibliometrically. An additional $10 million in NIH funding produces roughly 2.3 additional private-sector patents; a related summary of the same work reports 3.1. Frontier At the macro level, Fieldhouse and Mertens estimate large social returns to federal nondefense R&D, comparable to estimated returns to private R&D, and project that fully appropriating the CHIPS and Science Act's R&D increases would raise US productivity 0.2–0.4% after seven years with peak output effects exceeding $40 billion annually. Frontier Time-series identification on federal appropriations is much weaker than the NIH discontinuity, and a projection is a projection. The clean result is small and real; the large result is a forecast.

Frontier Investment subsidies buy jobs cheaply and buy no productivity, and the literature routinely cites the first while implying the second. Criscuolo, Martin, Overman and Van Reenen study the UK's Regional Selective Assistance programme using an instrument from pan-European state-aid rules that change area eligibility every seven years — genuinely exogenous variation in who qualifies. Results: positive effects on employment, investment and net entry, and none on total factor productivity. Effects concentrate in firms under 150 employees, with no significant effect for large firms, at a cost per job of roughly $6,300. Handwave Read that carefully. The programme worked, at low cost, at moving employment. It did not make anything more productive. If cultivating innovation means raising productivity, this is a null; if it means creating jobs in lagging regions, it is a success. Substituting one for the other is the step the policy argument does by assertion.

Frontier Grants to small firms do something real, and what they do is pay incumbent employees. Howell and Brown use a discontinuity on applicant rankings for government R&D grants and find the grants raise employment and revenue, with a rent-sharing elasticity of 0.07 at employee level and 0.21 at firm level, benefits concentrated in incumbent employees present before the award and rising with tenure. Their interpretation is a backloaded wage contract: constrained firms underpay early and true up when cash arrives. Established The design is clean and the finding is about wages rather than innovation. It is carried here because it establishes that the grants did something real to the firm, which is more than can be said for several instruments further down the list.

Frontier Public venture capital is where the evidence most directly contradicts the rationale offered for it. Lerner's founding SBIR evaluation found awardees grew significantly faster than a matched set of firms over ten years — with the positive effects confined to firms in zip codes with substantial venture capital activity. Frontier Brander, Du and Hellmann, on an instrumented international panel, find enterprises funded by both private and government venture capital outperform purely private-backed benchmarks when only a moderate fraction of funding is governmental and underperform when a large fraction is, with underperformance concentrating in periods of abundant private capital — and the outperformance applying only to funds supported but not owned outright by governments. Established Bai, Bernstein, Dev and Lerner map 755 public entrepreneurial finance programmes worldwide and find government programmes involve private investors more where state capacity is stronger, where the target is earlier stage, and where private venture markets are already well-developed. Frontier Three independent studies, one shape: public innovation finance is a multiplier on an existing ecosystem, not a seed for a new one. That is the exact inverse of the rationale usually offered for it.

Established On clusters, the surveys say plainly what the policy advice does not. Chatterji, Glaeser and Kerr, reviewing the spatial concentration of entrepreneurship and innovation, write that “our understanding of what works and how it works is quite limited” and recommend “extensive experimentation and careful evaluation”. Frontier Kerr and Robert-Nicoud's survey explicitly poses “the degree to which policy makers can purposefully foster them” as an open question; only its abstract could be fetched here, so its own answer is not characterised. Established Neumark and Simpson's review of place-based policies reports that “the evidence on enterprise zones is very mixed”, and finds more credible evidence for infrastructure spending and investment in higher education and university research — attributing that to the public-goods nature of those policies, which is a mechanism being understood rather than a research design being stronger.

Established And the agglomeration effects the whole edifice rests on are real and small. Moretti and Yi, using job displacement on administrative data, find a 1% increase in labour market size raises the one-year re-employment probability by 0.015 percentage points for high school graduates and 0.023 points for college graduates, with shorter non-employment spells, smaller earnings losses, less need to relocate and better match quality. Established A policy that doubled a local labour market — which no policy has ever done — would move re-employment probability by low single-digit percentage points. Frontier Giroud, Liu and Mueller find larger tech clusters raise the productivity of inventors and plants in other clusters, through corporate networks, with connectivity mattering more than physical distance. If the spillover travels through firm networks rather than through geography, the case for geographic cluster policy specifically is weakened by this result rather than strengthened, which is not how it is usually cited.

Speculative Can clusters be created at all, or only accelerated where the preconditions exist? The evidence for the second reading is the surveys' own admissions, the SBIR regional restriction, the government-venture dosage effect and agglomeration effects that are real but small. Speculative The competing position — held by cluster-development agencies and much of the regional-policy profession — is that the null results are underpowered rather than true, and rests on case studies that are mostly self-evaluated. Established What would settle either is a pre-registered evaluation with a credible counterfactual region. Forty years of cluster policy has not produced one.

Frontier Do the spillovers travel through places or through firms? The cross-cluster spillover evidence says corporate-network connectivity dominates physical distance. Frontier This is the most consequential recent result for cluster policy and the least discussed, because it relocates the policy target from a map to an ownership structure — and nobody knows how to subsidise an ownership structure. Speculative If it holds, then attracting a branch plant of a firm already inside a strong cluster does more for a region than co-locating unrelated firms, which is the opposite of the usual innovation-district design.

Frontier Can venture capital move a cluster, or only follow one? Chen and Ewens exploit a rule change that disproportionately affected US regions historically lacking venture financing: when venture capitalists faced funding restrictions they reduced investment, and startups moved out of the affected states. Their reading is that venture capital causes agglomeration rather than following it. Frontier This is the closest thing in the corpus to evidence that a financial intervention can shift where a cluster is — and it establishes it in the destructive direction, which is much easier. One rule change in one country; the identification is real and the asymmetry should not be glossed over.

Handwave Can the state direct technological development, or only fund it? The descriptive core of the entrepreneurial-state argument is well corroborated: public funding demonstrably produces private-sector patents at a measurable rate, social returns to public nondefense R&D are large, and publicly funded research is upstream of private innovation. Handwave The directional claim — that the state can choose which technologies to back and be right — is not tested by any fetched source. The NIH result is about funding volume in areas selected by peer review, not about political direction-setting. Frontier The return-capture claim runs directly into the government-venture evidence: high government funding shares are associated with underperformance, and state-owned rather than state-supported vehicles do worse. Handwave And the counterfactual — that the innovation would not have happened otherwise — is the hardest thing in the field to establish and is established nowhere in this corpus.

Established Has industrial policy actually been rehabilitated, and if so by what? Juhász, Lane and Rodrik's survey concludes that recent work is “a significant improvement over the earlier generation of empirical work, which was largely correlational and marred by interpretational problems” and offers “a more positive take on industrial policy” — while emphasising that the reassessment is grounded in improved empirical technique rather than ideological change. Frontier That is a genuine, recent and credentialed shift in professional opinion and it should be reported as such rather than dismissed. It is also not a rehabilitation of cluster-building, which is the instrument the improved techniques have least to say about.

Frontier Is the binding constraint talent, and is the public research base losing it? Akcigit and co-authors track roughly 42,000 AI specialists over two decades and find the top 1% of publishing industry scientists now earn about $1.5 million more annually than comparable academics, a fivefold increase since 2001, with movers to industry publishing less but patenting more. Frontier If the public research base is the mechanism through which policy cultivates innovation — which is what the two best-identified instruments in this brief both operate through — and that base is being drained at the top, then every other instrument here weakens simultaneously. This brief's assessment is that it is the most important under-discussed hypothesis in the subject.

Speculative And the fringe reading, stated because it is what the stratification looks like taken entirely at face value. On this view the field measures programme outputs because programme outcomes are unmeasurable, and the enterprise is closer to ritual than to policy. Its strongest evidence is that the defining paper on the ARPA model remains definitional after decades of ARPA-for-X agencies being founded on it, and that cluster policy has run for forty years without a single credible counterfactual. Frontier The counter is the top of the stratification: two instruments with clean designs and real measured effects. The fringe reading is wrong about the whole field and uncomfortably close to right about its most-announced half.

Established The binding methodological constraint is the absence of a counterfactual region. Every credible design in this brief exploits an administrative rule that assigns treatment nearly at random: an asset threshold, a peer-review rigidity, an applicant ranking, a state-aid eligibility rule that changes on a seven-year cycle. Cluster policy has no such rule, because a cluster programme is assigned to the place a government wants to help. Frontier That is not an oversight to be fixed by better econometrics; it is a property of the instrument. The policies with the best evidence have it because their eligibility rules were drawn for administrative convenience, not because anyone designed an evaluation.

Established The second bottleneck is that programme outputs are measured and programme outcomes are not. Patents, grants awarded, firms assisted, jobs created within the programme boundary. Patents are a noisy proxy for innovation, and an elasticity on R&D spending is not an elasticity on innovation. Frontier The one design that bridges the gap does so with a spillover result — innovation rising at technologically related firms that received nothing — and that bridge is what makes it the strongest evidence here. Almost nothing else in the corpus crosses from spending to consequence.

Established Third: the evaluations that do exist for the most-announced instruments are written by the agencies running them. National cluster-agency self-evaluations, and the OECD and EU cluster-excellence programme reviews, could not be fetched in this pass and are named without citation. Established They are interested parties evaluating their own programmes, and wherever they are eventually cited they must be marked as such. Frontier This is not an accusation of bad faith. It is a statement that an evidence base composed of self-evaluations has a known and large bias, and that no independent evaluation exists to bound it.

Established Fourth: the ARPA model's defining paper is conceptual. Azoulay, Fuchs, Goldstein and Kearney identify the model's features — organisational flexibility and substantial authority delegated to programme directors — and locate its domain as “mission-oriented research on nascent S-curves within an inefficient innovation system”. Established It is definitional and does not present empirical evidence that the features cause the outcomes. Given how many ARPA-for-X agencies have been founded on the strength of it, that is worth stating plainly rather than as a footnote.

Frontier And fifth, the gap this brief most wanted to fill and could not: the record of deliberate Silicon Valley replication. No fetched source systematically evaluates the attempts — not Skolkovo, not the various national “Silicon” initiatives, not Malaysia's Multimedia Super Corridor, not the EU cluster-excellence programmes. Established That is not evidence they failed. It is evidence that the most confidently offered policy advice in this subject rests on a literature this pass could not locate in peer-reviewed form, and stating the absence is more useful than filling it with anecdotes.

Established Nothing on this map produces a result this brief waits on, and no typed depends-on edge is claimed. What it waits on is one market condition and two institutional ones, recorded as typed requirements in section 13, and none of them is a discovery.

Frontier The first is a market condition, and it is the most important sentence in this brief. Public innovation finance works as a multiplier on a private venture market that already exists and works badly or not at all where one does not. Three independent studies converge on it: the SBIR effects confined to venture-rich zip codes, the non-monotonic dosage effect turning negative at high government funding shares, and the mapping showing public–private co-investment rising with private market development. Established That makes a functioning private venture market a precondition for the instrument rather than an objective of it, and it is a market structure rather than a research result.

Established The second is an evaluation design that does not exist: a pre-registered counterfactual region. Cluster programmes are assigned to the places governments want to help, which destroys the comparison. Committing in advance to a comparison region — by lottery among eligible places, by a scored eligibility threshold, or by staggered rollout — converts forty years of announcements into evidence. Frontier Nothing about it is technically hard. It requires a government to accept that some eligible places will be told to wait.

Frontier The third is a public research payscale that can retain the researchers the public instruments run through. Both of the best-identified instruments here operate by putting money into a research base and letting it spill over. If the top of that base is leaving — a $1.5 million annual premium for top industry scientists, up fivefold since 2001, with movers publishing less and patenting more — then the spillover channel narrows while the instruments stay the same. Speculative No fetched source measures the elasticity of the spillover to the composition of the public research base, so the size of this dependency is unknown and its sign is not.

Established The highest-value experiment is the one the field's own surveys have been asking for and nobody has run: randomise or quasi-randomise a cluster intervention. Eligible regions exceed available funding in every cluster programme in existence, which means a lottery among the eligible is administratively available and ethically defensible. Frontier A single such programme, pre-registered, would produce more usable knowledge than the entire self-evaluation literature. Its absence after forty years is a fact about political tolerance for being measured, not about method.

Established Second: replicate the R&D tax-credit discontinuity at other thresholds and on large firms. The elasticity of 2.6 is estimated at one threshold, on small firms, in one country under one tax regime, and it is being used everywhere. Frontier Most tax systems contain multiple size and expenditure thresholds with the same discontinuity structure, and the data are administrative. This is a replication programme, not a research frontier, and it would tell every finance ministry using this instrument whether the number they are citing applies to them.

Frontier Third: evaluate the state's technology bets including the failures. The entrepreneurial-state argument's directional claim requires an assessment of state technology choices with a counterfactual, and the case-study literature on both sides selects on the outcome — successes for the proponents, failures for the critics. Speculative A complete portfolio of a single agency's directed bets over a defined period, scored against a matched set of undirected allocations, is a hard but finite piece of archival work that nobody has done.

Frontier Fourth: randomise the ARPA features rather than founding agencies on them. Programme-director autonomy, portfolio flexibility and termination authority are separable and assignable within a single funding agency. Speculative A design that varies delegated authority across otherwise comparable programme portfolios would convert the model's defining paper from a description into a test, and the agencies that have adopted the model are exactly the ones positioned to run it.

Frontier Fifth: test the connectivity-versus-distance result as a policy instrument. If cross-cluster spillovers travel through corporate networks, then a region's return from attracting an establishment of a firm already embedded in a strong cluster should exceed its return from co-locating unconnected firms of the same size. Speculative Regional investment programmes make this choice constantly and record enough to test it retrospectively, and no fetched source has.

Established And sixth, cheapest of all: publish the eligibility rules. Every credible design here exists because an administrative threshold happened to be documented. Programmes that publish their scoring rules and cut-offs in advance generate their own evaluations at zero marginal cost; programmes that do not are unevaluable by construction. Frontier That is a procurement and transparency decision, and it is the single highest-leverage thing a government could do for this evidence base.

Established The working instruments in this brief are all pieces of administrative machinery, and their design details are the mechanism. An asset-based eligibility threshold. A peer-review rule rigid enough to be quasi-random. A numeric applicant ranking with a funding cut-off. An area-eligibility rule set supranationally on a seven-year cycle. Frontier Each works as policy for the same reason it works as identification: it applies a rule rather than a judgement, which limits the discretion that would otherwise be exercised in favour of the applicants best able to lobby for it.

Frontier The dosage engineering on public venture capital is unusually well specified and unusually ignored. The measured pattern is that co-funding outperforms at a moderate government share and underperforms at a high one, that the underperformance concentrates when private capital is abundant, and that support without outright ownership beats ownership. Established Those are three design parameters with signed effects — share, timing and control — and agencies routinely cite the positive half of the finding while setting all three the other way.

Established The cost side is knowable and rarely quoted. Roughly $6,300 per job for the regional investment subsidy with no productivity effect; roughly 2.3 private-sector patents per $10 million of public biomedical research funding. Frontier Those two numbers are in different units and cannot be compared directly, which is precisely the problem: the field has no common denominator, so instruments are chosen on narrative rather than on cost-effectiveness. Constructing one would be a substantial contribution and nothing in the fetched corpus attempts it.

Frontier And the place-based rehabilitation, read carefully, endorses a different machine than the one usually built. Austin, Glaeser and Summers argue against the traditional dismissal of place-based policy on the grounds that regional convergence has slowed, long-term non-employment is diverging across areas, male non-employment in the eastern US heartland has nearly tripled over fifty years, and — the analytical core — labour demand increases have larger employment effects in areas where non-employment has historically been high. Established Their recommendation is a regionally varying earned income tax credit. The strongest recent argument for place-based policy is an argument for a wage subsidy in depressed labour markets, not for innovation districts, cluster agencies or public venture funds. Frontier The heterogeneous-elasticity claim that carries the argument is asserted more than demonstrated in what could be fetched here.

Within this map: Future Labour Markets, which measures what happens to the workers the place-based instruments are aimed at; AI Driven Productivity, whose intangible-capital problem is why R&D spending and innovation output are so hard to connect; Digital Economies, where the market structures that concentrate innovation rents are assessed; Scientific Governance Models and Scientific Advisory Institutions, which own the machinery of research funding and advice; Artificial Scientists, on what happens to the research base's productivity itself; Megaproject Governance, whose evaluation problems are the same ones one tier up; and Future Public Administration, whose finding that evaluation is discretionary explains why forty years of cluster policy produced no counterfactual.

Outside it: the economics of innovation and the market-failure framework that justifies intervention; urban and regional economics, which supplies the agglomeration estimates; entrepreneurial finance, which supplies the venture-capital evidence; public finance, which supplies the tax-credit designs; and the new industrial-policy empirics, which is the live methodological argument this subject sits inside.

Established The institutional constraint that binds hardest is that the most-announced instruments are the least evaluable, and this is structural rather than accidental. A tax credit has a threshold because tax law needs one. A grant committee has rigidities because it processes applications at scale. A cluster programme has neither, because its whole purpose is to be directed at a chosen place. Frontier Evaluability is therefore correlated with how impersonal the instrument is, and the instruments politicians most want are the ones where discretion is the point.

Established Interested parties are unusually concentrated at the weak end of the evidence. National cluster agencies evaluating their own programmes, and the supranational cluster-excellence reviews, are the principal evidence base for the instrument with no credible design. None could be fetched here; all are named without citation and must be marked wherever they are eventually used. Established At the strong end the pattern reverses: the R&D tax-credit and NIH results come from academic teams with no stake in the programmes, and the investment-subsidy evaluation reports a null on the outcome its commissioning context would most have wanted.

Frontier The talent constraint is institutional and is being lost quietly. A $1.5 million annual pay gap at the top of the AI research distribution, up fivefold since 2001, with movers to industry publishing less and patenting more, is a transfer of the public research base's most productive people into settings where their output is less disclosed. Speculative Whether that reduces the spillover the public instruments rely on is untested. That it changes the disclosure regime of the resulting knowledge is measured, and disclosure is the channel every spillover estimate in this brief runs through.

Established And the professional-opinion shift is itself an institutional fact worth recording accurately. A survey by three prominent economists finds recent industrial-policy work a significant improvement on an earlier generation that was largely correlational, and takes a more positive view as a result — explicitly attributing the reassessment to technique rather than ideology. Frontier That is a real change and it is routinely cited in support of instruments the improved techniques have not been applied to. Better econometrics rehabilitated the instruments that could be evaluated, which is not the same as rehabilitating industrial policy in general.

Established The distributional case for intervening is stronger than the instrumental case for the usual instruments, and the two are constantly merged. Regional convergence has slowed, long-term non-employment is diverging across areas, and male non-employment in the eastern US heartland has nearly tripled over fifty years. Frontier Labour demand increases do appear to have larger employment effects where non-employment has historically been high, which is a real argument for spending in depressed places. Established It is an argument for a wage subsidy, and the instrument it is used to justify is usually an innovation district.

Frontier The grant-to-small-firm evidence raises an incidence question the programmes do not ask. The benefits concentrate in incumbent employees present before the award and rise with tenure, consistent with a backloaded wage contract being trued up. Speculative That is a defensible outcome — the workers who bore the firm's constraint are paid for it — and it is not what the programme's stated objective describes. A programme justified by innovation and delivering rent-sharing to existing staff is not a scandal; it is a mismatch between an instrument and its rationale, and it is invisible unless somebody measures wages.

Established The circularity has a distributional edge that is rarely stated. If public entrepreneurial finance works where venture capital already is and does not work where it is not, then a programme sold as a means of spreading opportunity systematically concentrates it. Frontier Every place that lacks the precondition is precisely the place the programme was announced for. Speculative Whether a different instrument would work in those places is unknown, because the instrument that has been tried is the one whose failure mode is this.

Frontier And the return-capture argument deserves its strongest hearing and its strongest counter. If the state funds the foundational research, there is a real equity claim to a share of the returns. Frontier The available evidence on the mechanism most often proposed — public equity stakes and state-owned vehicles — is that high government funding shares are associated with underperformance and that ownership does worse than support. Handwave The ethical claim and the instrumental evidence point in opposite directions, and the honest position is that the claim is strong and the proposed instrument is the one the evidence likes least.

Established The terminal position is that the framing is true of instruments nobody argues about and unproven for every instrument that gets announced. Innovation can be cultivated by policy: fix the price of research and firms do more of it, with spillovers onto related firms; fund the science and private patents follow at a measurable rate. Established Innovation ecosystems, in the sense of a place where innovation happens because a government decided it should, have never been shown to be creatable. Both halves are the same subject and they are almost never stated together.

Frontier The change of question that follows is from geography to conditions. The best recent spillover evidence says corporate-network connectivity dominates physical distance; the best entrepreneurial-finance evidence says public capital multiplies an existing private market; the best agglomeration estimates say the mechanism is real and small. Speculative Together those suggest the operative question is not where to put a cluster but what makes a place capable of being connected — which is a question about firms, capital markets and disclosure rather than about land use, and which nobody has instrumented.

Frontier The long-run risk is that the evaluable instruments are quietly starved while the unevaluable ones absorb the budget. Tax credits and research funding are unglamorous, have no ribbon to cut, and produce their effects diffusely; cluster programmes and public venture funds produce announcements. Established The stratification in this brief is a ranking of evidence quality, and it is close to the inverse of the ranking of political attractiveness. Speculative A polity that allocates on attractiveness will, over decades, shift spending from the instruments that work to the instruments that photograph well, and will never learn that it did, because the second kind is not measured.

Frontier And the deepest uncertainty is whether the research base that all of this runs through is being maintained. Both top-ranked instruments operate by funding public research and letting the results spill over. The top of that base is moving into settings that publish less and patent more, at a pay gap that has grown fivefold in two decades. Handwave Nothing measures what that does to the spillover, and a policy architecture whose two best instruments share a single upstream input should want to know. Naming the untested dependency is the most useful thing available here; asserting its size would not be.

These horizons track evaluation opportunities, appropriations and one talent trend rather than technology:

• 10 yr: Frontier The CHIPS and Science Act appropriations are the largest natural experiment in public research funding in a generation, and the projection attached to them — 0.2 to 0.4% of productivity after seven years — has a date on it and can be scored. Established Expect the R&D tax-credit discontinuity to be replicated in other tax regimes within this window, because the data are administrative and the design is portable. Speculative Expect no pre-registered cluster evaluation, because none has appeared in forty years and nothing has changed the incentive. Frontier Expect the AI talent gap to widen rather than close, since the mechanism driving it is a difference in the value of the same person's output to two different employers.
• 25 yr: Speculative If the connectivity result holds, regional innovation policy re-specifies around corporate networks and supply relationships rather than co-location, and the innovation-district form fades the way enterprise zones did. Frontier The alternative is that it does not, because the district is a real estate instrument with a constituency and the evidence is about something else. Handwave Choosing between those is forecasting a political economy, not extrapolating a measurement.
• 50 yr: Speculative At this horizon the interesting question is whether any state ever demonstrates the directional claim — picking technologies and being right, with a counterfactual. Handwave The record contains many assertions and no test, and a claim that has gone fifty years without one is more likely to be untestable than to be awaiting its instrument. Speculative The measurable alternative is that the funding claim keeps being confirmed and the directional claim keeps being asserted alongside it, which is the current arrangement extended.
• 100 / 250+ yr: Handwave Beyond useful forecasting. The base rates available are a handful of research systems whose institutional forms outlasted the states that built them, and their success is attributed after the fact by people with a thesis. Handwave Nothing in the fetched corpus supports a century-scale claim about cultivating innovation, and the honest position is that this brief has none to offer.

• Depends on Nothing on this map. No result another brief produces is a precondition for this one: the identification problems here are structural properties of the instruments rather than missing science, and the two best-evidenced instruments are already well estimated. No typed depends-on edge is claimed.
• Requires (not on this map) One market condition and two institutional ones, none of them a research result. First, a private venture market deep enough for public capital to multiply: three independent studies converge on public entrepreneurial finance working as a multiplier on an existing market and badly or not at all where none exists — the founding SBIR evaluation finding growth effects confined to zip codes that already had substantial venture capital activity, an instrumented international panel finding co-funded firms outperforming at a moderate government share and underperforming at a high one with support beating outright ownership, and a mapping of 755 public programmes worldwide finding private co-investment rising where private venture markets are already well-developed. That makes a functioning private market a precondition for the instrument rather than an objective of it, which is close to a circularity for the whole cultivation premise. Second, a pre-registered counterfactual region for a cluster programme: every credible design in this brief exploits an administrative rule that assigns treatment nearly at random, and cluster programmes have none because they are directed at the places governments choose to help — a lottery among eligible regions, a scored eligibility threshold or a staggered rollout would convert forty years of announcements into evidence, and nothing about it is technically hard. Third, a public research payscale that retains the top of the research base: both best-identified instruments here work by funding public research and letting it spill over, and the top 1% of publishing industry scientists in AI now earn about $1.5 million more annually than comparable academics, a fivefold increase since 2001, with movers to industry publishing less and patenting more. The elasticity of the spillover to that composition is unmeasured; the change in the disclosure regime is not.
• Enables In principle every technology on this map that expects public research funding or a supportive innovation system behind it depends on those instruments working. No typed enabling edge is claimed, and the reason is the brief's central finding: the instruments with clean evidence are narrow and general-purpose — a tax price and a research budget — while the instrument that would enable a named technology in a named place is the one with no credible design.
• Adjacent The economics of innovation and the market-failure framework; urban and regional economics, which supplies the agglomeration estimates; entrepreneurial finance, which supplies the venture evidence; public finance, which supplies the tax-credit designs; the new industrial-policy empirics; and within this map Scientific Governance Models, Future Labour Markets and AI Driven Productivity.

Speculative “Clusters can be built — look at the ones that exist.” No fetched source provides a credible causal design for deliberate cluster creation, and the field's own surveys say so: “our understanding of what works and how it works is quite limited”, with a recommendation of extensive experimentation and careful evaluation. Established The clusters that exist are the observation to be explained, not evidence that policy explains them. Frontier This pass could locate no systematic peer-reviewed evaluation of the deliberate replication attempts at all — not of Skolkovo, not of the national “Silicon” initiatives, not of Malaysia's Multimedia Super Corridor, not of the European cluster-excellence programmes. That is not evidence they failed; it is evidence that the most confident advice in the subject rests on a literature nobody in this pass could find.

Established “Regional investment subsidies raise productivity.” The best-identified evaluation of one finds the opposite: employment, investment and net entry up, total factor productivity flat, effects confined to firms under 150 employees, nothing significant for large firms, about $6,300 per job. Handwave The programme is routinely cited for the employment result while the productivity claim is implied, and the substitution is the step the argument does by assertion. As a job-creation instrument it is cheap and it works; as an innovation instrument the fetched evidence is a null.

Established “Agglomeration effects justify the policies built on them.” A 1% increase in labour market size raises one-year re-employment probability by 0.015 percentage points for high school graduates and 0.023 for college graduates. Established The effects are real, clean and measured on administrative data. They are also small enough that doubling a local labour market — which no policy has achieved — would move the outcome by low single-digit percentage points. Frontier Citing the existence of agglomeration externalities without their magnitude is the most common move in this literature.

Established “The ARPA model has demonstrated efficacy.” Its defining paper identifies features — organisational flexibility, delegated programme-director authority — and a domain, and does not present evidence that the features cause the outcomes. Frontier Many agencies have been founded on it. That is a description of institutional diffusion, not of validation, and the experiment that would validate it is separable, assignable and unrun.

Frontier “Government venture capital works.” It works non-monotonically. Co-funded firms outperform private benchmarks at a moderate government share and underperform at a high one; the underperformance concentrates when private capital is abundant; and support without outright ownership beats ownership. Established Agencies routinely cite the positive half and set all three parameters the other way. The founding evaluation's own qualification — effects confined to venture-rich zip codes — is the part that most often disappears in citation.

Handwave “The entrepreneurial state proves governments can pick winners.” The descriptive core is well corroborated and this brief carries it: public funding produces private patents at a measurable rate, social returns to public nondefense research are large, and public research is upstream of private innovation. Handwave The directional claim — that the state can choose which technologies to back and be right — is not tested by any fetched source, and the NIH evidence is about funding volume in peer-reviewed areas rather than about political direction-setting. Frontier The return-capture claim collides with the government-venture evidence. Handwave And the counterfactual, that the innovation would not have happened otherwise, is established nowhere in this corpus. Established The critics' side — selection bias in the case studies, neglect of failed state bets, conflation of funding with direction — could not be sourced in this pass and is named without citation rather than dropped.

Frontier “Industrial policy has been discredited by economists.” It has not, and the reason it has not is worth stating precisely. A survey by three prominent economists finds recent empirical work a significant improvement on an earlier correlational generation and takes a more positive view as a result — attributing the shift to technique rather than ideology. Established That rehabilitation applies to the instruments the improved techniques can reach, which are the top of the stratification, and not to the cluster and ecosystem-building instruments at the bottom. Citing it for the latter is citing a paper for the opposite of its methodological point.

Frontier “Proximity is the channel.” The best recent spillover evidence says corporate-network connectivity dominates physical distance in cross-cluster innovation spillovers. Frontier If that holds, geographic cluster policy is targeting the wrong unit, and the result — which is usually cited as support for cluster policy — is evidence against the specifically geographic version of it. Speculative And on the direction of causation, the one credible study finds venture capital causing agglomeration rather than following it, established in the destructive direction: restrict the financing and the startups leave. Moving a cluster by removing its capital is demonstrated; moving one by adding capital is not.