The Judgement Core

Summary: science-based ventures are moving their protected centre from the factory to the judgement core

1. Introduction: four observations that the classical theory of the firm explains poorly

Consider a decision that a growing number of food biotech founders now face. A team with a promising fungal protein has raised its seed round. A landlord offers a lease on a 5,000-litre fermentation suite that the team would run on its own schedule and could show investors as a tangible asset. For the same money, a contract manufacturer offers three rented production runs over twelve months, with a contract clause giving the start-up full rights to every data point those runs produce. Twenty years of management practice says take the lease. This essay argues that, for most such teams today, the second option is the better choice, and that the reasons tell us something general about how science-based ventures are changing.

Twenty years of management practice says take the lease. This essay argues that, for most such teams today, the second option is the better choice

The classical answer has a long pedigree. James D. Thompson's Organizations in Action (1967) placed a technical core, the production machinery that turns inputs into outputs, at the protected centre of every organisation, with buffering and boundary-spanning units arranged around it. Transaction-cost economics, from Ronald Coase (1937) to Oliver Williamson (1985), added that assets specific to one relationship should be owned, because renting them invites hold-up. A science-based venture with a proprietary process would, on both accounts, be expected to own its production as early as it could.

Four observations from the present decade sit uneasily with that expectation.

1. Owned steel has become a frequent companion of failure. Fermentation ventures that raised hundreds of millions of US dollars, including Meati (about USD 400 million raised, with its own production facility; High, 2025) and Motif FoodWorks (USD 345 million by mid-2021; Watson, 2024), closed or were restructured between 2024 and 2025, while venture funding for fermentation fell from USD 632 million in 2024 to USD 357 million in 2025 (Good Food Institute, 2026).
2. The cost of coordination is falling at an unusual speed. The length of task that frontier AI agents can complete with 50% reliability has doubled roughly every seven months since 2019, and roughly every 4.3 months since 2023 (Kwa et al., 2025; METR, 2026).
3. Capacity can be rented, but it is scarce and rarely rich in data. The global food fermentation capacity available to start-ups is estimated at a lower bound of about 16 million litres, with an investment of more than USD 250 billion needed by 2050 (Good Food Institute & Integration Consulting, 2023; Aminetzah et al., 2025).
4. Start-ups are getting smaller before agents have even matured. The average seed-stage company on Carta had 6.2 equity-holding employees in 2025, down from 10.3 in 2021 (Carta, 2025; Vidra, 2025).

None of these facts refutes Thompson or Williamson. Together, however, they suggest that something in the conditions those theories assumed has changed, and that the protected centre of the science-based venture may be moving.

Research questions

The essay addresses three questions.

• RQ1. What should a science-based venture keep inside its boundary when coordination is cheap and production capacity can be rented?
• RQ2. How should AI agents be arranged in relation to the human decisions that carry the venture's risk?
• RQ3. Under what conditions does renting capacity create value for the venture rather than leak it to the owner of the capacity?

Contribution and structure

The contribution is a theory, offered for testing, that connects the theory of the firm with three active conversations in innovation management and entrepreneurship: how new ventures in deep technology organise under uncertainty, how open and outsourced models of innovation distribute value, and how human judgement and AI are being combined inside organisations.

Section 2 sets out the research design and the logic of inference. Section 3 assembles the evidence and grades its strength. Section 4 places the argument within established theory. Section 5 presents the theory itself, Section 6 its propositions, and Section 7 its practical implications. Section 8 states where the theory stops and how it could be tested.

2. Research design: a theory adaptation built by abduction and exposed to falsification

2.3 Sources and grading of evidence

The essay weights evidence by its source. It uses four tiers and grades every central claim as strong, moderate or speculative.

2.4 Choice of context

Food biotech, and precision and biomass fermentation in particular, serves as a revelatory context: a setting in which all the pressures the theory concerns act at once. Capacity is scarce, regulation is slow, capital is selective, and AI is already entering design and process work. A theory that holds here should be easier to test elsewhere in physical and regulated sciences; one that fails here is unlikely to hold anywhere else.

2.5 Threats to validity and the author's position

Four threats deserve naming. Confirmation bias: an author who proposes a theory will notice supporting evidence first, which is why each proposition carries a stated refutation. Survivorship: published fermentation results and visible failures are not a random sample of ventures. Analogy risk: the semiconductor precedent in Section 4 illustrates a mechanism; it does not prove that fermentation will follow the same path. Positionality: I founded FoodEdge, which keeps a record of 172 food biotech funding rounds and 47 venture post-mortems; I serve MycoVibes Biotech as fractional Chief Strategy and Commercial Officer; and I evaluate proposals for the European Commission. These roles give access to the phenomenon and create interests in how it is described. They are disclosed here, and no claim in this essay rests on confidential information.

The essay does not test its propositions statistically. It states them in a form that can be tested, and Section 8 describes how.

3. The phenomenon in numbers: strong evidence on capability and capacity, weaker evidence on organisation

This section separates evidence from interpretation. The table records what is known, from what kind of source and with what confidence. The interpretation follows it and is marked as such.

3.1 Evidence

Two kinds of evidence are deliberately absent. Market-size forecasts for precision fermentation, which range from about USD 9 billion (360iResearch, 2026) to about USD 78 billion (Maximize Market Research, 2026) for 2032, are left out because their spread exceeds their information. And no figure in the table comes from the FoodEdge record, which remains to be analysed in the way Section 8 describes.

3.2 Interpretation

Five inferences follow from the table. Each is labelled with the confidence it deserves.

1. The cost of coordination and analysis is falling quickly (strong). Capability measurements and productivity studies agree in direction, though the size of the effect inside science-based ventures is not yet measured.
2. The gains from AI shrink as experiments approach industrial scale (moderate). The pattern from bench to tank is consistent across the studies available, but they are few and differ in method. This is the basis of the Bench-to-Tank Discount, treated here as a hypothesis.
3. Physical capacity, regulatory time and capital are the binding constraints (moderate to strong). Each is documented independently; their joint effect on venture speed is an inference.
4. Early ownership of capacity carries high risk in the current capital cycle (moderate). Visible failures and falling funding point the same way, but visible failures are not a random sample.
5. Owners of scarce production capacity can capture a large share of value (moderate, by analogy). The semiconductor case shows the mechanism; whether it applies in fermentation is an open question.

Taken together, these inferences describe a world in which analysis and coordination are becoming cheap while capacity, regulatory approval and capital still move slowly.

The next section asks what established theory says about how firms should be organised in such a world.

4. Theoretical foundations: six bodies of theory each explain one part, and none explains the whole

The argument draws on six bodies of work in the theory of the firm, innovation management and entrepreneurship. Each answers one of the questions a science-based venture faces. None, on its own, answers them together.

4.1 Where the boundary of the firm lies

Coase (1937) located the boundary of the firm where organising one more task internally costs as much as buying it in the market. Williamson (1985) added that specific assets favour ownership. Malone, Yates and Benjamin (1987) predicted that information technology would shift activity from hierarchies towards markets by lowering coordination costs. Recent work applies this logic to AI agents: an NBER working paper by Shahidi, Horton and colleagues (Shahidi et al., 2025) asks whether agents could push transaction costs towards zero, Klein and Wieczorek (2026) argue that agentic AI produces a "headless" firm with an hourglass architecture, and Kim and Koning (2026) find that AI-native start-ups are smaller and flatter than comparable firms.

What this leaves open: almost all of this work concerns information businesses. Klein and Wieczorek state that their model does not hold where processes are tightly coupled to physical state, which is exactly the condition of a fermentation venture.

4.2 What sits at the protected centre

Thompson (1967) placed the technical core at the centre, protected by buffering and boundary-spanning units. Mintzberg (1979) placed a strategic apex at the top of the organisation, defined by hierarchy. Agrawal, Gans and Goldfarb (2018) argue that as AI makes prediction cheap, the value of human judgement, deciding what to do with a prediction, rises. Hydari and Muzaffar (2026) propose that the boundary of an AI-assisted firm follows its accountability assets: whatever makes an output legitimate and assignable to a responsible party.

What this leaves open: none of these accounts says where judgement should sit when the technical core can be rented, or how AI agents should be arranged around it.

4.3 How science-based ventures create capability

Gary Pisano's Science Business (2006) argued that biotechnology struggles with three problems at once: managing deep uncertainty, integrating knowledge across disciplines, and learning from slow, expensive experiments. Cohen and Levinthal (1990) showed that a firm's ability to use outside knowledge, its absorptive capacity, depends on knowledge it already holds inside. Teece, Pisano and Shuen (1997) described dynamic capabilities as the firm's ability to sense opportunities, seize them and reconfigure its resources.

What this leaves open, and what it suggests: if production and much analysis move outside the firm, absorptive capacity and dynamic capabilities must live somewhere inside it. The judgement core is a candidate answer: the place where absorptive capacity is concentrated, which is why its members must be able to judge the science themselves.

4.4 How value is captured when innovation is opened up

Teece (1986) showed that innovators often lose value to the owners of complementary assets such as manufacturing. Chesbrough (2003) described open innovation, in which firms deliberately source and sell knowledge across their boundaries. Sturgeon (2002) documented modular production networks in which brands no longer own factories. Christensen and Raynor (2003) proposed that profit migrates to the part of a value chain that remains hard to copy. The semiconductor industry provides the clearest case: after TSMC opened as the first pure-play foundry in 1987, companies that design chips without owning plants flourished, and the foundry itself came to hold 69.9% of its market by 2025 (Focus Taiwan, 2026).

What this leaves open: these theories were written before rented production generated large volumes of data that software could learn from. They do not say who should own that data, or what happens to value when the renter's learning depends on it.

4.5 How entrepreneurs act under uncertainty

Four influential accounts of entrepreneurial and innovation practice point the same way. The Lean Startup (Ries, 2011) treats a venture as a loop of building, measuring and learning, judged by the speed of validated learning. Effectuation (Sarasvathy, 2001) describes expert entrepreneurs who commit only what they can afford to lose. Real-options reasoning (McGrath, 1999) values small, reversible investments that keep future choices open. Stage-Gate systems (Cooper, 1990) organise innovation as a sequence of decisions at defined gates.

What this leaves open, and what it suggests: all four favour small, reversible commitments and fast learning, yet none addresses physical experimentation that must be rented and data that must be secured. Renting capacity can be read as an affordable-loss and real-options strategy; loop velocity is the physical counterpart of validated learning; and gate latency measures how quickly the decisions at each gate are actually taken.

4.6 How work is divided between people and machines

Parasuraman, Sheridan and Wickens (2000) distinguished the types and levels of automation. Shneiderman (2020) argued that high automation and strong human control can coexist. Hung and colleagues (2024) proposed levels of laboratory autonomy from L0 to L5. Microsoft's Frontier Firm (Spataro, 2025) and McKinsey's agentic organisation (Sukharevsky et al., 2025) describe human-agent teams in practice.

What this leaves open: these frameworks set autonomy task by task. None offers an organisational rule for allocating autonomy according to how close a task sits to the decisions that carry a venture's risk.

4.7 Synthesis

5. The theory: a judgement core at the centre, three orbits of agents and a rented belt

5.1 Constructs

A theory needs constructs precise enough to be measured and distinct from their neighbours. Table 5.1 defines the nine used here.

Two further measures describe the configuration as a whole: the belt ratio, the share of physical capacity spend that is contracted rather than owned, and a firm's Copernican position, estimated jointly from belt ratio, memory share and judgement ratio.

5.2 Architecture

The architecture has a direction. Data flows inwards: a run ends in a rented tank, the coordination orbit receives the results, the state orbit sets them against what is already known, and the experimental orbit turns them into ranked proposals for the core. Decisions flow outwards along the same path, ending in a booked run in the belt. The speed of this cycle is loop velocity.

5.3 Two ideal types

The names borrow from the history of astronomy, but the meaning comes from philosophy. In the preface to the second edition of the Critique of Pure Reason (1787), Immanuel Kant described his own inversion as Copernican: instead of knowledge conforming to objects, objects conform to the way they are known. In the Ptolemaic Firm, judgement conforms to production, because the plant one owns sets the limits of strategy. In the Copernican Firm, production conforms to judgement. Unlike some recent uses of the metaphor in writing on AI, which move people away from the centre, this one moves the factory away and keeps human judgement in the middle.

5.4 Five mechanisms

1. Boundary mechanism. When the cost of coordination falls faster than the specificity of assets rises, the efficient boundary of the venture contracts towards its judgement core.
2. Inverse autonomy. The closer a layer of agents sits to the core, the more consequential its errors, so the fewer final decision rights it should hold. Autonomy can rise everywhere over time; the claim concerns its relative distribution.
3. Bench-to-tank discount (moderator). As experiments approach industrial scale, the cost per data point rises and the gain per iteration falls, weakening the effect of agent capability on progress.
4. Triad constraint. The speed of the venture is limited by the slowest of three clocks: data-rich capacity, regulatory time and capital backed by real orders.
5. Memory clause. When production is rented, value accrues to whoever holds rights to the data production generates.

5.5 Failure modes

6. Six propositions, each with its premises, its mechanism and the result that would refute it

Each proposition below is written as a chain: the premises it rests on (with the evidence grade from Section 3), the mechanism that links premises to prediction, the prediction itself, and the observation that would count against it. The propositions apply within the boundary conditions set out in Section 8.

6.1 How the propositions relate

P1, P5 and P6 describe properties of the judgement core itself. P2 and P4 describe the Copernican configuration around it. P3 describes the process of moving towards that configuration, and the cost of reversing it. A single venture that showed a high judgement ratio, hybrid founders, high memory share and rented capacity, yet progressed no faster than its Ptolemaic peers, would be a serious problem for the theory as a whole.

7. Implications: own the decisions and the data, rent the steel, delegate the coordination

A theory of organisation earns practical value when it changes decisions. This section translates the propositions into decision rules for five groups. The rules are interpretations; where the propositions remain untested, so do the rules.

7.1 A decision rule: own, rent or delegate

7.2 For founders

The most important hire is often a co-founder, and the most important qualification is the ability to judge the core science, whatever that person's commercial role (P5). Before signing any manufacturing contract, read the data clause first and the price second (P4). Sequence regulatory routes so that faster approvals generate revenue and data while slower ones proceed in parallel; in the United States, the FDA aims to answer a GRAS notification within 180 days (21 C.F.R. § 170.265), while the first EU authorisation of a novel mycoprotein took six years (Watson, 2026). And measure from the first month the five numbers below, because a company that cannot see its own loop velocity cannot manage it.

Before signing any manufacturing contract, read the data clause first and the price second.

7.3 For investors

Due diligence can be shortened to four questions.

The theory also implies that capital for science-based ventures will increasingly split in two: venture capital to fund learning, and infrastructure or project finance, backed by binding orders, to fund steel. Funds that mix the two in one instrument may be encouraging the Ptolemaic relapse they later write off.

7.4 For corporate innovation managers

Large food and pharmaceutical companies are, almost by construction, Ptolemaic: their plants sit at the centre of their accounts and their careers. The theory leaves their existing organisation in place and suggests building new ventures as Copernican units from the start, in the spirit of the structural ambidexterity described by O'Reilly and Tushman (2004): separate units with their own judgement core, renting capacity from the parent company's own plants under a contract that gives the unit rights to its data. The parent then becomes a supplier with a strong interest in the unit's learning, which is a healthier relationship than ownership of an experiment it does not understand.

7.5 For contract manufacturers

For manufacturers, the theory points to a choice of business model. One option is to sell litres and keep the data, which captures value in the short term and invites customers to leave as soon as they can. The other is to sell instrumented capacity with data rights attached, and to build trust on the principle that made TSMC the leading foundry: never compete with a customer. The second option is slower to monetise and harder to copy.

7.6 For policymakers

If regulatory time is one of the three clocks, regulatory design becomes innovation policy. Structured scientific dialogue before submission, regulatory sandboxes that include novel foods, and shared, instrumented pilot facilities would each shorten a clock that no amount of private investment in AI can shorten.

7.7 A dashboard of five indicators

8. Limitations, boundary conditions and a research agenda: what would have to be done to turn a plausible theory into a tested one

8.1 Boundary conditions

The theory is proposed for ventures that meet five conditions: their product must be grown, made or tested in the physical world; their market is regulated; experimentation is expensive and slow; they are at an early stage, before demand is confirmed by binding orders; and a market exists in which capacity and expertise can be rented. Food biotech meets all five. Industrial biotechnology, advanced materials and parts of pharmaceutical development meet most of them.

The theory should not be expected to hold for pure software companies, where the belt is thin and the clocks are fast; for established companies as a whole, where it applies first to new units; or in jurisdictions where data rights in contracts cannot be enforced, since the Memory Clause then has nothing to rest on.

8.2 Limitations

Five limitations should be stated plainly. First, the study is conceptual: its propositions are derived, not tested. Second, part of the evidence is analogical; the semiconductor case shows a mechanism that fermentation may or may not repeat. Third, the evidence for the bench-to-tank discount rests on a small number of studies with different methods. Fourth, visible failures in fermentation are not a random sample of ventures, so they illustrate P3 without testing it. Fifth, the author's roles in FoodEdge, at MycoVibes Biotech and as an evaluator create interests in how the phenomenon is described, which is why independent analysis is part of the agenda below.

A sixth limitation concerns the speed of the field. Between February and June 2026 alone, several working papers proposed new accounts of AI and the boundaries of the firm, including the headless firm (Klein and Wieczorek, 2026), the minimum viable firm (Li and Shi, 2026) and accountability boundaries (Hydari and Muzaffar, 2026). The claims of novelty made here (the epistemic layering of agents, the judgement core in physical and regulated science, and the Memory Clause) will need to be checked again against the literature before formal publication.

8.3 Research agenda

The first step for P3 can begin immediately, because the data already exist. It should be carried out with an academic partner who has no stake in FoodEdge or in any venture studied, and the coding rules should be published before the results.

9. Conclusion: the protected centre has moved, and the question now is who holds it

The essay set out to answer three questions. Its answers, offered for testing, are these.

RQ1. What should a science-based venture keep inside its boundary? Its judgement core, the consequential decisions and the competent, accountable people who make them, together with the rights to the data its experiments produce. Most of the rest, including production, can and usually should be rented until demand is confirmed.

RQ2. How should AI agents be arranged? In orbits distinguished by the kind of knowing they do, experimentation, keeping state and coordination, with authority to act rising with distance from the core and falling as decisions become irreversible.

RQ3. When does renting create value rather than leak it? When the venture holds rights to the data from rented runs, when the capacity it rents is instrumented to return that data, and when its partners are bound by a relationship of trust that makes competing with their customers unattractive. Without these, renting transfers the venture's learning to the owner of the steel.

For half a century, organisation theory told companies to protect the factory at their heart. For a growing class of science-based ventures, that advice no longer describes where the risk lies. The factory can now be rented; judgement still has to be owned. The Copernican Firm is a name for what follows when a company accepts that change and arranges itself around it. Whether the theory survives contact with data is the next question, and the most useful thing an essay of this kind can do is make that question easy to ask.

The factory can now be rented; judgement still has to be owned.

References

All references checked against their DOI or source page on 29 October 2026.

Companion essays: Judgement Is Not for Rent; From Technical Core to Judgement Core: The Copernican Firm.