The Valley of Death Is a Legibility Problem

Every large technology organization has a research lab. Most research labs produce papers, benchmarks, and prototypes that never become products. The standard explanation is a funding gap — not enough capital at the critical transition point between research demonstration and commercial viability. Branscomb and Auerswald gave this diagnosis its canonical form in 2003, and the policy community has been building programs around it ever since.

The funding-gap diagnosis is empirically weak. Ferguson’s 2014 analysis found the perception “inconsistent with empirical evidence.” Beard and colleagues argued that government funding of basic research paradoxically creates the gap by inflating early-stage output beyond what private markets will absorb. EARTO found that even where governments increased financing for the valley-of-death phase, “the challenge of crossing it and securing first major commercialisation investment persists.” The money arrives. The transitions don’t.

If not funding, then what? The cross-domain evidence — from pharmaceuticals, defense, energy, agriculture, and AI — points to a structural cause that the standard frameworks not only fail to address but actively worsen. The valley of death is a legibility problem. The tools organizations use to manage research transitions — Technology Readiness Levels, stage-gate processes, innovation review boards — decompose a coupled optimization problem into sequential, uncoupled evaluations. Each pass treats the others’ outputs as fixed. The coupling between technical feasibility, market viability, and production readiness — the interaction effects that determine whether research actually ships — becomes invisible. The frameworks that make the portfolio manageable are the same frameworks that make the most important information disappear.

But this diagnosis comes with a critical qualifier that the integration literature typically elides: legibility is stage-dependent. Premature legibility — forcing research to articulate commercial value before the underlying capability is mature — is at least as destructive as failed transition. The challenge isn’t making all research legible to product. It’s knowing when the transition from protected exploration to structured translation should occur, and having mechanisms that work at each stage.

Why the standard fixes don’t work

Technology Readiness Levels are the most widely adopted framework for managing research transitions. Developed by Stan Sadin at NASA in 1974, TRL assigns each technology a maturity score from 1 (basic principles observed) to 9 (proven in operational environment). The scale was designed for 1970s space programs: massive, sequential, fully specified systems where the path from lab to deployment was linear and predictable.

When applied to iterative, exploratory R&D — which is most R&D — TRL becomes a bureaucratic proxy for progress rather than a meaningful assessment of it. A 2020 empirical study by Olechowski and colleagues surveyed practitioners across seven organizations and identified fifteen critical shortcomings. The top-ranked problems were structural: TRL cannot distinguish critical components from non-critical ones, cannot assess integration maturity separate from component readiness, and cannot account for parallel development paths. The authors found that the scale’s sophistication “gradually diminished as its usage spread outside its original context.”

Stage-gate processes have the same structural flaw. Cooper designed them to reduce risk by adding decision points with go/no-go criteria. For incremental product improvements, this works. For research transition, a stage-gate process evaluates technology at Gate 1 against criteria that can only be meaningfully assessed at Gate 3, then penalizes early-stage work for failing to answer late-stage questions. Cooper himself acknowledged that the approach makes overlapping stages “all but impossible.” A Strategos analysis concluded that stage-gate’s controls “tend to weed out the true outliers — those game-changing innovations.”

James C. Scott’s concept of legibility — introduced in Seeing Like a State — provides the precise diagnosis, and it’s one that no published academic paper has yet made explicitly for technology transfer. Scott showed that when complex, locally-adapted systems are forced into standardized, centrally-legible forms, the resulting simplification enables management but destroys adaptive capacity. The state’s cadastral survey makes property legible for taxation but erases the informal use-rights that actually govern the land. Scientific forestry makes timber yield legible but destroys the ecosystem diversity that sustains the forest.

TRL is an almost perfect Scottian legibility project. It takes the messy, nonlinear, context-dependent process of technology development and renders it as a neat 1–9 linear scale readable by managers and funders, systematically destroying the practitioner mētis — the tacit, contextual knowledge — that actually drives successful transitions. The deeper problem is that both TRL and stage-gate are sequential evaluation mechanisms. They assess research readiness, then market readiness, then production readiness — each pass treating the others’ outputs as fixed inputs. This would be fine if the three questions were independent. They aren’t.

Sequential answers to coupled questions

The claim that sequential evaluation loses critical information is not speculative. The pattern has been identified and measured across multiple domains.

In compiler optimization, production systems apply transformation passes sequentially: optimize operator fusion, then memory checkpointing, then device parallelism, then numerical precision. Each pass treats the others’ outputs as fixed. A 2023 MLSys paper demonstrated that making checkpointing decisions aware of downstream fusion opportunities improves throughput by 12% and enables up to 1.75× larger batch sizes — gains that come not from better individual passes but from reasoning about the interaction effects between them.

In supply chain management, a meta-analysis comparing sequential planning against integrated planning found an average cost reduction of 11% from integration — not from better individual decisions, but from capturing the coupling between production, distribution, and inventory.

In military operations, the pattern was so costly that Congress legislated against it. Before the 1986 Goldwater-Nichols Act, U.S. military services planned operations sequentially — Army, Navy, Air Force each optimizing independently, then “deconflicting” at the end. The result was catastrophic coordination failures that stemmed directly from sequential planning’s inability to capture interdependencies. Congress mandated joint planning and created career incentives for joint service — a structural fix to a structural problem. The parallel to research-product transitions is direct: Goldwater-Nichols required joint duty experience as a criterion for career advancement, which is precisely the mechanism the “research engineer” role at organizations like Anthropic now provides.

The research-to-product transition has exactly this structure. “Is the research technically sound?” is coupled to “Is there a market?” because what counts as technically sufficient depends on what the market requires. “Can we build it?” is coupled to “What’s the competitive positioning?” because build cost depends on performance targets, which depend on where you compete. These aren’t three independent questions answered in sequence. They’re a system of mutual constraints that must be solved jointly.

Chatterjee and colleagues quantified what this costs: they found that “commercial viability explains 39% of innovation success variance” and “organizational performance improves 71% when companies navigate the valley of death through appropriate operational structures.” Getting the structure right matters more than any single factor — more than funding, more than talent, more than executive attention.

Premature legibility and the maturity problem

But here the argument requires an uncomfortable qualifier, because the integration literature — including the first draft of this essay — typically ignores the mirror failure mode.

Premature commercialization is at least as destructive as failed transition, and the evidence is devastating. IBM Watson Health destroyed over $5 billion in value by rushing NLP capabilities to clinical deployment before the technology could support real-world use. Internal documents revealed a doctor telling IBM that Watson for Oncology was “a piece of shit.” The expert systems bubble of the 1980s — over 300 AI companies shut down by 1993 — was driven by the same dynamic. Cleantech 1.0 lost over $25 billion when software VCs applied 3–5 year return expectations to companies requiring 10+ year R&D cycles. In each case, the failure wasn’t that research lacked legibility to product. It’s that the legibility was premature — it made immature capability look product-ready.

Deep learning itself — the most consequential AI breakthrough of the century — required roughly fifteen years of protection from commercial pressure. From 1997 to 2012, the field had no obvious product application. Hinton moved to Canada explicitly to escape the constraints of military-funded, application-oriented AI research. A commercially integrated evaluation of deep learning in 2006 would have killed it. The work needed time to develop outside the vocabulary of market viability.

This creates a genuine tension with the legibility thesis. If forcing early joint evaluation of coupled questions destroys the most important research, then the mechanism I’m describing can’t be universally applied. The question isn’t whether to make research legible to product — it’s when.

Trading zones and boundary objects

If the problem is sequential evaluation of coupled questions, and the constraint is that forcing legibility too early destroys research quality, then the solution can’t be “make all research legible to product.” It has to be something more precise.

Peter Galison’s concept of trading zones offers the right frame. Galison studied how physicists, engineers, and instrument-makers — groups with genuinely incommensurable paradigms — managed to collaborate on particle physics experiments. His finding: they didn’t resolve their epistemic differences. They didn’t learn each other’s languages. Instead, they developed local pidgins — simplified shared vocabularies and material practices that enabled coordination without requiring deep agreement. The trading zone is the space where exchange happens, mediated by artifacts that both sides can use even though they interpret them differently.

Susan Leigh Star and James Griesemer formalized the coordinating artifacts as boundary objects: things “plastic enough to adapt to local needs and the constraints of the several parties employing them, yet robust enough to maintain a common identity across sites.” A prototype is a boundary object — researchers see a proof of concept, product leaders see a feature demo, engineers see a system architecture. Same artifact, different readings, productive coordination.

Amazon’s PR/FAQ is a boundary object, not just a legibility device. It’s usually described as a “customer-focused” planning tool: write a fictional press release announcing the finished product, then answer frequently asked questions. This description is accurate but insufficient.

The PR/FAQ works because it forces the writer to articulate the dependency structure between research capability, customer value, and business viability simultaneously, within a single narrative. You cannot write a credible press release without answering, in the same document: What does this do for the customer? What technology makes it possible? Why will this work at scale? Why should the company invest? Each answer constrains the others. The FAQ section makes the tensions explicit by forcing the writer to address objections from engineering, finance, and business development — all within the same document review.

But the PR/FAQ functions as a coupling device only when constructed as a cross-functional team activity — researcher, product manager, engineer, and business development negotiating the narrative jointly. The document becomes the site of negotiation, not a translation from one vocabulary into another. The coupling happens in the process of writing, not in the finished artifact.

Two degraded forms are common. When a researcher writes the PR/FAQ alone, it becomes a one-directional translation device — legibility flowing from research to product, with no constraint traveling back. When a product manager writes it alone, it becomes justification backfill: a product narrative that borrows research credibility without being constrained by technical reality. This is premature legibility wearing a different mask. The document looks like joint optimization but is actually single-perspective justification dressed in the coupling format.

The artifact isn’t the mechanism. The mechanism is the cross-functional negotiation that the artifact structures. The PR/FAQ is a protocol for forcing joint articulation, and it degrades to sequential evaluation whenever the protocol collapses to a single perspective.

This is joint optimization forced by narrative structure — when the protocol is intact. The PR/FAQ is a coupling device — a mechanism that prevents the decomposition of a coupled problem into uncoupled subproblems. But it’s also a boundary object: researchers read it as a translation of their capability, product leaders read it as a feature proposal, executives read it as an investment case. Same document, different interpretations, productive coordination.

But a boundary object only works in a trading zone — a space where the exchange actually happens with people who have decision authority. This requires structured visibility: the document has to circulate through a defined audience on a cadence they already attend. The most brilliant PR/FAQ sitting in a researcher’s Google Doc has the same commercial impact as a brilliant paper sitting on arXiv — zero, until it reaches someone who can act on it.

And even a well-distributed boundary object requires something prior to all of this: demand. DARPA’s transition data — from a GAO study that should be better known — found that demand for the technology, not funding or organizational structure, was the primary predictor of successful transition. The Heilmeier Catechism forces every DARPA program manager to answer “Who will take the technologies and turn them into a capability or product?” before a program begins. Not “is this technically interesting?” Not “can we build it?” But “does someone want it?” Legibility without demand is an essay with no reader.

The mechanism, fully specified, has four preconditions: the research must be mature enough that commercial evaluation won’t destroy it; demand must exist (or be articulable) on the product side; a boundary object must force joint articulation of the coupled questions; and the boundary object must circulate through a trading zone with decision authority.

Two cases, two domains, the same 6×

I ran a version of this mechanism at AWS’s AI research organization — roughly 150 researchers across graph neural networks, deep learning frameworks, causality, reinforcement learning, NLP, computer vision, and large-scale distributed training. The standard failure mode was familiar: world-class research that never became product because nobody with a product budget understood what it could do.

The mechanism had three components. First, systematic evaluation of research maturity — examining code repositories, competitive benchmarks, and existing integrations to identify which projects had moved beyond proof-of-concept into something that could plausibly support a product. This was the maturity gate: not “is this good research?” but “is this ready for commercial evaluation without being destroyed by it?”

Second, narrative translation via the PR/FAQ — working with research teams to articulate, in product language, what their capability made possible. Not “we achieved state-of-the-art on benchmark X” but “this enables customers to do Y, which they currently can’t do or pay Z to do manually.”

Third, structured visibility through the business review cadence — putting the translated narrative in front of the people who allocate product budgets, in a format they could evaluate, on a cadence they already attended.

Neptune ML illustrates the full cycle. Amazon’s AI Lab included a team working on graph neural networks applied to knowledge graphs — strong academic results, no obvious productization path. “Graph neural networks for knowledge graph enrichment” is a research description, not a product description. The translation step converted it: customers using Amazon Neptune could now run ML-powered predictions directly on their graph data — link prediction, node classification, knowledge graph completion — without moving data out of Neptune, without managing ML infrastructure. The business review mechanism put this narrative in front of the Neptune service team and their leadership chain. Neptune ML shipped as a production feature.

The aggregate result was a roughly 6× increase in the rate of research projects entering formal product planning pipelines. This was not because the research improved or the budget increased. What changed was the legibility of research value to the people who made product decisions.

But the same organization provides a case that exposes the mechanism’s limits — and the limits are more instructive than the success.

Amazon’s first serious attempt at building large language models received paltry investment: roughly thirty headcount and approximately $15 million in annual compute, at a time when scaling laws and the transformer architecture’s potential were already legible to anyone reading the field closely. The team was training 500-million-parameter models and distilling to 70 million for production deployment while the frontier was at 175 billion parameters. Not because the researchers didn’t understand scaling — their roadmap explicitly targeted 100-billion-parameter models — but because the funding mechanism couldn’t support that ambition. AWS AI research operated on a patronage model: service general managers funded the research they wanted to consume. Resources flowed proportional to near-term service value, not strategic potential.

The team’s strategy was rational within these constraints: mature the research by building LLMs for specific internal use cases — improving search relevance, duplicate product detection, ad click prediction, collaborating with the voice assistant team on a multilingual teacher model. This is exactly the boundary-object approach the essay advocates. Make research legible by building things service teams can evaluate. And it worked for what it was designed to do: real measurable improvements on specific tasks, significant training cost reductions through cross-team collaboration, demonstrable product value.

But the patronage model shaped the science. The framing positioned the work entirely as proprietary entity representations — not general-purpose language models. The team argued against GPT-3-style scaling, noting “diminishing returns” from parameter increases and optimizing instead for economically viable deployment via distillation. This wasn’t wrong as science — there are legitimate efficiency arguments, and distillation remains valuable — but it was wrong as strategy, and wrong in the specific way that patronage funding encouraged. The GMs who funded the work wanted deployable models with tight latency constraints, not frontier capability research. The funding structure selected for exactly the research questions that fit within existing product frames, and filtered out the questions that didn’t.

Meanwhile, the same organization’s annual planning process — itself a well-constructed PR/FAQ ecosystem, cross-functional and data-driven — correctly allocated the lion’s share of resources to the managed ML infrastructure platform, which had clear revenue (over $500 million ARR), clear customers, and clear competitive positioning. The LLM team’s case was illegible under that framework: no direct revenue, uncertain customer, unclear competitive moat, a value proposition that only made sense if you believed the entire AI services layer would be rebuilt on top of foundation models. The most revealing signal: senior leadership saw the opportunity clearly enough to invest in Anthropic — then still in stealth — as a way to ride the wave without taking on internal risk. The same organization that couldn’t allocate a hundred engineers to its own LLM effort wrote a check to a company whose entire thesis was that foundation models would reshape AI. This was venture investment as an end-run around the organization’s own legibility constraints. Leadership didn’t miss the opportunity. The funding mechanism couldn’t express it.

This is the PR/FAQ failing not through incompetence but through correct operation of its own logic. The planning process did exactly what it was designed to do: prioritize investments with clear customer demand, measurable outcomes, and defensible competitive positioning. Foundation models failed every one of those tests in 2021. The mechanism worked perfectly and produced the wrong answer — because the evaluative frame couldn’t accommodate a paradigm shift that made the existing frame obsolete. The patronage model is Fraunhofer’s demand-pull without Fraunhofer’s base funding: it captures the integration benefits but has no structural protection for research that exceeds what current customers can evaluate.

Independently and in a completely different domain, AstraZeneca achieved the same magnitude of improvement through a structurally identical intervention. Their 5R Framework changed drug development success rates from 4% to 23% — also roughly 6× — by changing what questions were asked before committing resources. The 5R forces joint evaluation of scientific validity, clinical feasibility, and commercial viability within a single assessment. Same coupling device, different vocabulary, same result.

The convergence on 6× across domains with radically different failure rates (pharmaceutical development at 6.7% baseline, research-lab transitions at a lower but less precisely measured baseline) suggests this may be the approximate ceiling for organizational intervention against coupled-question failure. What organizational design can’t overcome — and AstraZeneca’s remaining 77% failure rate makes this clear — is genuine scientific uncertainty. Most hypotheses are wrong. The mechanism doesn’t change that. It prevents the organizational structure from making it worse.

What the current AI labs are testing

The contemporary AI industry is running a live experiment in research-product organizational design, and the early results complicate every simple narrative.

OpenAI’s fully product-dominant model has produced extraordinary commercial results — $12 billion in annualized revenue — but significant research attrition. Jan Leike, former head of alignment, left for Anthropic, stating that “safety culture and processes have taken a backseat to shiny products.” The GPT-4o safety team received one week for testing before launch. This is premature legibility at institutional scale: research forced into product cadence before the underlying questions are resolved.

Meta FAIR’s separate-lab model produced extraordinary research — PyTorch, Llama, foundational work in self-supervised learning — but Yann LeCun himself acknowledged that “FAIR was extremely successful in the research part. Where Meta was less successful is in picking up on that research and pushing it into practical technology and products.” PyTorch succeeded because it was a tool — legible as useful infrastructure with adoption metrics product leadership could evaluate. LeCun’s research on world models remained illegible to anyone optimizing for Meta’s engagement metrics. Same lab, same researcher, different legibility outcomes depending on whether the research output was itself a boundary object.

Google DeepMind’s forced merger of previously competing labs was structurally a legibility intervention — eliminating the autonomy buffer that had allowed DeepMind to operate in its own evaluative language. Whether this was wise depends on what you’re optimizing for. DeepMind’s separation produced AlphaFold, arguably the most important scientific result of the decade. A commercially integrated DeepMind might never have pursued it.

Anthropic’s approach is the most structurally interesting. Constitutional AI represents genuine research-product alignment: the research technique directly improves the product. But Anthropic is also migrating to a dual-track model — a scaled product organization alongside an experimental Labs incubator. The “research engineer” role, where the company deliberately blurs the researcher/engineer divide, may be the most consequential organizational innovation: it creates boundary-spanning positions that function as permanent residents of the trading zone between research and product. This is the individual-level version of what Goldwater-Nichols did at the institutional level — making the boundary crossing a career requirement rather than an exception. But the research engineer role works when the research is at the right maturity for integration — it creates permanent residents of the trading zone. It cannot compensate for a funding structure that prevents research from reaching that maturity in the first place. The structural precondition is a funding mechanism that protects exploratory work, not just a role that bridges the gap once the work is ready to cross.

The UC Berkeley finding on Pasteur’s Quadrant researchers — people who both publish papers and file patents — provides quantitative support: these dual-purpose researchers produce work that is more novel and more influential than those who stick to one activity, and their numbers have increased 350% since 1980. But they concentrate in wealthy geographic areas and remain rare (14% of inventors, 4% of scientists), suggesting the conditions that produce them are institutionally fragile.

The hardest open question

The maturity-dependent thesis — protection early, integration late — creates a problem it doesn’t solve: when should an organization switch from protecting research to demanding legibility?

No empirical framework answers this. The evidence suggests the switching point depends on at least three factors: whether the research has stabilized enough that commercial evaluation won’t destroy its optionality (deep learning was unstable until roughly 2012; graph neural networks for knowledge graphs were stable by the time we evaluated them); whether demand exists or is articulable on the product side (DARPA’s primary predictor); and whether the research output can be expressed as a boundary object that both communities can use without resolving their epistemic differences.

The Fraunhofer model offers one institutional answer: the 30/70 funding split (30% government base, 70% contract research) forces dual orientation structurally, without requiring anyone to determine the switching point for individual projects. Fraunhofer presence boosts patent output of local firms by at least 13%. The institutional DNA is captured in the phrase “If you cannot sell it, it’s not Fraunhofer.” This works because the funding structure itself creates a permanent trading zone — researchers must always be partially legible to industry, but base funding protects exploratory work from premature commercialization.

The 30% base matters precisely because it is the structural element that pure patronage lacks. In a patronage model — like the one that shaped the LLM investment at AWS — there is no institutional floor that says “you can pursue directions that no service GM would fund yet.” Every dollar requires a sponsor with a near-term use case. The maturity switch this essay describes (protect early, integrate late) requires a funding mechanism that allows the protection phase. Patronage doesn’t have one. The Fraunhofer base is what makes the switching point a gradient rather than a cliff: researchers who are 30% exploring and 70% translating live in both worlds simultaneously, rather than crossing from one to the other at a discrete moment that someone must correctly identify in advance. The funding structure creates permanent boundary-spanners by default, which is structurally superior to requiring anyone to get the timing right on individual projects.

The Green Revolution offers another answer from a radically different domain. Borlaug’s dwarf wheat varieties moved from research plots to deployment in roughly twenty years, saving an estimated one billion lives. Six institutional design features explain the success: seeds from non-profit institutions (building trust), unpatented and distributed at cost (eliminating IP barriers), shuttle breeding that cut development time in half, technology bundled with complementary infrastructure, government demand pull, and Cold War geopolitical urgency. The contrast with GM crops — where 84% of all GMO acres remain in four Western Hemisphere countries after twenty-plus years — confirms that institutional design and social trust matter as much as technical capability. Proprietary seeds from corporations face barriers that public-good seeds from trusted non-profits never did. Same science, different institutional legibility, radically different deployment.

What I don’t yet know — and what the literature doesn’t yet answer — is whether the switching point can be identified in advance rather than recognized in retrospect. The pharma industry’s progressive decline in approval rates (6.7% and falling) suggests that even sophisticated maturity assessment can’t keep pace with increasing scientific complexity. The 6× improvement is real and replicable, but the ceiling it hits may be getting lower as the problems get harder.

A note on what this is and isn’t

This essay argues that the valley of death between research and product is primarily a structural problem — coupled questions answered sequentially, through frameworks that impose legibility at the cost of destroying the coupling information that matters most. The evidence supports this diagnosis across pharmaceuticals, defense, energy, agriculture, and AI, with organizational interventions achieving consistent 2–6× improvements in transition rates.

It does not argue that organizational design can overcome genuine scientific uncertainty. Most hypotheses are wrong, and no coupling device changes that. It does not argue that integration is always superior to separation — deep learning needed fifteen years of protection, PARC’s autonomy produced revolutionary research, and premature commercialization has destroyed more value than failed transition. And it does not argue that legibility is the only factor — demand must exist, the research must be mature, and the trading zone must include people with decision authority.

What it does argue is that the standard mechanisms — TRL, stage-gate, sequential evaluation — are not merely insufficient but actively harmful, and that the alternative is not “more integration” but “better boundary objects circulating through well-designed trading zones at the right stage of research maturity.”

The through-line connects to work I’m pursuing in adjacent domains — claim-dependency tracking in AI safety evaluation, joint optimization in ML compilation, real-time state awareness in contemplative practice. Each involves the same structural pattern: dependency structures that exist, that must exist because the systems have logical structure, but that live in implicit, manually-maintained forms and fail when conditions change faster than manual maintenance can keep up. Whether this pattern genuinely generalizes or whether I’m overfitting to a structural metaphor is an open question I’m not yet equipped to answer. But the valley of death is the most economically visible instance, and the evidence there is strong enough to act on.

Sources and Evidence Notes

The TRL critique draws on Olechowski et al. (2020), “Technology readiness levels: Shortcomings and improvement opportunities,” Systems Engineering. The valley of death literature draws on Markham (2002), Research-Technology Management; Branscomb & Auerswald (2003), Journal of Technology Transfer; and Ferguson (2014). The funding-gap counterevidence includes Beard et al. on government funding paradoxes and EARTO’s analysis of European transition programs. James C. Scott’s Seeing Like a State (1998) provides the legibility framework. The claim that Scott’s framework has not been applied to technology transfer in published academic work is based on the research report’s independent confirmation — I welcome corrections.

Peter Galison’s concept of trading zones appears in Image and Logic (1997). Star and Griesemer’s boundary objects framework appears in “Institutional Ecology, ‘Translations’ and Boundary Objects” (1989), Social Studies of Science. Takeuchi and Nonaka’s critique of sequential development appears in “The New New Product Development Game” (1986), Harvard Business Review. Cooper’s stage-gate acknowledgments are drawn from his own published revisions.

The DARPA transition evidence draws on GAO studies of program outcomes. The Goldwater-Nichols analysis draws on the legislative history and operational assessments of Desert Storm. The AstraZeneca 5R framework evidence draws on published success-rate data from Morgan et al. The pharma failure data draws on Norstella/Citeline’s 2024 analysis and Eroom’s Law (Scannell et al.). The Fraunhofer evidence draws on econometric studies of patent output effects. The Green Revolution analysis draws on CGIAR impact assessments. The Bell Labs analysis draws on Jon Gertner’s The Idea Factory and Watzinger & Schnitzer’s finding that the Bell breakup increased overall innovation.

The Pasteur’s Quadrant finding draws on a November 2024 UC Berkeley study published in Science. AI lab organizational analysis draws on published reporting and public statements from OpenAI, Anthropic, Google DeepMind, and Meta.

The AWS AI Lab case study is drawn from my direct experience as the technology transition lead. Project names (Neptune ML, DGL) are publicly documented AWS features. The “6×” figure refers to the rate of research projects entering formal product planning pipelines during the period I managed the transition function, compared to the prior baseline.