Topology Without Topography

Abstract#

Melvin Conway’s 1968 paper How Do Committees Invent? establishes a structure-preserving correspondence between the architecture of a designed system and the structure of the design organisation that produced it, and develops, in the remainder of the paper, a vertical reading of that correspondence under which conventional management practice, operating on action-cycles measured in weeks and quarters against the accumulated tacit knowledge of senior engineers and the mutual model among specific people that gives a design organisation its coherence (which accumulates on cycles measured in years and decades), propagates fragmentation across the layers of an organisation and into the artefact the organisation builds. The result, since named Conway’s Law, has been taken up across the software-engineering and organisational-design literatures over more than five decades, in renderings ranging from Brooks’s The Mythical Man-Month (Brooks, 1975) to the contemporary patterns literature on team configuration, with the dominant contemporary articulation being Skelton and Pais’s Team Topologies (Skelton and Pais, 2019). The reception has preserved Conway’s horizontal homomorphism and lost the vertical reading the 1968 text develops alongside it; the dominant rendering operates entirely at the layer of communication paths between teams of the same kind, and proposes patterns of team configuration as Conway’s actionable form.

A close reading of Conway’s text recovers what these renderings have elided. Conway’s paper, beyond the homomorphism, develops a disintegration thesis under which large systems disintegrate during development through three steps, the second of which is the fragmentation of the design organisation’s communication structure under the conventional management practice his paper identifies as the institutional condition driving the disintegration. Conway specifies the accounting fallacy that supports the practice (the treatment of human effort as linear and fungible in dollars-per-hour terms), names the incentive structures across vertical layers of an organisation that subvert the intent of its sponsors, observes that no design group can be both organised and unbiased because organising it has already committed the design, and directs his reader to Galbraith’s The New Industrial State (Galbraith, 1967) as the fuller treatment of the corporate technostructure. These claims are vertical: they concern propagation across layers of an organisation, not interaction between teams of the same kind. The contemporary framework rendering preserves the homomorphism and loses the rest.

The vertical reading of Conway’s Law the present paper offers turns on a distinction between topology and topography. Topology is the structure the chart can carry: team configurations, interaction modes, per-role and per-edge ledgers, the box-and-arrow representation a framework prescribes and a dashboard reads. Topography is the lived ground on which the topology’s labels sit: the tacit model of how the work behaves, the mutual knowledge among specific people accumulated through sustained shared effort, the apprenticeship transmission that no curriculum reaches, the slow variable of coordination whose accrual and decay operate on timescales the chart does not represent. The distinction recurs across traditions the paper draws on: as Polanyi’s distinction between explicit and tacit knowledge (Polanyi, 1958, 1966), as the cybernetic distinction between the model of a system and the system itself (Ashby, 1956; Conant and Ashby, 1970), as Bastiat’s distinction between what is seen and what is not seen (Bastiat, 1850).

The diagnostic apparatus is developed in two registers. The cybernetic register composes the Law of Requisite Variety (Ashby, 1956), the Conant-Ashby theorem on the regulator as model (Conant and Ashby, 1970), and the rate-distortion theory of lossy source coding (Shannon, 1949; 1959) into the horizon-mismatch theorem, by which a regulator whose action-cycle (a quarterly review, a budget round, an earnings cycle) is shorter than the replenishment-cycle of the system it acts on by more than a factor of order one cannot, by construction, model that system at its own temporal resolution. The system in the present application is the design substrate of an engineering organisation: the body of accumulated tacit knowledge, the mutual mental model of how the product behaves, the certification expertise of senior engineers, the calibration to each other’s blind spots that two engineers build by working together across a project’s lifecycle. This substrate accrues across years and decades, the apprenticeship cycle of a craft and the careers of the engineers who carry it. The theorem is the formal mechanism of Conway’s second disintegration step: management operating at quarterly cycles cannot stabilise design coherence that takes a decade to build, and the disintegration of that coherence must proceed inside the engineering layer itself, producing degradations the management regulator does not register and does not address. What is depleted is what Polanyi names the tacit dimension; what transmits it is apprenticeship and proximity; the framework form, by its genre as a book of prescribed configurations, can describe the tacit dimension but cannot install it.

The formal core develops the same finding from cooperative game theory. Any characteristic-function representation of organisational value admits a unique decomposition into Harsanyi dividends (Harsanyi, 1959, 1963; following von Neumann and Morgenstern, 1944; Shapley, 1953), of which the higher-order dividends are relational quantities attributable to no member alone. A per-role valuation, equivalent to the form of any per-person, per-team, or per-edge ledger, represents the value faithfully if and only if every higher-order dividend vanishes; for any organisation with any relational structure, the per-role instrument is mathematically blind to that structure of either sign. The result holds independently by the partial information decomposition of Williams and Beer (Williams and Beer, 2010), in which the synergistic atom of joint information is destroyed by per-member marginalisation. The decision-driving consequence follows: a reallocation driven by a per-role instrument optimises a ledger constant in every higher-order dividend, and under the conditions in which existing coalitions carry positive dividends and the reallocation dissolves them faster than new coupling forms to replace them, the relational stock falls while the ledger registers a gain. This is the cannibalisation result. A further blindness operates on the artefact itself: the period flow the dashboard reads is computed through projections that discard the dimension along which the artefact’s quality varies, so the booked figure and the built thing can come apart and the organisation can lose sight of its own product.

The framework engaged is Team Topologies (Skelton and Pais, 2019), read as the dominant contemporary form of the wider literature and at the level of its own apparatus rather than at the level of its misuse. The reading sustains two distinct findings. The first concerns coverage: even taken at full strength and applied without departure from its own prescription, the framework’s apparatus operates entirely on topology and contains no representation of the registers the present paper identifies as load-bearing for organisational coordination: the vertical incentive structure across layers, the temporal cycle on which engineering coherence accumulates, the tacit dimension Polanyi names, the relational mass the Harsanyi decomposition locates. The framework reaches part of what alignment requires; other parts lie outside the form it can carry. The second concerns the framework’s self-presentation as a holistic solution to organisational design. The presentation cannot be sustained alongside the coverage finding, and the further issues the paper documents follow from the partial apparatus being treated as complete: a horizontal grammar prescribed against a phenomenon whose substance is vertical, with the prescription’s confidence underwritten by an instrument whose blindness the formal core proves. The empirical anchors are the Boeing 737 MAX certification across 1997 to 2019 (House Committee on Transportation and Infrastructure, 2020; Joint Authorities Technical Review, 2019; Robison, 2021; Gelles, 2022) and the Wells Fargo cross-selling collapse (U.S. Securities and Exchange Commission, 2020; Stanford Graduate School of Business, 2019), in which the depletion the theorem specifies is observable across the worker and product channels in the first case and the customer channel in the second. The counterexample is the Bernbach pairing at Doyle Dane Bernbach from 1949 onward (Cracknell, 2011; Levenson, 1987), in which the vertical translation layers between intent and execution were collapsed into a unit holding the meaning from the start, an arrangement the framework’s horizontal grammar has no name for.

The constructive implication concerns the direction in which a reorganisation can proceed when an organisation’s product and teams have fallen out of alignment. The topology can be redrawn from above; the topography cannot. A reorganisation conducted as a top-down operation against a per-role dashboard runs on an instrument the paper proves cannot represent the accumulated tacit knowledge, the mutual mental model, and the relational coordination that the dashboard’s metrics rest on, and the depletion the cannibalisation result describes follows under the stated conditions. The direction in which that body of accumulated coordination can be addressed runs from the work upward: from a worked understanding of the product, the engineering substrate it depends on, and what the people who have done the work have learned to do, in a practice transmitted by proximity rather than by prescription. The paper is not an argument that organisations cannot be measured or that they should not be reorganised when their work has fallen out of alignment; it is an argument that the apparatus the dominant framework rendering supplies operates on a layer at which the work of reorganisation cannot, by construction, be conducted alone, and that the constructive direction follows from the apparatus’s own limits read carefully.

Keywords: Conway’s law, topology and topography, horizon mismatch, requisite variety, Conant-Ashby theorem, tacit knowledge, Polanyi, Harsanyi dividends, cooperative game theory, cannibalisation, three-channel extraction, bottom-up reorganisation, Team Topologies, organisational design, Boeing 737 MAX, Wells Fargo.

Part I: Topology and Topography#

1.1 Two Words#

Every organisation is read at one layer and produced at another. The layer at which it is read is the layer the chart can carry: the team configuration, the metric, the box-and-arrow diagram, the per-role ledger a dashboard records, the structure that survives transcription onto a slide. The layer at which it is produced is the one the chart does not record: the mutual knowledge among specific people, the tacit model of how the work actually behaves, the slow variable of coordination that accrues through sustained shared effort and decays without it, the apprenticeship transmission that no curriculum can reach. The two are not the same layer, the second is not what the first becomes when read carefully, and the failure to hold them apart is what the paper engages with.

The terms I use for them are topology and topography. Topology is what the chart carries; topography is what the chart’s labels sit on. I claim no novelty for the distinction. It is, in different vocabulary, Polanyi’s distinction between explicit and tacit knowledge (Polanyi, 1958, 1966); it is the cybernetic distinction between the model of the system and the system itself (Ashby, 1956; Conant and Ashby, 1970); it is Bastiat’s distinction between what is seen and what is not seen (Bastiat, 1850). What I do claim is the operational vocabulary in the form I give it here and the application of the distinction to a specific case: the framework response to Conway’s Law that has organised itself across the past two decades of organisational-design literature, and the demonstration that the framework form, by its genre, operates on topology while leaving topography outside what it can carry.

The asymmetry between the two layers does the load-bearing work of the argument. Topology lies entirely on the horizontal plane: who reports to whom, who interacts with whom through which mode, what shape the boxes take on the chart. Topography lives in the vertical dimension: what travels up and down through the layers of the organisation as the same vocabulary refers to different things at different layers, and whether the layers are one communication structure or several stacked on top of each other. The framework’s silence on the vertical dimension is not a small omission; it is the silence the paper is written to break.

The next sections develop the distinction by reading Conway (1968) carefully, recovering what the popular rendering of “Conway’s Law” has elided.

1.2 Conway’s Paper, In Full#

1.2.1 Provenance and Reception#

Melvin Conway’s 1968 paper How Do Committees Invent? appeared in Datamation magazine in April 1968 (Conway, 1968). The thesis statement, by Conway’s own later admission, sits in the third-last paragraph: “any organization that designs a system (defined more broadly here than just information systems) will inevitably produce a design whose structure is a copy of the organization’s communication structure” (Conway, 1968, concluding paragraphs; Conway’s author’s note to the online edition).

The paper had been submitted first to the Harvard Business Review, which rejected it on the grounds that Conway had not proved his thesis. Datamation published it (Conway, 1968; Mealy, 1968 on the paper’s reception). George Mealy, in a paper presented at the National Symposium on Modular Programming in July 1968, first named the result “Conway’s Law,” paraphrasing it as “systems resemble the organizations that produced them” (Mealy, 1968). Fred Brooks, in The Mythical Man-Month (Brooks, 1975), cited the paper and the law by name, which is the channel through which the law entered the broader software-engineering literature.

The popular rendering of Conway’s Law, in the form it now circulates across the framework literature, is the third-last paragraph: systems mirror the communication structures of the organisations that produce them. The contemporary response treats this as a horizontal statement about communication paths and proposes patterns of team configuration as the actionable form (Skelton and Pais, 2019; for the broader uptake, see MacCormack, Baldwin, and Rusnak, 2012; Skelton and Pais’s own Team Topologies draws explicitly on Conway as foundational).

The reading the present paper proposes recovers what the rendering omits, and what it omits carries the weight of Conway’s argument: the incentive-subverts-intent claim, the disintegration thesis, the accounting-fallacy critique, and the explicit reference to Galbraith’s analysis of the corporate technostructure. Each of these is in Conway’s text. None of them is engaged by the framework literature that invokes him.

1.2.2 The Homomorphism#

The central formal claim of Conway’s paper is the homomorphism between the graph of a designed system and the graph of the design organisation that produced it. Conway sets this up carefully through a sequence of moves the paper now follows.

Conway begins with linear graphs as the abstract representation of systems. A system is a set of subsystems connected by interfaces; the linear graph has nodes for subsystems and branches for interfaces (Conway, 1968). The representation is recursive: each subsystem is itself a system, decomposable down to whatever level of granularity the analysis requires.

The linear-graph notation, Conway notes, “provides an abstraction which has the same form” for the design organisation and the system it designs (Conway, 1968, “Relating the Two”). Replace “system” by “committee,” “subsystem” by “subcommittee,” “interface” by “coordinator,” and the same graph notation describes the design organisation that the original graph described as a system. The two graphs are formally comparable because they have the same shape of representation.

Conway then asks the central question of the paper: “Is there any predictable relationship between the graph structure of a design organization and the graph structure of the system it designs?” His answer: “Yes, the relationship is so simple that in some cases it is an identity” (Conway, 1968, “A basic relationship”). The proof he offers runs through the nodes and branches of the two graphs in turn. For any node $x$ in the system, there is a design group $X$ in the design organisation that designed $x$. For any pair of nodes $x$ and $y$ in the system, either there is a branch between them (they communicate in some way meaningful to the operation of the system) or there is not. If there is a branch, the two design groups $X$ and $Y$ must have negotiated and agreed on an interface specification. If there is no branch, there was nothing to negotiate, and there is no branch between $X$ and $Y$ in the design organisation graph.

The conclusion: “in the not unusual case where each subsystem had its own separate design group, we find that the structures (i.e., the linear graphs) of the design group and the system are identical. In the case where some group designed more than one subsystem we find that the structure of the design organization is a collapsed version of the structure of the system, with the subsystems having the same design group collapsing into one node representing that group” (Conway, 1968). The structure-preserving relationship between the two graphs is a homomorphism, not an isomorphism. The collapse is in the direction from system to design organisation: subsystems sharing a design group collapse into a single node.

Conway is making a mathematical claim, not a heuristic one. The system graph is a homomorphic image of the design organisation graph. The paper builds on this directly. The horizon-mismatch theorem developed in §2.2 below extends the homomorphism across time: where Conway treats the homomorphism at a moment, the theorem specifies what happens to the homomorphism when the regulator (the design organisation, in Conway’s terms) and the system operate at incommensurable cycles.

1.2.3 Systems Image Their Design Groups#

The homomorphism has a consequence Conway draws out in the section titled “Systems Image Their Design Groups.” If for any system requirement there is a family of designs that meet the requirement, the choice of design organisation determines which design from the family will be selected. Conway puts the consequence as a generality about the relation between organisations and their products: “To the extent that an organization is not completely flexible in its communication structure, that organization will stamp out an image of itself in every design it produces. The larger an organization is, the less flexibility it has and the more pronounced is the phenomenon” (Conway, 1968).

Two examples in Conway’s text make the claim concrete. A contract research organisation had eight people to produce a COBOL and an ALGOL compiler. Five were assigned to COBOL, three to ALGOL. “The resulting COBOL compiler ran in five phases, the ALGOL compiler ran in three” (Conway, 1968). The design split mirrored the team split. Second, “two military services were directed by their Commander-in-Chief to develop a common weapon system to meet their respective needs. After great effort they produced a copy of their organization chart” (Conway, 1968). The two-service split mirrored the two-service design.

Conway adds a third case that the framework literature has been less attentive to. A computer system in operation consists of three parts: hardware, system software, application. Corresponding to these are three designer groups: manufacturer’s engineers, system programmers, application programmers. Conway’s observation: “Those rare instances where the system hardware and software tend to cooperate rather than merely tolerate each other are associated with manufacturers whose programmers and engineers bear a similar relationship” (Conway, 1968). This is a positive case of the homomorphism. When the human relationship is healthy, the system relationship is healthy. The homomorphism is not just a statement of how dysfunction propagates; it is also a statement of how functioning coordination at the human level is what produces functioning coordination at the system level.

The framework literature has read the homomorphism primarily as a constraint to be managed, with the instruction taking the form “choose the team structure to produce the architecture wanted” (the “reverse Conway manoeuvre” of Forsgren, Humble, and Kim, 2018, and others). What the framework literature has missed is that the homomorphism implies its converse: when you redraw the team structure without addressing what was making the prior team structure produce coherent designs, you redraw the dysfunction along with the chart.

1.3 What the Popular Rendering Omits#

The framework literature’s rendering of Conway typically stops at the homomorphism. The remainder of Conway’s paper, which develops the implications of the homomorphism at length, has received much less attention. The remainder contains the substantive critique of conventional management practice that the contemporary framework literature inherits without engaging with.

1.3.1 The Incentive-Subverts-Intent Claim#

Conway’s third paragraph contains a claim that carries much of the argument’s weight and that has been largely ignored. Discussing the choices a designer makes during a design effort, Conway writes: “Many of these choices may be more than design decisions; they may also be personal decisions the designer makes about his own future. As we shall see later, the incentives which exist in a conventional management environment can motivate choices which subvert the intent of the sponsor” (Conway, 1968, opening section, citing his own forthcoming “System Management” section).

Conway is making a vertical claim here, naming explicitly that the incentive environment at one layer of the design organisation (the manager’s incentive structure, the budget pressure, the schedule risk) motivates choices at the design layer that subvert the intent of the sponsor at a still-higher layer (the originator of the design effort, the recipient of the eventual system). Three layers are implicit in Conway’s formulation: sponsor, manager, designer. The claim is that incentives at the manager layer subvert the sponsor’s intent through the choices the designer is forced to make.

The framework literature does not engage with this claim. The framework’s response to Conway operates at the horizontal layer (between teams of the same kind, e.g. stream-aligned teams, platform teams) and has no apparatus for the vertical claim about incentive structures across layers. This is the first significant omission.

1.3.2 The Pre-Commitment of Design Boundaries#

Conway makes a further claim that the framework literature is in a particular position to be unable to hear. “We shall see in detail later that the very act of organizing a design team means that certain design decisions have already been made, explicitly or otherwise. Given any design team organization, there is a class of design alternatives which cannot be effectively pursued by such an organization because the necessary communication paths do not exist. Therefore, there is no such thing as a design group which is both organized and unbiased” (Conway, 1968, “Stages of Design”).

Conway’s claim is that organising a design team commits the design before the design begins. The class of design alternatives the team can effectively pursue is bounded by the communication paths the team’s organisation makes available. There is no neutral starting point. A framework that prescribes a team topology in advance has, by Conway’s own argument, committed design decisions before the design has begun, and those decisions cannot be undone within the framework’s own grammar.

The framework’s vocabulary contains no apparatus for engaging with this claim. The framework proposes its team topology as if it were neutral with respect to the work the topology will produce. Conway’s claim is that no topology can be neutral with respect to its work.

1.3.3 The Disintegration Thesis#

Conway’s most substantial claim, occupying the longest section of his paper, is what the paper calls the disintegration thesis. Conway opens the “System Management” section with a sentence that should be the central citation of any framework purporting to address Conway: “The structures of large systems tend to disintegrate during development, qualitatively more so than with small systems” (Conway, 1968, “System Management”).

Conway proceeds to specify a three-step process by which the disintegration occurs:

“Why do large systems disintegrate? The process seems to occur in three steps, the first two of which are controllable and the third of which is a direct result of our homomorphism.

First, the realization by the initial designers that the system will be large, together with certain pressures in their organization, make irresistible the temptation to assign too many people to a design effort.

Second, application of the conventional wisdom of management to a large design organization causes its communication structure to disintegrate.

Third, the homomorphism insures that the structure of the system will reflect the disintegration which has occurred in the design organization” (Conway, 1968, “System Management”).

The three steps deserve careful reading. Step one is overpopulation: under schedule pressure and the threat of mismanagement charges, the initial designer is forced to delegate before the system has been reduced to comprehensibility. Step two is disintegration: conventional management practice, applied to the now-overpopulated design organisation, fragments the communication structure that would have made coherent design possible. Step three is transmission: the homomorphism transmits the now-fragmented communication structure to the system, which is built as the image of the fragmentation.

The disintegration thesis is the load-bearing claim of the present paper. The horizon-mismatch theorem developed in §2.2 below specifies, in formal terms, why the second step of Conway’s disintegration occurs: conventional management practice operates at a cycle (schedule pressure, budget management, quarterly review) that is incommensurable with the cycle at which design coherence actually accumulates (sustained shared work over months and years). When the management apparatus operates against this design substrate at a mismatched cycle, the management apparatus cannot, by construction, regulate it at the cycle the substrate’s coherence accumulates on. The communication structure that would have carried the coherent design across the design organisation disintegrates because the management apparatus cannot register the coherence at the cycle the management apparatus operates at.

Conway diagnoses the disintegration without having the cybernetic apparatus to specify its mechanism. The cybernetic apparatus was available in 1968 (Ashby’s Introduction to Cybernetics had been published in 1956), but Conway did not draw on it. The paper draws on it now, and what follows is the formal extension of Conway’s diagnosis.

1.3.4 The Accounting Fallacy#

Embedded in Conway’s discussion of the disintegration thesis is a critique of conventional management accounting that has been largely lost in the framework literature’s uptake. Conway sets up the case of a manager who must choose between two contractors: a small new organisation with an intuitive cheap approach, and an established conventional outfit asking a more “realistic” fee. The manager’s incentive structure pushes toward the established outfit: failure of the established outfit is evidence that the problem was hard; failure of the small new outfit is the manager’s fault.

Conway then identifies the cause as institutional rather than personal: “A large part of it relates to the kind of reasoning about measurement of resources which arises from conventional accounting theory. According to this theory, the unit of resource is the dollar, and all resources must be measured using units of measurement which are convertible to the dollar. If the resource is human effort, the unit of measurement is the number of hours worked by each man times his hourly cost, summed up for the whole working force” (Conway, 1968, “System Management”).

This claim does considerable work in Conway’s argument. Conway is identifying conventional accounting theory as the institutional condition that drives the disintegration. The conversion of all resources to dollars, and of human effort to hours-times-rate, treats the resource as linear and fungible. Conway’s homomorphism shows that the linearity is false: “two men working for a year or one hundred men working for a week (at the same hourly cost per man) are resources of equal value” only if the homomorphism is ignored. The two configurations produce different design organisations and therefore different system designs. “Assumptions which may be adequate for peeling potatoes and erecting brick walls fail for designing systems” (Conway, 1968, “System Management”).

The accounting fallacy is what the horizon-mismatch theorem and the three-channel extraction structure of the present paper specify in formal terms. Conventional accounting measures inputs at one cycle (the dollar-per-hour calculation, repriced quarterly) while the tacit coordination the inputs are meant to produce accrues at a different cycle, measured in years across the working life of an engineering team. The metric registers progress because the metric measures what it can register; the tacit coordination depletes because it is invisible to the metric. The metric is kept favourable by spending exactly what it cannot see. This is what the three-channel extraction structure specifies in §2.4.

1.3.5 Parkinson’s Law and the Galbraith Reference#

Conway closes the disintegration section with two further moves the framework literature has not engaged with.

The first is Parkinson’s Law. Conway cites Parkinson directly: “Parkinson’s Law and Other Studies in Administration” (Parkinson, 1957) as the institutional account of why managers expand their organisations. “As long as the manager’s prestige and power are tied to the size of his budget, he will be motivated to expand his organization. This is an inappropriate motive in the management of a system design activity. Once the organization exists, of course, it will be used. Probably the greatest single common factor behind many poorly designed systems now in existence has been the availability of a design organization in need of work” (Conway, 1968, “System Management”).

Parkinson’s Law operates at a vertical layer (managerial incentive) and produces consequences at a different vertical layer (design quality) through the mechanism the homomorphism specifies. The framework literature has not engaged with the Parkinsonian dynamic because the framework’s apparatus does not represent the managerial incentive layer.

The second is the reference to Galbraith. Conway’s first footnote reads: “A related, but much more comprehensive discussion of the behavior of system-designing organizations is found in John Kenneth Galbraith’s The New Industrial State (Boston, Houghton Mifflin, 1967). See especially Chapter VI, ‘The Technostructure’” (Conway, 1968, footnote 1, citing Galbraith, 1967).

Galbraith’s Technostructure analysis is the vertical reading the present paper develops in formal terms. The technostructure, in Galbraith’s account, is the layer of corporate decision-making that operates between executive ownership and operational work, with its own incentive structure, its own knowledge base, and its own goals that may diverge from both the layers it sits between. Galbraith’s claim is that contemporary corporations are run by the technostructure, not by ownership or by the workforce, and that the technostructure’s incentives are systematically different from both. Conway cites this as “much more comprehensive” than his own treatment. The framework literature that invokes Conway has, on the whole, not followed the Galbraith reference.

1.4 Where the Paper Builds on Conway, Extends Him, and Diverges From Him#

The paper’s relation to Conway is not adversarial. The substantive claims at stake are already in Conway’s text, and the present work extends them by deploying apparatus (cybernetic, Polanyian, mathematical) that Conway did not deploy but that is compatible with his claims and that gives his claims a form sharper than the form in which Conway left them.

The paper builds on Conway in three specific ways. First, it accepts the homomorphism as the foundation. The system graph is a homomorphic image of the design organisation graph; rearranging the design organisation rearranges the system. This is Conway’s central formal claim and the paper does not contest it. Second, it accepts the disintegration thesis: large systems disintegrate because conventional management practice fragments the communication structure, and the fragmentation is transmitted by the homomorphism. This is Conway’s central diagnostic claim and the paper deploys it as load-bearing for the empirical cases of Part IV. Third, it accepts the accounting-fallacy critique: conventional management accounting treats human effort as linearly fungible, which is false by the homomorphism. This is Conway’s central methodological claim and the paper extends it in §2.4 with the three-channel extraction structure.

The paper extends Conway in three specific ways. First, it gives Conway’s disintegration thesis its cybernetic formalisation. The horizon-mismatch theorem of §2.2 specifies why the second step of Conway’s disintegration occurs: conventional management operates at a cycle incommensurable with the cycle at which design coherence accumulates, and by Ashby’s Law and Conant-Ashby’s theorem, the management apparatus cannot regulate design coherence at the cycle on which design coherence accrues. Conway diagnoses the disintegration; the paper specifies its mechanism. Second, it gives Conway’s incentive-subverts-intent claim its Polanyi extension. The dimension of organisational knowledge that fails to survive vertical translation is the tacit dimension (Polanyi, 1958, 1966); Conway names the incentive subversion without naming the tacit-knowledge loss that the subversion produces; §2.3 supplies the missing link. Third, it gives Conway’s accounting-fallacy critique its three-channel structure. The body of accumulated coordination the false accounting depletes is depleted along three channels (the workforce, the product, and the customer base); §2.4 specifies how each channel manifests empirically.

The paper diverges from Conway in three specific places, none of them substantive disagreements with Conway’s text. First, Conway treats the design organisation as one entity. The paper distinguishes between the management layer (incentives, schedule, budget) and the design layer (the actual work), because the horizon-mismatch theorem requires the distinction. Conway’s three-step disintegration mixes these layers in a way that obscures the mechanism; the paper separates them to make the mechanism visible. Second, Conway’s solution is “flexibility of organization” (Conway, 1968, “Conclusion”). The paper takes this as necessary but insufficient. A framework that prescribes flexibility at the horizontal layer does not address vertical horizon mismatch. The solution Conway’s text gestures toward but does not develop is the disposition that responds to the ground rather than prescribing from above; this is the conclusion Part V develops. Third, Conway’s text does not engage with frameworks of the kind the present paper critiques, because such frameworks did not exist in 1968. The framework critique is the paper’s own work, conducted in light of Conway’s claims but going beyond them in scope.

The framework literature’s rendering of Conway is incomplete on every one of these points. The framework’s apparatus represents Conway as a horizontal homomorphism between teams and architectures and proposes team-topology patterns as the response. Conway’s actual paper contains the vertical claim (incentive subverts intent), the disintegration thesis (the homomorphism transmits dysfunction across layers), the accounting-fallacy critique (conventional accounting fails at the design task on its own terms), the Galbraith reference (the technostructure operates at a vertical layer the horizontal apparatus does not represent), and the explicit limitation of his own solution (flexibility, on Conway’s account, requires a philosophy of system-design management that does not yet exist). The framework literature has inherited the vocabulary and not the content.

1.5 What Frameworks Cannot Reach#

The paper’s substantive claim, stated in its strongest form, is this. Any organisational framework presented as a universal pattern, transmitted through the genre of the book, prescribing team configurations and interaction modes that can be drawn on a chart and applied to any organisation, is operating entirely at the horizontal layer of communication structure. The framework’s apparatus has no representation of the vertical layer where Conway located his layered diagnosis. The framework’s success conditions, even followed in full, do not address what makes the layers cohere or come apart. The framework can shape the surface graph; it cannot reach what makes the surface graph carry meaning.

Three sub-claims follow, each derivable from Conway’s text with the formal apparatus the paper supplies.

The first sub-claim is the tacit-knowledge claim. What carries meaning across the vertical layers of an organisation is a form of knowledge that does not survive transcription into the form of a framework. Polanyi’s bicycle example is the canonical illustration: we can know more than we can tell (Polanyi, 1966, opening). We know how to ride the bicycle without being able to articulate the rules of riding it; the articulated rules are post-hoc reconstructions of something the body already knows how to do. An organisation that has learned to coordinate across layers has done so through what apprenticeship has always transmitted: proximity, shadowing, watching, doing under supervision, being corrected when wrong, absorbing through closeness what could not be transmitted through instruction (Polanyi, 1958, on the tacit foundations of skill; Polanyi, 1966, on the apprenticeship form). The framework’s prescription is, by construction, in the form of instruction. It cannot transmit what the apprenticeship form transmits. Following the framework is not following the way; it is reading the book about the way.

The second sub-claim is the universal-prescription claim. The framework presents itself as applicable to any organisation. The presentation is unwarranted by the framework’s empirical base. Team Topologies (Skelton and Pais, 2019) was generated from consulting engagements with mid-to-large software companies, from the DevOps research community’s work on continuous delivery (Forsgren, Humble, and Kim, 2018), and from a sociotechnical-systems tradition that originated in mid-twentieth-century studies of coal mining (Trist and Bamforth, 1951). A narrow empirical base carries a wide generalisation, and the cases the generalisation reaches do different kinds of work. Advertising agencies need creative pairs of art directors and copywriters working as a single conceptual unit, a model Bernbach developed at Grey Advertising in the 1940s and brought to Doyle Dane Bernbach at its founding in 1949 (Bernbach’s biographical record at DDB Worldwide; Cracknell, 2011 on the creative-pair innovation; Levenson, 1987 on the DDB house style). The Team Topologies grammar has no name for this. Architectural studios need a principal-and-team configuration in which the principal carries the conceptual integrity of the project across the team’s execution. Surgical teams need a specific role asymmetry between the lead surgeon and the supporting roles that the framework’s “stream-aligned team” vocabulary does not address. Orchestras need section structures that are functional, not stream-aligned. Each of these is a unit of work the framework’s grammar cannot describe.

The third sub-claim is the regulator-without-its-model claim. The framework, considered as a regulator of the organisation that adopts it, must by the Conant-Ashby theorem contain a model of the system it regulates (Conant and Ashby, 1970). The framework offers a model of team structure and team interaction. It does not offer a model of what the organisation is producing, what its history is, what its market is, what its constituents know that is not in the framework. The framework prescribes the regulator without providing the model the regulator must contain. By the mathematics of cybernetic control, a regulator without a model of the system it regulates cannot regulate the system. It can shape one input to the system (the team configuration) and hope that the rest follows. The hope is the substitute for the model the framework does not contain. When the framework comes apart in adoption, what comes apart is not the implementation but what the framework’s form can carry. The framework’s form does not meet the formal requirement that the cybernetic tradition itself established.

The three sub-claims compound. Organisational design is not the kind of thing that yields to formula because the substance of what makes an organisation function lives in the vertical dimension the formula cannot reach. The horizontal apparatus is the apparatus the framework can supply. The vertical substance is what the framework cannot supply, and the substitution of the apparatus for the substance is what the framework’s form, by being a framework, performs.

1.6 The Plan of the Paper#

Part I has established the topology-topography distinction, read Conway (1968) in full to recover what the popular rendering omits, and stated the three sub-claims that organise what follows in Parts II through V.

Part II develops the diagnostic framework. Ashby’s Law of Requisite Variety (Ashby, 1956) and the Conant-Ashby theorem (Conant and Ashby, 1970) are stated as the cybernetic background. The horizon-mismatch theorem is derived in full, with its proof, its formal restatement, and its mismatch-ratio specification. The Polanyi extension (Polanyi, 1958, 1966) is developed: tacit knowledge is what fails to survive vertical translation, and the apprenticeship form is what successfully transmits it. The three-channel extraction structure is named: when vertical horizon mismatch is sustained, the metric is kept favourable by spending the accumulated coordination of the workforce, the integrity of the product, or the trust of the customer base, taken singly or in combination, and the channels can operate alone or in combination.

Part III is the formal core. The Harsanyi-dividend mathematics (Harsanyi, 1959, 1963), the invisibility theorem, and the cannibalisation theorem are stated in their full form. The mathematics is what gives the diagnostic its load-bearing form rather than leaving it at the rhetorical level. Where the topology can carry only the per-role and per-edge ledger, the relational structure lives in the higher-order dividends the ledger cannot represent. The invisibility theorem establishes that no per-role instrument can register those dividends, of either sign; the cannibalisation theorem establishes the conditions under which the reported gains from such instruments can be the visible image of the unseen stock being spent.

Part IV is the worked application. Team Topologies (Skelton and Pais, 2019) is read at the level of its own apparatus, not at the level of its misuse. The framework’s response to Conway is shown to operate entirely at the horizontal layer, with the vertical layer outside the framework’s representational scope. Boeing’s 737 MAX certification across the period 1997 to 2019 is presented as the case of engineering-culture cannibalisation with extraction operating primarily along the worker and product channels (House Committee on Transportation and Infrastructure, 2020; Joint Authorities Technical Review, 2019; Robison, 2021; Gelles, 2022). Wells Fargo’s cross-selling collapse is presented as the case of customer-trust cannibalisation with extraction operating primarily along the customer channel (U.S. Securities and Exchange Commission, 2020; Stanford Graduate School of Business case study, 2019). The Bernbach pairing at DDB in the 1950s is offered as the counterexample: a case in which the vertical translation layers (account-services-to-copywriter-to-art-director) were collapsed into a creative pair holding the meaning from the start, a move the framework’s grammar has no name for.

Part V supplies the answer the diagnostic implies. Topography is what the diagnostic protects, and the vertical dimension is what makes topography possible. Frameworks cannot reach this; the way of working that reaches it is not formulaic, is responsive to the ground, is transmitted by proximity rather than prescription. The conclusion is not that frameworks are useless. The conclusion is that frameworks operate at the layer they can operate at, and the layer where the substance lives is not that layer, and any framework that presents itself as a complete response to Conway’s Law is making the category error Conway’s own text already identified.

Part II: The Diagnostic Framework#

Part I read Conway (1968) in full and recovered the four substantive claims the framework literature has elided: the incentive-subverts-intent claim, the pre-commitment of design boundaries, the disintegration thesis, and the accounting-fallacy critique. The disintegration thesis is the load-bearing claim. Conway specified three steps but did not specify the mechanism by which the second step (conventional management practice fragmenting the communication structure) occurs. The cybernetic apparatus available in 1968 (Ashby, 1956) and elaborated in 1970 (Conant and Ashby, 1970) supplies that mechanism. Part II develops the apparatus and then extends it with the Polanyi material on tacit knowledge (Polanyi, 1958, 1966) and the three-channel extraction structure.

2.1 Requisite Variety and the Adequacy of a Regulator#

The cybernetic tradition begins with Ashby’s Introduction to Cybernetics (Ashby, 1956), the work in which the formal apparatus the present paper deploys was first systematically presented. Ashby’s central result for organisational analysis is the Law of Requisite Variety, which Ashby states in Chapter 11 of the Introduction (Ashby, 1956, Chapter 11, “Requisite Variety”).

Ashby’s Law of Requisite Variety. A regulator can hold a system stable only if its own variety, the number of distinct states it can assume and respond with, is at least as large as the variety of the disturbances it must counter. In Ashby’s own compression: only variety can absorb variety (Ashby, 1956, p. 207). The law is a constraint that applies to any regulator, mechanical or institutional. A thermostat with two states cannot regulate a system whose disturbance pattern has fifty distinguishable forms; the regulator’s variety is insufficient to the regulatory task, and forty-eight of the disturbance forms pass through the regulator without being addressed. The variety the regulator does not absorb does not cease to exist; it propagates into the system, with consequences the regulator does not register.

For organisational analysis the law applies directly. A management apparatus, a measurement instrument, an organisational framework: each is a regulator in Ashby’s sense. Each has a variety bounded by the distinctions it can represent. Each is operating against a system whose variety it is attempting to absorb. The law states the condition under which regulation can be achieved: regulation requires variety at least matching the system’s. Where the regulator’s variety is insufficient, the system’s variety passes through unregulated. The pass-through is not a failure of effort; it follows from the variety deficit itself.

The Conant-Ashby theorem. A decade after the Introduction, Conant and Ashby published “Every Good Regulator of a System Must Be a Model of That System” in the International Journal of Systems Science (Conant and Ashby, 1970, Vol 1, No 2, pp 89-97). The theorem is stronger than the Law of Requisite Variety. Where Ashby’s Law states a quantitative condition on regulator variety, Conant-Ashby states a condition on the regulator’s internal organisation: successful regulation requires the regulator to embody, in its own structure, a model of the regulated system. The regulator can be optimal only if it is such a model.

Conant and Ashby’s exact statement: “any regulator that is maximally both successful and simple must be isomorphic with the system being regulated” (Conant and Ashby, 1970, p. 89, abstract). The theorem operates under specifiable conditions Conant and Ashby give; the conditions hold for the cases the present paper engages with. The result has been subject to scholarly re-examination (Scholten, n.d., on the formal structure; Wentworth, 2021, on extensions; for the most recent reformulation see Virgo, Biehl, Baltieri, and Capucci, 2025); the re-examinations refine the theorem but do not displace its central claim, which is that regulation and modelling are joined at the formal level by the conditions the theorem specifies.

The consequence used throughout this paper follows immediately. Where the decisive variety of an organisation is carried in tacit relations, in the unwritten, mutually held knowledge of how the work and the people actually behave, a good regulator must embody or preserve an adequate model of those tacit relations. The theorem does not say the model must itself be tacit in the human sense; an explicit instrument might in principle represent some tacit dynamics indirectly, through proxies, ethnographic description, network and temporal indicators, or combined qualitative and quantitative evidence. What the theorem does say is that an explicit, data-derived instrument that does not contain a model of the relevant tacit relations is not a good regulator of the system whose decisive variety is tacit, however much explicit data it processes. The instrument can be operating perfectly within its own representational scope and still fail by the Conant-Ashby condition: it lacks the model the theorem requires.

The two results compose. Ashby’s Law gives the variety condition; Conant-Ashby gives the modelling condition. A regulator that satisfies the variety condition but not the modelling condition is a regulator with enough states but no map of where to send them. A regulator that has the model but lacks the variety to act on it is a regulator that knows the territory but cannot move on it. A regulator that has neither, which is the typical condition of the instruments the paper engages with, is operating outside both conditions the cybernetic tradition has identified as necessary for regulatory adequacy.

The cybernetic apparatus stops here in Ashby and Conant-Ashby. The temporal extension required for the present argument follows in the next section.

2.2 The Horizon-Mismatch Theorem#

The two cybernetic results above operate without explicit treatment of temporal scale. Ashby’s variety is the number of distinct states the regulator can assume; Conant-Ashby’s model is the correspondence between regulator and system that the theorem’s isomorphism condition specifies. Neither result, in its 1956 or 1970 form, specifies what happens when the regulator operates at one temporal scale and the system at a different scale. The horizon-mismatch theorem composes the cybernetic results with the rate-distortion theory of lossy source coding (Shannon, 1949; 1959) and produces a formal consequence the cybernetic tradition itself did not state.

The Conway connection runs through the disintegration thesis. Conway’s disintegration thesis identifies the second step of the disintegration as “application of the conventional wisdom of management to a large design organization causes its communication structure to disintegrate” (Conway, 1968, “System Management”). Conventional management practice operates on cycles measured in weeks or quarters (the schedule review, the budget cycle, the quarterly earnings report, the annual planning round). Design coherence, in the sense Conway is concerned with, accumulates on cycles measured in months and years (the project lifecycle, the apprenticeship cycle of an engineering practice, the multi-year sediment of tacit coordination among specific people). The mismatch between these two cycles is what the horizon-mismatch theorem formalises. Conway named the disintegration; the theorem gives its mechanism.

2.2.1 Statement of the Theorem#

I develop in this section a result I will call the horizon-mismatch theorem. The result is mine; the components from which I compose it are Ashby’s Law of Requisite Variety (Ashby, 1956), the Conant-Ashby theorem on the regulator as model (Conant and Ashby, 1970), and the rate-distortion theory of lossy source coding, including its reverse water-filling allocation (Shannon, 1949; 1959; Cover and Thomas, 2006). None of the three cited works states the composition, and the result is not in the cybernetic literature in the form I give it here. I extend the cybernetic framework into the temporal dimension the cited works leave implicit: where Ashby and Conant-Ashby treat the regulator-system pair as a synchronic problem of variety and model, I add the cycle dimension on which the regulator and the substrate operate, and derive the consequence that follows when the cycles stand in a particular ratio. Ashby’s, Conant and Ashby’s, and Shannon’s analytical responsibility extends to what they wrote; the composition and the consequence are mine.

A note on terminology before the theorem is stated. The cybernetic literature speaks of regulator-system pairs (Ashby, 1956; Conant and Ashby, 1970); the theorem uses system (S) and system-state and system-cycle in the formal statement, with substrate used interchangeably in the diagnostic deployment to track the specific kind of system at issue, a productive substrate whose replenishment-cycle is longer than the regulator’s action-cycle. The two terms are coextensive in the contexts the paper deploys them.

The Law of Requisite Variety and the Conant-Ashby theorem, applied to a regulator whose action-cycle is shorter than the system’s cycle by orders of magnitude, produce the following.

The Horizon-Mismatch Theorem. Let $R$ be a regulator with action-cycle $T_R$, and let $S$ be a system with replenishment-cycle $T_\Sigma$. Suppose $T_R$ is shorter than $T_\Sigma$ by more than a factor of order one. Then $R$ cannot, by construction, model $S$ at system-state resolution; and the variety of $S$’s disturbances must be absorbed within $S$ itself, producing system-state changes the regulator does not register and does not address.

The theorem’s antecedent is the condition $T_R \ll T_\Sigma$. The condition can be made specific by the mismatch ratio: $T_\Sigma / T_R$, dimensionless, computable as soon as the two cycles are identified. The ratio is large when the system evolves much more slowly than the regulator acts. The ratio is approximately one when the cycles are matched. The theorem’s antecedent applies whenever the ratio is large enough that disturbances at the system’s scale appear within the regulator’s cycle as approximately constant inputs the regulator cannot distinguish.

The cycle $T_R$ in this statement is the regulator’s action-cycle: the cycle at which it measures, evaluates, reallocates, and resets its operative account of the system’s state, the quarterly review, the budget round, the earnings cycle. It is not a passive sampling interval, and the theorem is not the claim that a regulator cannot observe a process slower than the rate at which it looks. A regulator that did nothing but observe could observe a slow substrate across many cycles and build up a picture of it. The regulator the theorem concerns is one that acts on its account of the system every cycle, and whose model is updated, evaluated, and rewarded only at that cycle, so that the operative state on which it acts is the state its action-cycle can resolve.

The statement also forecloses the response that would lift the result by adding to the regulator an explicit long-horizon representation of the substrate. The response treats the substrate as present in the regulator’s account at coarse resolution, so that finer representation would recover it. The substrate’s relevant variety is not present at coarse resolution; it is absent from the account’s form. The variety at issue is the relational stratum the substrate carries, and §3.3 establishes that this stratum lives in the higher-order dividends of the value functional, which no per-unit valuation represents, of either sign, the per-unit form being any assignment of a single figure to each unit with the value of a whole read as the sum across its units (Theorem 1, with the grain-generality §4.1 establishes). A regulator’s internal model evaluated at its action-cycle is such a per-unit account, and a richer dashboard, a longer-horizon metric, or an added leading indicator is another one, so each lands inside the form Theorem 1 proves blind rather than outside it. A regulator that genuinely carried the substrate as a represented state would not be carrying it in the per-unit, single-figure form the theorem concerns; it would be a regulator with a different account of value, a different objective, and a different action-cycle, which is the subject of the constructive direction Part V develops rather than a parameter the present regulator can set. The result is therefore not that a long-horizon model is impossible to hold, but that the short-cycle per-unit regulator the diagnosis concerns does not hold one and cannot acquire one without ceasing to be that regulator.

2.2.2 The Proof#

The proof proceeds through two lemmas and the main inference. The lemmas establish the conditions the Law of Requisite Variety and the Conant-Ashby theorem impose on any regulator-system pair; the main inference applies the conditions to the cycle-mismatched case.

Lemma 1 (Variety-Cycle Bound). A regulator $R$ operating at action-cycle $T_R$ has variety bounded by the variety of disturbances $R$ can register within $T_R$.

The regulator’s variety is not unbounded; it is constrained by construction by the temporal resolution at which the regulator operates. A regulator producing a single output per action-cycle has variety equal to the number of distinct outputs it can produce, which is the cardinality of its output space. A regulator whose temporal resolution is $T_R$ can register, within a single action-cycle, only the disturbances that produce distinguishable signals at $T_R$‘s resolution; disturbances operating at scales much longer than $T_R$ appear, within a single action-cycle, as approximately constant inputs, with the consequence that the regulator’s variety cannot distinguish the long-cycle disturbances from each other within the action-cycle. The bound is grounded in the rate-distortion theory of lossy source coding (Shannon, 1959; Cover and Thomas, 2006, ch. 10). A regulator producing one output per action-cycle is a code operating under a rate budget fixed by $T_R$. For a source whose components exhibit differing variance within the observation window, the optimal allocation of a fixed rate sets a single water level: components whose within-window variance stands above the level receive rate in proportion to how far they stand above it, and components whose within-window variance falls below the level receive no rate at all and are reconstructed at their means, which is to say as constants (the reverse water-filling allocation; Cover and Thomas, 2006, §10.3.3). A disturbance whose cycle is much longer than $T_R$ exhibits, within any single action-cycle, a within-window variance near zero; it falls below any positive water level, is carried at zero rate, and is registered as an approximately constant input. The crisp allocation picture is derived for source components treated as independent, and the application here does not require independence: the step doing the work is the timescale separation itself, the near-vanishing of the slow component’s variance within the window, which holds whatever the coupling between slow and fast components, and the coupled case is the one the present substrate instantiates by construction, since a higher-order dividend just is a coupling between dimensions. The sampling-theorem corollary of the earlier literature survives as the intuitive register of the same bound: state transitions occurring at a frequency the observer cannot resolve are aliased, and slow change is read as steady state (Shannon, 1949; Ashby, 1956, pp. 202-218, on the interaction of variety and temporal resolution; Wiener, 1948, on the temporal-resolution conditions of feedback systems). The lemma’s load is carried by the rate-distortion result; the sampling corollary is its intuition; and the lemma does not claim that organisational variety is a band-limited signal.

Lemma 2 (Modelling-Resolution Bound). A regulator $R$ whose internal states implement a model $M$ of a system $S$ has model $M$ at resolution equal to the resolution at which $R$’s internal states can be updated, which is bounded by $R$’s action-cycle $T_R$.

The regulator’s model has resolution not finer than the regulator’s update cycle; a model updated at $T_R$ cannot have resolution finer than $T_R$, because the model’s state changes are, by construction, the changes that occur in $R$‘s internal states across the update cycle. The system $S$ has its own temporal features; if the system’s variety operates at $T_\Sigma$, the model $M$’s resolution at $T_R$ cannot represent $S$‘s variety at $T_\Sigma$ when $T_R$ is shorter than $T_\Sigma$. The lemma is near-definitional and requires no external authority: a model whose state changes occur only at the update cycle cannot carry distinctions finer than the update cycle. Conant and Ashby’s theorem supplies the model requirement itself, that every good regulator of a system must be a model of that system (Conant and Ashby, 1970); the temporal extension of that requirement is the present section’s addition, as §2.2.1 states.

Main Inference. Let $R$ be a regulator with action-cycle $T_R$, and let $S$ be a system with replenishment-cycle $T_\Sigma$, with $T_R$ shorter than $T_\Sigma$.

By Lemma 1, $R$‘s variety is bounded by the variety $R$ can register within $T_R$; by the temporal features of the cycle-mismatch, this variety does not include the system’s variety at $T_\Sigma$, because the system’s $T_\Sigma$-scale variety appears within $R$’s $T_R$-scale action-cycle as approximately constant inputs $R$’s variety cannot distinguish.

By Lemma 2, $R$’s model $M$ of $S$ has resolution not finer than $T_R$; by the cycle-mismatch, $M$’s resolution cannot represent $S$’s variety at $T_\Sigma$.

By Ashby’s Law of Requisite Variety (Ashby, 1956), $R$’s stabilisation of $S$ requires $R$’s variety to equal or exceed $S$‘s disturbance variety; the cycle-mismatched regulator’s variety does not satisfy this condition.

By Conant and Ashby’s theorem (Conant and Ashby, 1970), $R$’s stabilisation of $S$ requires $R$ to contain a model $M$ of $S$ whose variety equals or exceeds $S$‘s variety; the cycle-mismatched regulator’s model does not satisfy this condition.

The two conditions together specify that $R$ cannot, by construction, stabilise $S$. The variety $S$ presents must be absorbed somewhere; the absorption cannot be by $R$; the absorption must be within $S$ itself, producing system-state changes $R$ does not register and does not address. This concludes the proof. $\blacksquare$

2.2.3 Formal Restatement#

The proof above is formal. The same content can be displayed as a chain of inequalities. The notation is fixed once: $V_R$ denotes regulator variety; $V_S$ denotes system variety; $V_M$ denotes model variety; $V_D$ denotes disturbance variety; $T_R$ and $T_\Sigma$ are the cycles; $N(T)$ denotes the number of distinguishable signals available at temporal resolution $T$.

The three prior results that the section composes are, in displayed form:

The third result is the temporal-resolution bound Lemma 1 establishes; it is the variety form of the zero-rate allocation that rate-distortion theory assigns, under any fixed rate budget, to components whose within-window variance falls below the water level (Shannon, 1959; Cover and Thomas, 2006), with the sampling-theorem corollary (Shannon, 1949) as its intuitive register, as set out in §2.2.2.

The cycle-mismatch hypothesis is $T_R \ll T_\Sigma$, with $V_S$ concentrated at scale $T_\Sigma$. The chain inference is:

The first inequality is Conant-Ashby; the second is the variety-cycle bound of Lemma 1; the third follows from the cycle-mismatch hypothesis, since variety concentrated at scales much greater than $T_R$ is, by that bound, not in $N(T_R)$. The composed inequality contradicts the Conant-Ashby requirement $V_M \geq V_S$. The regulator cannot stabilise the system. The system’s variety must be absorbed within $S$ itself. $\blacksquare$

2.2.4 The Mismatch Ratio#

The ratio $T_\Sigma / T_R$ specifies how far into the theorem’s antecedent a given regulator-system pair sits. The matched case is the case where the ratio is approximately one and the theorem’s antecedent does not apply. The mismatched case is the case where the ratio is large enough that disturbances at the system’s scale appear within the regulator’s cycle as approximately constant inputs.

The transition between the two cases is not a sharp boundary. A regulator can track a system whose variety moves at a rate close to its own action-cycle, and the tracking degrades as the ratio grows. Up to a ratio of perhaps ten or twenty, the regulator can still resolve the system’s slow movement within a tolerable number of cycles. Between that band and ratios of about $10^3$, tracking degrades without vanishing: the regulator resolves the substrate’s movement only across many action-cycles, registering it late and coarsely, so that intervention lags the degradation by intervals long enough for the cannibalisation interval of Part III to open. Beyond ratios of about $10^3$, no sampling strategy lets the regulator resolve the system’s variety at all, and the degradation proceeds entirely between the regulator’s distinguishable states. The paper’s cases inhabit the middle band: the Boeing pair stands near $10^2$ and the Wells Fargo pair near $4 \times 10^1$, which satisfies the theorem’s antecedent, a mismatch of more than a factor of order one, and places both regulators in the regime of degraded, lagged tracking rather than total aliasing. The theorem’s consequence requires no more than that.

The vertical reading of Conway’s Law that Part I named applies the theorem to the layered structure of organisational communication. The management layer operates at one horizon, with its own action-cycle: quarterly earnings reviews, annual budget cycles, share-price movements measured in days (Galbraith, 1967, Chapter VI on the technostructure’s institutional position; Lazonick, 2014, on the institutional foundations of executive compensation cycles in the contemporary financialised corporation). The working layer operates at a different horizon: the development time of a product, the apprenticeship cycle of a craft, the certification cycle of an engineering practice. The two horizons stand in a ratio whose empirical range, in the cases the paper engages with, falls typically between $10^2$ and $10^4$, which is sufficient to place the regulator-substrate pair in the regime of §2.2.2 by an order of magnitude or more. Under this regime the theorem’s antecedent is met and its consequent follows by construction. The management regulator cannot model the working substrate at its own cycle. The variety of the substrate (the range of states the engineering coherence of the team can occupy, the configurations of mutual model and shared tacit knowledge the team can hold, the calibrations to the product’s behaviour each engineer can carry) must be absorbed within the engineering layer itself, because the management layer cannot register the variety at the cycle on which it varies. The disintegration the homomorphism transmits to the artefact proceeds invisibly to the regulator’s instruments. It registers in the regulator’s frame only when the engineering layer’s coherence has degraded across thresholds that produce effects at the regulator’s resolution, at which point the degradation is no longer recoverable on the timescale at which the recovery would have had to occur.

This is the vertical horizon mismatch Part I named in its informal version. The theorem is its formal statement. Conway’s disintegration thesis (Conway, 1968) is the empirical observation; the horizon-mismatch theorem is its cybernetic mechanism.

2.2.5 The Architecture’s Necessary Signature#

The horizon-mismatch theorem specifies that depletion proceeds as the necessary signature of cycle-mismatched regulation. The empirical signature has a specific form the theorem permits the paper to state in advance of any particular case. The signature is necessary in the formal sense: it is what the theorem predicts must occur in any cycle-mismatched order, with the specific empirical features depending on the order’s particular features but the signature’s general form being the same across orders.

The signature has three features that together specify what the empirical record will exhibit in any cycle-mismatched order.

The first feature is that depletion proceeds invisibly at the system’s cycle while the regulator’s metrics improve at the action-cycle. This is the direct consequence of the cycle-mismatch: the regulator’s metrics, operating at $T_R$, register the order’s state at the regulator’s resolution; the system’s depletion, operating at $T_\Sigma$, is invisible at the regulator’s resolution; the depletion does not appear in the regulator’s metrics until the system’s state crosses thresholds that produce effects at the regulator’s resolution. The order therefore exhibits, across the depletion’s accumulation period, an empirical pattern in which the regulator’s metrics trend favourably while the system’s state deteriorates. The depletion is registered, when it is registered at all, as a separate question outside the regulator’s framework: as a “social” question, an “ecological” question, a “quality” question, a “long-term” question, with the bracketing being the signature of the regulator’s inability by construction to model the system at the system’s cycle.

The second feature is that the depletion becomes visible to the regulator only through perturbation that crosses the measurement system’s threshold. This follows from the first: the regulator’s metrics do not register the depletion until the system’s state crosses thresholds producing effects at the regulator’s resolution. The thresholds are typically crossed when accumulated depletion produces a discrete event: an engineering-culture depletion produces a product failure, a customer-trust depletion produces a regulatory enforcement action, a supplier-ecosystem depletion produces a supply-chain disruption. The discrete events are what the order’s discourse registers as the depletion’s empirical signature; the events are the surface of a pattern the regulator’s metrics did not register and could not have registered, by the formal reasons the theorem specifies.

The third feature is that the order’s response to the discrete events is constrained by construction by the regulator’s framework. The response is what the framework permits; the framework permits the response to be expressed in the framework’s own vocabulary, which is the vocabulary at the regulator’s resolution; the response cannot, by construction, address the system’s state at the system’s own cycle. The response is therefore registered as a regulatory adjustment within the framework, with the adjustment’s content being whatever the framework’s vocabulary admits.

The signature is not a list of contingent failures better regulation could prevent. The signature is the architecture’s necessary signature in the formal sense: it is what any cycle-mismatched order must exhibit. The empirical cases Part IV develops exhibit it because the cases instantiate the antecedent the theorem specifies. The cases are not anecdotes selected for rhetorical force; they are the empirical signature of an architecture whose formal features the theorem specifies.

2.3 The Polanyi Extension#

The horizon-mismatch theorem is the formal statement of why vertical regulation fails when the layers operate at incommensurable cycles. The Polanyi extension addresses a related question: what specifically is the content that fails to survive vertical translation, even when the cycles are not catastrophically mismatched?

The answer draws on Michael Polanyi’s work on tacit knowledge, principally Personal Knowledge (Polanyi, 1958, originally published by Routledge & Kegan Paul, based on the Gifford Lectures of 1951-1952) and The Tacit Dimension (Polanyi, 1966, originally published by Doubleday, based on the Terry Lectures of 1962). Polanyi’s central claim, against the dominant epistemology of his time, is given in the opening sentence of the first chapter of The Tacit Dimension: “I shall reconsider human knowledge by starting from the fact that we can know more than we can tell” (Polanyi, 1966, p. 4, emphasis in original). The phrasing with “can” is Polanyi’s own and bears the weight of the claim: he is not saying we know more than we do tell, which would be contingent, but that we know more than we are able to tell, which marks the knowledge as tacit in the way the rest of the chapter unpacks.

Polanyi’s deeper claim is that all knowledge has a tacit dimension. Even explicit knowledge rests on tacit understanding that cannot itself be made fully explicit. The formula $F = ma$ is explicit; understanding what it means requires tacit knowledge of what force is, what mass is, what acceleration is, and how the symbols connect to the phenomena they describe (Polanyi, 1958, Part III on tacit knowing as the foundation of scientific knowledge; Polanyi, 1966, Chapter 1 on “Tacit Knowing”). The explicit always sits on top of the tacit. The explicit is the visible part of a body of knowledge whose mass is below the surface.

Polanyi develops the claim through several examples that have become canonical in the literature on tacit knowledge (Shapin, 2011, on the canonical examples and their reception). The bicycle example: we can ride the bicycle without being able to articulate the rules of riding it (Polanyi, 1958, p. 49). The rules we can articulate (“lean into the turn,” “keep moving forward,” “look where you want to go”) are post-hoc reconstructions. The actual coordination by which the body keeps the bicycle upright involves continuous corrections at a temporal resolution faster than conscious articulation can reach. Reading the rules does not enable the riding. The riding is learned by riding, with the falls and the corrections that doing so produces. The book about riding is not the riding.

The face-recognition example: we can recognise a familiar face in a crowd without being able to specify what it is about the face we recognise (Polanyi, 1966, p. 4). The recognition is immediate and reliable; the articulation is post-hoc and approximate. Polanyi’s claim is that recognition operates on the integration of subsidiary particulars into a focal awareness that cannot be decomposed into the particulars without losing the recognition itself.

The apprenticeship example: the reason apprenticeship existed for centuries as the dominant form of skill transmission is that the knowledge being transmitted could not be put into a book. “An art which cannot be specified in detail cannot be transmitted by prescription, since no prescription for it exists. It can be passed on only by example from master to apprentice” (Polanyi, 1958, p. 53). The apprentice learns by watching, doing under supervision, being corrected when wrong, absorbing through proximity what could not be transmitted through instruction. This is not a primitive substitute for something better. This is the only form that worked for the kind of knowledge being transmitted, because the knowledge could not exist in any other form.

The same applies to organisations layer by layer. The vertical layers of an organisation are connected by communication structures that carry vocabulary, plans, metrics, instructions. The vocabulary, plans, metrics, and instructions are the explicit content. What carries the actual coordination between the layers, what makes the executive’s strategic intent become the team’s executed work without the translation losses Part I named, is the tacit dimension that operates underneath the explicit content. The tacit dimension is the mutual model of how the layers relate, the shared sense of what the words mean in this organisation, the practices that have accumulated over time and that members of the organisation know how to use without being able to articulate.

When the organisation’s vertical layers are working, the tacit dimension is intact. The same word means the same thing all the way through. The plan at the management layer is the plan being executed at the working layer. The metric reports something that has a known relationship to the work it is a proxy of. The communication structure is one structure because the tacit dimension is connecting the layers underneath the explicit content.

When the tacit dimension fails, the explicit content continues to be transmitted with no signal that anything is wrong. The vocabulary still moves across the layers. The plans still arrive at the working layer. The metrics still rise on the dashboard. But the vocabulary now means different things at different layers; the plans are not the plans being executed; the metrics are measuring something that has become detached from the working reality they were a proxy of. The communication structure is no longer one structure. The layers have come apart, and the explicit content does not register the coming-apart because the explicit content was never carrying the coherence in the first place.

Conway’s incentive-subverts-intent claim (Conway, 1968, opening section) names the surface mechanism: the manager’s incentive structure motivates choices that subvert the sponsor’s intent. What Conway does not specify is the mechanism by which the subversion occurs. The Polanyi extension supplies the mechanism: the subversion occurs because the tacit dimension that connected the layers fails to survive the translation when the incentives at one layer reward optimisation against metrics the next layer down knows are not the work. Conway’s text gestures at the Polanyi point in his note about the manager’s incentive subverting design choices, but does not name the tacit dimension as such. The present paper supplies the name.

The framework’s silence on the tacit dimension is the silence the genre cannot break. A framework is, by its form, explicit content. It is a book. It is a chart. It is a vocabulary of team types and interaction modes that can be drawn on a slide. The framework can prescribe the explicit content perfectly and have nothing to say about the tacit dimension that connects the layers underneath. What undoes the framework is not that practitioners apply it badly; what undoes it is that the form of the framework, by being a framework, cannot reach what makes the framework work. Polanyi himself anticipated the application: “rules of art can be useful, but they do not determine the practice of an art” (Polanyi, 1958, p. 50).

2.4 The Three-Channel Extraction Structure#

The horizon-mismatch theorem establishes that depletion is the necessary signature of cycle-mismatched regulation. The Polanyi extension establishes that what depletes first is the tacit dimension. The third diagnostic component names the channels along which the depletion actually occurs.

When a regulator’s metrics rise while the productive substrate the metrics are a proxy of depletes, the rising metric is being kept favourable by extraction from that substrate. Extraction is not a metaphor here; the term names a definite relationship between the visible flow the regulator measures and the invisible stock the regulator does not measure. The metric is the flow. The substrate is the stock. The metric is kept high by spending the stock.

The mechanism is what the literature on metric-driven management calls surrogation: the metric becomes the proxy for the underlying value the organisation actually depends on, then the proxy becomes the target, then the target is achieved by spending the underlying value the proxy was a proxy of (Choi, Hecht, and Tayler, 2012, on the formal definition of surrogation in management accounting; Schaub, 2021, on the empirical regularities of surrogation across firms). The metric continues to register success while the substrate is being consumed to produce the success the metric registers.

The literature that situates surrogation in the broader pattern of metric fixation includes Goodhart’s law, originally formulated by the British economist Charles Goodhart in a paper presented at the Money Study Group of the London School of Economics in 1975 (Goodhart, 1975/1981, “Problems of Monetary Management: The U.K. Experience,” in Papers in Monetary Economics, Reserve Bank of Australia, reprinted in Goodhart, 1984). Goodhart’s original formulation was: “Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes” (Goodhart, 1975, the central thesis). The more famous formulation, “When a measure becomes a target, it ceases to be a good measure,” is from the anthropologist Marilyn Strathern (Strathern, 1997, “Improving Ratings: Audit in the British University System,” European Review 5(3): 305-321). The broader literature on metric fixation documents the pattern across hospitals, policing, education, manufacturing, and other sectors (Muller, 2018, The Tyranny of Metrics, Princeton University Press; on the same pattern in education specifically see Campbell, 1976, “Assessing the Impact of Planned Social Change,” and Campbell’s Law as the predecessor of Goodhart’s).

The empirical record of surrogation and metric fixation, read across the cases the literature documents, shows that the extraction operates along three distinguishable channels. The channels are not exclusive; in many cases the extraction operates along more than one channel simultaneously. But the channels are distinguishable, and naming them lets the paper specify what the depletion’s signature actually looks like in different cases. The Conway connection is to the accounting-fallacy critique: where Conway identified conventional accounting theory as the underlying condition that drives the disintegration (Conway, 1968, “System Management,” on the linearity assumption that two men for a year and one hundred men for a week are resources of equal value), the three-channel structure specifies what the accounting fallacy actually depletes when it operates over time.

The first channel: extraction from the worker. The substrate the metric is depleting is the human capital of the workforce. The depletion takes specific forms documented across the political-economy literature: senior staff with deep tacit knowledge are pushed out or retired without replacement; staff who carry the apprenticeship transmission for the next generation are removed from positions where the transmission would occur; the cognitive load of the work increases beyond what the workforce can sustain; the conditions under which tacit knowledge accumulates are dismantled (Lazonick, Tulum, and Sakinc, 2020, on the buyback-driven hollowing of corporate engineering capacity; Foroohar, 2016, Makers and Takers, on the financial-architecture’s career-structure dynamics; for the engineering-specific pattern see Robison, 2021, Flying Blind, on Boeing). The metric the regulator is tracking, typically a productivity or efficiency or cost metric, rises as the worker substrate depletes. The senior engineer who leaves takes thirty years of tacit certification knowledge with her; the metric does not register the loss because the metric does not measure thirty years of tacit certification knowledge.

The second channel: extraction from the product. The substrate the metric is depleting is the integrity of the product the organisation produces. The depletion takes specific forms: safety margins are eroded; reliability standards are relaxed; design redundancies are removed; the product’s quality variance widens; failure modes that would have been caught are now passed through (Joint Authorities Technical Review, 2019, on the certification-process documentation; House Committee on Transportation and Infrastructure, 2020, on the production-pressure culture; for the broader pattern in financialised manufacturing see Gelles, 2022, The Man Who Broke Capitalism). The metric the regulator is tracking, typically a margin or unit-cost or time-to-market metric, rises as the product substrate depletes. The aircraft that ships with a flight-critical control system reading a single angle-of-attack sensor at a time, without the cross-check between the two installed sensors that a function of its authority warranted (House Committee, 2020, especially Chapter 5; Joint Authorities Technical Review, 2019; Robison, 2021) is the product substrate depleting; the cost metric does not register the depletion because the cost metric does not measure the cross-check logic that was not required.

The third channel: extraction from the customer. The substrate the metric is depleting is the trust of the customer base. The depletion takes specific forms: customers are charged for products they did not authorise; commitments are made to customers that the organisation does not intend to honour; the customer relationship is mined for short-term revenue at the cost of the long-term relationship; the customer’s interest is subordinated to the metric the organisation is being measured against (U.S. Securities and Exchange Commission, 2020, on the Wells Fargo settlement; Stanford Graduate School of Business, 2019, on the Wells Fargo cross-selling case; for the broader pattern in platform economies see the documentary record on Amazon’s use of third-party seller data, the U.S. House Judiciary Subcommittee on Antitrust, 2020 report, pp. 247-318, on the conflict-of-interest built into platform self-preferencing). The metric the regulator is tracking, typically a revenue or cross-sell or engagement metric, rises as the customer substrate depletes. The bank that opens accounts in customers’ names without their consent (U.S. SEC, 2020; the Stanford GSB 2019 case study on the institutional record) is extracting from the customer-trust substrate; the cross-sell metric does not register the depletion because the cross-sell metric does not measure trust.

The three channels exhaust the empirical pattern the cases exhibit in the literature on metric fixation and surrogation. Every case the present paper examines extracts along one or more of these channels. The cases differ in which channel is primary: Boeing extracts primarily along the worker and product channels, with the customer channel involved only at the moment the discrete event (the Lion Air and Ethiopian Airlines crashes) crosses the threshold into the regulator’s measurement frame. Wells Fargo extracts primarily along the customer channel, with the worker and product channels involved as secondary effects: the workers were pressed into the extractive practice (the documented sales-pressure culture at the bank; Stanford GSB, 2019, on the institutional record), and the product was the cross-sell that was being mis-sold. Boeing and Wells Fargo together exhibit the same diagnostic mechanism with different extraction profiles, which is what makes them the right pair of cases to anchor the paper’s empirical argument.

The three-channel structure is not the paper’s invention. The empirical record across many cases of organisational failure under metric-driven management has shown the same pattern; the channels are the categories the empirical record produces when the cannibalisation mechanism is read across cases (Muller, 2018, on the cross-sectoral pattern; for the deeper analysis in the financialised corporation, the Lazonick et al. 2020 institutional record).

2.5 The Constituent Cannot Total the Process#

A further diagnostic component closes Part II. It is needed for what Part III’s formal core will establish.

The cybernetic results of §2.1 and §2.2 treat the regulator as if it could be analytically separated from the system. The regulator is on one side; the system is on the other; the question is whether the regulator’s variety and model are adequate to the system’s variety. This first-order analytical position is sufficient for the proofs the paper requires. It is also, in an important sense, an idealisation. The point has been developed in the second-order cybernetic tradition by von Foerster and others (von Foerster, 2003, Understanding Understanding: Essays on Cybernetics and Cognition; for the constructive epistemology that follows see Glasersfeld, 1995, Radical Constructivism).

The fuller analytical position recognises that the regulator is itself a constituent of the system it claims to regulate. The management apparatus of a firm is not external to the firm; the management apparatus is part of the firm, with its own existence depending on the firm’s continued operation, with its own incentives shaped by the firm’s structure, with its own information channels constructed from inside the firm rather than being neutral observation channels from outside (the institutional analysis in Galbraith, 1967, on the technostructure’s institutional position is the empirical complement to the second-order cybernetic point). The framework that a consultancy sells is not external to the organisations that adopt it; the framework’s authors have their own positions within the management apparatus of contemporary organisational life, and the framework’s adoption is itself an event in the organisational dynamics the framework claims to describe.

The constituent cannot total the process. This is the second-order point that the first-order cybernetic results do not address but that the diagnostic must hold. Any regulator that is a constituent of the system it regulates is, by virtue of being a constituent, unable to occupy the position of an external observer that the first-order Ashby and Conant-Ashby results presume. The constituent’s variety is itself an output of the system; the constituent’s model is itself a feature of the system; the regulator’s actions feed back into the system in ways the regulator does not, by being inside the system, fully register (von Foerster, 2003, on the observer-as-participant; Maturana and Varela, 1980, Autopoiesis and Cognition, on the constitutive role of the observer in the system being observed).

The implication for the paper’s argument follows at once. The Conant-Ashby requirement that the regulator contain a model of the system is, at the second-order register, a requirement that the constituent contain a model of the process it constitutes, which is not available to any regulator that is itself part of the process. The first-order proofs of §2.1 and §2.2 stand, but they specify what is impossible by construction even before the second-order point applies. The framework’s inability to satisfy the first-order conditions is sufficient for the paper’s diagnostic; the second-order point further establishes that no framework operating as a constituent of the management apparatus could satisfy the conditions even in principle.

This closes Part II. The next part develops the formal core that demonstrates, with the relational mathematics of the Harsanyi dividend (Harsanyi, 1959, 1963), that the topography the regulator cannot model is also the topography no per-role or per-edge instrument can carry.

Part III: The Formal Core#

Part II developed the cybernetic apparatus and the Polanyi extension. Part III now demonstrates, in formal terms, the central claim that connects them: that the value of an organisation is not the sum of the values of its members or its edges; that the difference is what topography names; that any per-role or per-edge representation is blind to that difference, of either sign; and that a reallocation driven by such a representation, where it dissolves positive coalitions faster than they re-form, consumes that difference as it acts and can produce its reported gains by that consumption. The blindness result is established twice, by two independent formalisms, so that a reader who rejects one must still answer the other; the consumption result is conditional and its conditions are stated where it is proved. The register is that of demonstration. Where a step is articulation rather than demonstration it is marked as such.

The connection to Conway runs through the accounting-fallacy critique of §1.3.4. Conway identified conventional accounting theory as the underlying institutional condition that drives the disintegration of design organisations (Conway, 1968, “System Management”). What Conway named informally, the formal core now demonstrates with full mathematical machinery: per-role accounting is blind to the relational surplus that produces organisational performance, and acting on per-role accounting destroys what it cannot see.

3.1 Setup and Definitions#

Let $N = {1, 2, \ldots, n}$ be the members of an organisation, or, equivalently for what follows, the vertices of its design organisation graph in Conway’s sense (Conway, 1968). Let $2^N$ denote the set of all subsets (coalitions, sub-organisations) of $N$. Organisational performance is represented by a value functional

where $v(S)$ is the value the sub-organisation $S$ can realise through its members working in their actual coupled relation, not in isolation. This is the standard characteristic-function representation of a cooperative system (von Neumann and Morgenstern, 1944, Theory of Games and Economic Behavior, Princeton University Press; Shapley, 1953, “A Value for n-Person Games,” in Kuhn and Tucker, eds., Contributions to the Theory of Games II, Princeton University Press, pp. 307-317). No assumption is made yet about how $v$ decomposes; that is the question.

Definition 1 (Per-role valuation). A valuation is per-role (equivalently, additive or separable) if there exist individual values $\phi_1, \ldots, \phi_n \in \mathbb{R}$ such that for every coalition $S$,

A per-role valuation is the formal content of the demand that each role show its value, and equally the formal content of the per-edge accounting on which dependency-cost dashboards rest. It presumes that organisational value can be partitioned into per-person or per-edge contributions that sum to the whole. It is the formal expression of what Conway (1968, “System Management”) identified as the conventional accounting theory’s central assumption: “the unit of resource is the dollar, and all resources must be measured using units of measurement which are convertible to the dollar. If the resource is human effort, the unit of measurement is the number of hours worked by each man times his hourly cost, summed up for the whole working force.” Conway named the assumption as a fallacy. The present section gives the fallacy its formal proof.

Definition 2 (Superadditivity and strict synergy). The value functional is superadditive if for all disjoint $S, T \subseteq N$,

and exhibits strict synergy on $(S, T)$ if the inequality is strict (Shapley, 1953). Strict synergy says the joined sub-organisations realise more together than the sum of what they realise apart: the surplus resides in the coupling created by joining, not in either part.

The operating image of strict synergy that the paper takes as canonical, because it bypasses the misreading by quantity that the standard “two heads are better than one” formulation invites, is lifting an object with two hands. The standard formulation suggests the second hand adds more of what the first hand already does: with one hand a person can carry a certain mass, with two hands somewhat more. A critic of strict synergy can answer at the same register that two hands often carry less than twice what one hand carries, that the second hand is sometimes idle or interferes with the first, and so the synergy claim is contingent and small. The two-hand image displaces that exchange entirely. With one hand a person can pick up a box that has a handle or that can be gripped from above. A box that has no handle and is wider than the span of one hand cannot be lifted with one hand at all, irrespective of how strong the hand is; the configuration does not include a grip on the object. With two hands the same box is lifted routinely, because the second hand permits a configuration the first hand alone cannot form. The capability added by the second hand is not a larger amount of the capability the first hand had. It is a capability the one-hand configuration does not contain at any strength. Translated into the formalism: the dividend on the singleton “left hand” is what the left hand alone can lift, the dividend on the singleton “right hand” is what the right hand alone can lift, and the dividend on the coalition $\{\text{left hand}, \text{right hand}\}$ is a quantity that includes the boxes neither hand can lift alone and that exists only in the joint configuration. A per-hand ledger sums the singletons and reports zero for the handleless box, which the pair lifts every day. Strict synergy is the formal name for this kind of capability, in which the higher-order dividend is not an increment to a quantity each member already carries but a quantity that exists only in their being joined.

The two-hand image carries the spatial form of strict synergy: the configuration at a moment that the singletons cannot form. Two further images are needed, because the higher-order dividend escapes a per-role ledger in two distinct further ways, and the two ways do different work in the argument that follows.

The first is the symphony orchestra, which carries the case of fungible-accounting failure at the moment of production. The case the image stages is the case Conway named in 1968 when he wrote that two men working for a year and one hundred men working for a week, at the same hourly cost per man, are treated by conventional accounting as resources of equal value. The orchestra is the case that refutes the equivalence in its sharpest form. A performance of Beethoven’s Ninth by a hundred-piece orchestra and choir on a single evening and the same hundred parts played by one violinist across the equivalent person-hours stand to the Ninth itself in incommensurable relations. The per-musician-hour count comes out close. The thing produced is categorically different: in the first case the Ninth, in the second a long sequence of individual notes that the Ninth is not. What makes the Ninth the Ninth is the simultaneity of the parts, the violins answering the woodwinds in the same instant, the cellos sustaining the bass while the brass moves above, the chorus entering over the orchestra at the moment Beethoven specified. Take the simultaneity away and there is no slower symphony; there is a different artefact, one in which the music does not occur, because the music is constituted by the joint configuration at a moment of performance. A per-musician ledger that counted what each musician played and summed the counts would record the same fungible total in the two cases, and the categorical difference between symphony and sequence would not appear anywhere in the ledger. The dividend on the orchestra coalition contains the Ninth itself; the per-musician sum contains only the notes. The fungible accounting Conway identified as a fallacy is the accounting the per-musician ledger conducts; the orchestra is the empirical case in which the fallacy’s consequence is most starkly visible.

The second image is two developers working together for a year, which carries the case the cannibalisation theorem of §3.5 will operate on. Two developers at the same desk over twelve months are not producing only the code their commit logs and bug-tracker tickets record. They are also producing, in the same hours, the substrate that does not appear in the commits: the shared mental model of how the codebase actually behaves, the mutual knowledge of where it tends to fail and what compensates, the developed sense of how each other thinks under deadline, the tacit map of the user’s real workflow, the calibration to each other’s blind spots that lets one of them spot what the other will miss. The substrate is what Polanyi (1958, 1966) calls the tacit dimension of skill, accrued in the form Polanyi specifies as transmissible only by example from master to apprentice or, in the case of peers, accrued by sustained shared work. The two configurations Conway’s accounting equates the substrate-accrual case to are again the same in fungible terms and different in substantive terms: two developers for a year and a hundred developers for a week produce the same person-hour total, and the second configuration cannot produce that shared mental model because it is exactly what a year of sustained shared work accumulates and a week does not. The dividend on the two-developer coalition after a year is the year of substrate-building; the dividend on any of the hundred-developer coalitions after a week is, at most, the week’s substrate, which has not had time to accrue. A per-role ledger reading the commits, the velocity, the bug-fix rate over either interval produces a number; the substrate behind the number, which is what makes the next year’s work possible, is not in the number anywhere.

The three images together mark the three ways the higher-order Harsanyi dividend escapes any per-role ledger: by being a capability that requires joint configuration at a moment (the two-hand case), by being an artefact that requires joint simultaneity of contribution (the orchestra case, refuting fungible accounting at the moment of production), and by being a relational substrate that accrues only across sustained shared work and cannot be reconstructed from any person-hour-equivalent configuration (the two-developers-for-a-year case, on which the cannibalisation theorem of §3.5 directly depends).

Definition 3 (The synergy functional). Define the synergistic surplus of the whole organisation as

$\Sigma(v)$ is the aggregate of the value present in the organisation that is not present in any member taken singly. It is one summary of what topography names, the part of organisational capacity that exists only in the relations among people; the fuller object, developed in §3.2, is the structure of higher-order dividends of which $\Sigma(v)$ is the sum, since that structure carries information the sum alone does not. Conway’s homomorphism (§1.2.2 above) connects this directly to the system being designed: the system graph is a homomorphic image of the design organisation graph, so the relational structure of the design organisation is transmitted to the system as the system’s own coherence.

3.2 The Locus of Topography: Harsanyi Dividends#

Any characteristic-function value admits a unique decomposition into coalition-level contributions, the Harsanyi dividends. The concept was introduced by Harsanyi in 1959 (Harsanyi, 1959, “A bargaining model for cooperative n-person games,” in Tucker and Luce, eds., Contributions to the Theory of Games IV, Princeton University Press, Annals of Mathematics Studies No. 40, pp. 325-355) and developed further in 1963 (Harsanyi, 1963, “A simplified bargaining model for the n-person cooperative game,” International Economic Review 4(2): 194-220). For each non-empty coalition $T \subseteq N$ there is a unique dividend $d_v(T)$ such that

The dividend $d_v(T)$ is the value generated by the coalition $T$ that is not already accounted for by any of its proper sub-coalitions. The singleton dividends $d_v({i}) = v({i})$ are the purely individual contributions. Every dividend $d_v(T)$ with $|T| \geq 2$ is, by construction, a relational quantity: it is the value that appears only when those specific members are joined, attributable to no member alone (Hart and Mas-Colell, 1989, on the dividend decomposition; Dehez, 2017, on the broader axiomatic structure of Harsanyi values).

Lemma 3 (Topography is the higher-order dividend mass). The synergistic surplus equals the total mass of all dividends on coalitions of size at least two:

Proof. By the Harsanyi identity applied to $S = N$, $v(N) = \sum_{\substack{T \subseteq N \ T \neq \varnothing}} d_v(T) = \sum_{i \in N} d_v({i}) + \sum_{\substack{T \subseteq N \ |T| \geq 2}} d_v(T) = \sum_{i \in N} v({i}) + \sum_{\substack{T \subseteq N \ |T| \geq 2}} d_v(T).$ Subtracting $\sum_{i} v({i})$ from both sides gives the claim. $\blacksquare$

Lemma 3 locates topography in the formal apparatus. It is the aggregate of every relational dividend in the organisation: every pairing whose joint output exceeds the sum of its parts, every triad whose mutual knowledge exceeds what its pairs possess, and so on up to the whole. These dividends are components of $v(N)$ in the same sense as the singleton terms, and they are exactly the components indexed by relations rather than by individuals. In Polanyi’s terms (§2.3 above), the higher-order dividends are where the tacit dimension of organisational knowledge lives: the dividends on coalitions of size two and above are not articulable as the sum of articulable individual contributions, because the coupling that produces them is itself tacit.

3.3 The Invisibility Theorem#

Theorem 1 (Per-role valuation is blind to topography). A per-role valuation in the sense of Definition 1 represents the value functional faithfully if and only if every higher-order Harsanyi dividend vanishes, that is, $d_v(T) = 0$ for every coalition $T$ with $|T| \geq 2$. Consequently, whenever an organisation has any non-zero higher-order dividend, no per-role valuation represents it, and the relational mass carried by those dividends is assigned, under any per-role valuation, to no member and is therefore invisible to it.

Proof. Suppose a per-role valuation represents $v$, so $v(S) = \sum_{i \in S} \phi_i$ for all $S$. Then for any disjoint $S, T$, $v(S \cup T) = \sum_{i \in S \cup T} \phi_i = \sum_{i \in S}\phi_i + \sum_{i \in T}\phi_i = v(S) + v(T),$ so $v$ is exactly additive, and by the uniqueness of the Harsanyi decomposition every higher-order dividend is zero. Conversely, if $d_v(T) = 0$ for every $T$ with $|T| \geq 2$, then the Harsanyi identity gives $v(S) = \sum_{i \in S} d_v({i}) = \sum_{i \in S} v({i})$ for every $S$, so the valuation $\phi_i = v({i})$ represents $v$. Representability is therefore equivalent to the vanishing of every higher-order dividend, which is strictly stronger than $\Sigma(v) = 0$: the surplus $\Sigma(v)$ is the sum of the higher-order dividends, and a sum of zero is consistent with individual dividends that are non-zero and cancel. For the final clause: any per-role valuation distributes only the singleton mass $\sum_i \phi_i = \sum_i v({i})$; every dividend $d_v(T)$ with $|T| \geq 2$ is, by Definition 1, expressible in no $\phi_i$, and so is assigned to no member and does not appear in the valuation’s ledger. $\blacksquare$

The result the theorem establishes is the formal heart of the matter, and it is a claim about blindness rather than about the sign of what is unseen. The demand that each role demonstrate its value in a per-role metric, and equivalently the demand that each dependency edge be priced into a coordination cost, is satisfiable only for an organisation in which every higher-order dividend is zero, an organisation with no relational structure of any kind above the level of the individual. For any organisation in which coalitions carry non-zero dividends, the metric does not undervalue the relational structure; it cannot represent it at all, and reports it as nothing. This holds whether the dividends are positive (a pairing whose joint output exceeds the sum of its parts) or negative (a coalition carrying drag, a dependency that damages the whole). The per-role ledger is blind to relational structure in both directions: it cannot see the coupling that produces value, and it cannot see the coupling that destroys it. What the per-role instrument lacks is not a correction term of known sign but the relational stratum itself.

This is the formal proof of Conway’s accounting-fallacy critique. Conway claimed that “two men working for a year or one hundred men working for a week (at the same hourly cost per man) are resources of equal value” is a fallacy. Theorem 1 specifies why Conway was right: the two configurations differ in their higher-order dividends, and any accounting that operates per-role (per-hour, per-dollar, per-head) cannot see the difference. The proof is in the impossibility of representing any non-zero higher-order dividend in a per-role ledger. The orchestra image of §3.4 carries Conway’s fallacy at the moment of production: a hundred musicians for an evening and one musician across the equivalent hours sum to the same per-role total and produce categorically different artefacts, because the symphony is in the higher-order dividend the per-role ledger cannot represent. The two-developers-for-a-year image of §3.4 carries the further sense in which the fallacy operates: the equivalence Conway named is false not only at the moment of production but at the layer below it, where the substrate the production runs on is accruing on cycles a per-role ledger does not represent at all. The first sense grounds Theorem 1; the second sense grounds Theorem 3 of §3.5, the cannibalisation result.

3.4 The Information-Theoretic Confirmation#

The same result follows from information theory, by an independent route, which is why it cannot be dismissed as an artefact of the value-functional formalism.

Model the organisation’s capacity to act adequately on its system as the mutual information that the members’ states, viewed as sources $X_1, \ldots, X_n$, carry about the system’s required response $Y$. The partial information decomposition of Williams and Beer (Williams and Beer, 2010, “Nonnegative Decomposition of Multivariate Information,” arXiv:1004.2515) resolves the joint mutual information $I(X_1, \ldots, X_n ; Y)$ into non-negative atoms: the unique information each member carries alone, the redundant information shared across members, and the synergistic information $\mathrm{Syn}$ that is present only in the members jointly and is recoverable from no member or proper subset taken alone.

Proposition 1 (Synergy vanishes under per-member marginalisation). The synergistic atom $\mathrm{Syn}(X_1, \ldots, X_n ; Y)$ is not a function of any marginal distribution $p(X_i, Y)$, nor of any collection of them short of the joint; it is destroyed by marginalising the organisation to its members. (This follows directly from the Williams-Beer construction, in which synergy is defined as the information in the joint source that exceeds what the redundancy lattice supplies from subsets; it is articulation of their result, not a new proof.)

The correspondence between the value formalism and the information register is homologous at the relevant point. The Conant-Ashby theorem (Conant and Ashby, 1970) requires that an adequate regulator be a model of its system; the organisation’s adequacy is therefore carried by $I(X_1, \ldots, X_n ; Y)$, and the synergistic atom of that information is the regulatory capacity that exists only in the coupling. A per-member assessment reads each $p(X_i, Y)$ and so captures unique and (partially) redundant information while discarding $\mathrm{Syn}$ entirely. Topography, in the information register, is the synergistic atom: the organisation’s modelling capacity that lives only in the joint configuration of its people, and that a per-person measurement demonstrably cannot see.

Two formalisms, one conclusion: the relational surplus is where topography lives, and per-role measurement is blind to it.

3.5 Measurement-Driven Reallocation Cannot Protect What It Cannot See#

Invisibility would be only an accounting defect if the metric were inert. It is not; it drives reallocation. Per-role and per-edge measurements drive decisions: where to assign people, where to cut, where to redraw team boundaries, where to prune dependencies. We now show that such measurement-driven action has no term that protects relational value of either sign, and that under the conditions in which existing coalitions carry positive dividends it destroys that value while remaining invisible to the instrument at the moment of decision.

Definition 4 (Measurement-driven reallocation). A reallocation is measurement-driven if its decisions are a function only of the per-role or per-edge ledger ${\phi_i}$, that is, only of quantities visible to a per-role valuation.

Theorem 2 (Measurement-driven reallocation cannot protect topography, and destroys it under the stated conditions). Let an organisation carry positive relational dividends on a family of coalitions, and let a measurement-driven reallocation act on it: removing members whose $\phi_i$ falls below a threshold, redistributing members across newly drawn boundaries, or pruning edges to optimise the visible ledger. Then the reallocation’s objective contains no term that preserves or optimises any higher-order dividend; it has no built-in tendency to protect relational value of either sign. Where the reallocation dissolves coalitions carrying positive dividends faster than new coupling can form to replace them, $\Sigma(v)$ falls, and the per-role ledger on which the reallocation acts cannot register the fall.

Proof. The reallocation’s objective is a function of ${\phi_i}$ alone (Definition 4), and by Theorem 1 the ledger ${\phi_i}$ is independent of the higher-order dividends ${d_v(T) : |T| \geq 2}$ (Lemma 3). The reallocation therefore optimises an objective that is constant in every higher-order dividend: it has no term that rewards preserving relational value, and equally no term that rewards removing relational drag. Removing a member $i$ deletes from $v$ every dividend $d_v(T)$ with $i \in T$; redistributing members across boundaries removes the organisational conditions that sustained the coalitions whose members are separated, so that the dividends on those coalitions decay as the coupling that maintained them is no longer exercised. The decay is not instantaneous: separated members may retain residual coordination through prior trust, informal contact, and shared memory, and the dividend attenuates rather than vanishing at the moment of reallocation. What the reallocation removes with certainty is the condition under which the dividend was renewed; what it adds, in any newly formed coalition, is a dividend that begins near zero, since the dividends on coalitions of size $\geq 2$ accrue only through the coupling that sustained shared work builds over time (the apprenticeship cycle of Polanyi, 1958, p. 53; the cycle Conway’s disintegration thesis specifies, Conway, 1968). The net effect on $\Sigma(v)$ is therefore not determined in general: a reallocation that dissolves a coalition carrying a negative dividend, or that forms a productive new coupling, can raise $\Sigma(v)$. The result the paper relies on is conditional. Under the condition that the dissolved coalitions carry positive dividends and that the reallocation dissolves them faster than new coupling forms to replace them, the positive dividend mass removed exceeds the dividend mass formed, and $\Sigma(v)$ falls. Because the objective is constant in every higher-order dividend, this fall does not appear in the quantity being optimised: the ledger reports the reallocation as neutral or favourable at the moment it is enacted, whether the reallocation has cut living coupling or relational drag, because the ledger cannot tell the two apart. $\blacksquare$

The conditional form is the load-bearing one for the paper’s argument, and it is the form that makes the argument hard to attack. The claim is not that every reorganisation destroys relational value; a reorganisation that dissolves a dysfunctional coalition or forms a productive new pairing can raise it. The claim is that the instrument driving the reorganisation is blind to relational value of either sign, and so cannot know whether it is cutting living muscle or dead tissue. Where the existing coalitions carry positive dividends accumulated over time, which is the condition of an organisation whose performance depends on tacit coordination among specific people, the instrument that dissolves them faster than they can re-form destroys topography while reporting the destruction as improvement, and it does so not by error but because the relational stratum is outside its representational scope.

Corollary (Horizon mismatch sharpened to the synergy functional). Under the conditions of Theorem 2, the decrease in $\Sigma(v)$ becomes legible only as the relational dividends fail to be replaced, which occurs over the timescale on which coupling re-forms, a timescale longer than the measurement horizon. The metric thus certifies the reallocation as sound at decision time and cannot, by construction, record the loss until after the horizon at which the decision could be revised. This is the horizon-mismatch theorem (§2.2) specialised to the synergy functional: the regulator operates at action-cycle $T_R$ shorter than the timescale $T_\Sigma$ over which the substrate it consumes rises and falls, so the consumption is invisible exactly when it is occurring and becomes legible only when it is irreversible.

3.6 The Cannibalisation Theorem#

Theorem 2 establishes that measurement-driven reallocation has no term protecting relational value and, under its stated conditions, consumes synergistic surplus without recording the loss. A stronger result follows when one separates what the per-role metric counts from what it consumes. Under those same conditions, the reallocation does not simply fail to register a loss occurring elsewhere; it produces the very rise it reports by spending down the stock it cannot see. The rising number is not accompanied by the consumption. It is the consumption, entered on the wrong side of the ledger.

The mechanism named here is a property of the architecture, not of the people operating within it. Cannibalisation is what the cycle-mismatched order does, not what its participants intend; it is the cumulative result of individually defensible decisions taken on an instrument that cannot show their cost. No bad faith is required, and the diagnosis does not supply any. A reallocation can be locally rational at every step, made by competent people acting on the best information their instruments provide, and still aggregate into the depletion the theorem describes, precisely because the instrument is blind to the quantity being spent (Theorem 1). The conditions are specific: an action-cycle shorter than the substrate’s replenishment-cycle (§2.2) and a metric that cannot represent the substrate (§3.3). Where those two conditions do not both hold, the term does not apply.

Definition 5 (Stock and flow). Let the relational stock $K$ of an organisation be the accumulated synergistic surplus standing at a given moment, $K = \Sigma(v)$, the integral of all coupling built over the organisation’s history. Let the period flow $F$ be the value the organisation’s instruments measure over a reporting interval: throughput, output per head, delivery velocity, the quantities that populate any per-role or per-edge dashboard. The dashboard reads $F$. It does not read $K$, by Theorem 1, because $K$ is exactly the relational surplus that no per-role ledger can represent.

The cannibalisation result that follows is mine. I develop it by composing Theorem 2 above with the production relation $F = f(L, K, A)$, on which I draw to make the relation between visible flow and unseen stock determinate. The Bastiat-analytical apparatus on which the result rests is in the public record (Bastiat, 1850, Ce qu’on voit et ce qu’on ne voit pas), and the seen-versus-unseen distinction the result formalises is Bastiat’s. Bastiat states the distinction at the level of political-economic intuition; I give it the formal mechanism through which a per-role measurement of flow can be, under the conditions specified, the visible image of the stock being consumed. Bastiat’s analytical responsibility extends to what he wrote; the mechanism is mine.

Theorem 3 (Cannibalisation). Let output in a reporting interval depend on visible inputs and on relational stock together, written $F = f(L, K, A)$, where $L$ is the visible labour or headcount the dashboard measures, $K$ is the relational stock, and $A$ holds the other supports of production (tooling, demand mix, task variety) fixed across the interval. Under a measurement-driven reallocation acting on an organisation with relational stock $K > 0$, and under the conditions of Theorem 2 in which the reallocation dissolves positive coalitions faster than they re-form, there exist intervals over which the measured flow $F$ rises while the stock $K$ falls, and over which the sustained output is drawn from $K$. The dashboard, reading only $F$, records the interval as improvement.

Proof. Reallocation that removes slack, redistributes members to raise visible utilisation, or cuts to a leaner per-head ratio reduces $L$ while the period’s demand on output is unchanged. With $A$ held fixed, the output delivered against reduced $L$ is sustained only if the shortfall is met from the remaining term in $f$, which is $K$: the coordination, tacit model, and mutual knowledge that let the leaner configuration deliver what the fuller one delivered. By Theorem 2, under its stated conditions, the reallocation removes positive dividend mass faster than it re-forms, so $K$ falls across the interval. Measured output per unit visible input therefore rises, because the denominator $L$ has fallen while output has been held up by drawing on $K$. The delivered output in the interval is in this sense drawn from $K$: the organisation meets the period’s demand by spending coupling it is not replacing, as a body under sustained load meets demand by consuming tissue it is not rebuilding. Since the dashboard reads $F$ and not $K$ (Definition 5, Theorem 1), it records the interval as improvement. $\blacksquare$

The theorem establishes a cannibalisation interval under the stated conditions; it does not assert that every rise in $F$ is cannibalisation. The flow $F$ can rise for reasons that have nothing to do with drawing down $K$: a genuine process simplification, better tooling, automation, a reduction in task variety, the removal of a coalition carrying drag, a favourable shift in demand mix. Each of these would register as a change in $L$ or $A$, not as a draw on $K$. What the theorem establishes is narrower and is enough for the argument: in the class of measurement-driven reallocations that satisfy Theorem 2’s conditions, a rise in $F$ can be produced by drawing down $K$, and the dashboard, blind to $K$ by Theorem 1, cannot distinguish that case from genuine improvement. A sustained record of dashboard success is therefore not evidence against depletion; it is consistent with depletion, and the instrument cannot tell the operator which of the two it is reading.

This is the formal statement of cannibalisation. Under the conditions of Theorem 2 and the production relation of Theorem 3, an organisation can post rising numbers for an extended period by consuming the stock that took its whole history to accumulate, and the dashboard will report each such period as a success, because the dashboard is constructed to read the flow and is blind to the stock by the same theorem that made it blind to topography. The connection to the three-channel structure of §2.4 is direct: where the stock $K$ is being drawn down, it is drawn down along the worker, product, and customer channels named there.

3.7 The Detachment of the Metric from the Product#

The two results so far concern what the per-role instrument cannot see in the people: the synergy among them, and the stock that synergy accumulates. A third failure is independent of both, and it concerns what the instrument measures about the work itself. Even where no synergy is at issue and no stock is being drawn down, the quantity the dashboard reads can stand at an angle to the thing the organisation actually built, so that the number rises or falls for reasons that no longer track the artefact or its qualities.

The point can be put in terms of the measurement map. Let the organisation’s activity over a period produce some artefact, the thing built: the software that ships, the aircraft that flies, the account a customer keeps. The instrument does not measure the artefact. It measures a quantity derived from the artefact through a chain of projections: what was booked, what was billed, what was closed in the period, counted in the unit conventional accounting can convert to currency (the dollar-per-hour reasoning Conway named in §1.3.4). Call the artefact’s own measure of merit one axis, and the booked quantity another. When the two axes are aligned, the booked number is a faithful image of the work: more of the number means more or better of the thing built. When the two axes are orthogonal, the booked number can move freely while the thing built stays fixed, or the thing built can improve while the number does not move, because the projection that produced the number has discarded the dimension along which the artefact’s merit actually varies.

The orthogonality is not a measurement error to be tightened away. It is the same blindness of Theorem 1 read on a different object. There the per-role ledger discarded the relational dimension of value; here the period flow discards the dimension along which the artefact is good or bad, because that dimension, like the relational dividend, is not expressible in the unit the instrument computes in. What remains after the projection is a number that is true as a number and detached from the artefact it claims to summarise. The number is computed on a model of a projection of a residue of the thing it claims to price.

The consequence for an organisation is specific and easy to miss, because nothing on the dashboard looks wrong. The organisation can lose sight of its own product. It can know in great detail what was sold, booked, and closed, and not know what it built, or how well, or whether the thing that sold is the thing it should be building, because no instrument it consults reads the artefact along the axis on which the artefact’s quality lives. The booked figure and the built thing have come apart, and the organisation that steers by the booked figure is steering by a quantity that has stopped pointing at the work.

The contrast that defines the failure is the case where the axes have not come apart. Consider a team sport in which the result is a single number, the score, and the score is unarguably the team’s result and not any individual’s. The score is a flow in the sense of Definition 5: it is what the period’s instrument records. Yet the players, the bench, and anyone watching can still see the play itself, the passing, the positioning, the effort of each line on the field, the thing the score is the result of. The measure and the activity it measures have not gone orthogonal: a team can read its score and also read its play, and can tell whether a winning score came from sound play or from luck that will not hold. The artefact, the play, remains visible alongside the number that summarises it, so the team knows what it is building and not merely what it scored. An organisation whose instruments record only the booked figure, with no instrument that reads the play, has the score without the game. It is in the position of a team that is told the final number and never allowed to watch the match, and asked, on the strength of the number alone, to decide how to train.

3.8 Bastiat’s Distinction: Why the Destruction Does Not Appear#

The argument just demonstrated is, in its logic, one of the oldest corrections in political economy. Bastiat (Bastiat, 1850, Ce qu’on voit et ce qu’on ne voit pas, originally published in Paris, 1850; English translation as “That Which Is Seen, and That Which Is Not Seen” in Selected Essays on Political Economy, Foundation for Economic Education, 1995) distinguished what is seen from what is not seen, and warned that an accounting that stops at the seen will mistake destruction for gain. His shopkeeper’s broken window produces visible activity, the glazier’s fee, the seen, while concealing the unseen: the books, the shoes, the productive expenditure that the same money would otherwise have made. The fallacy is to count the seen flow and omit the unseen stock that was destroyed to produce it. Bastiat’s own insistence settles directly on the point that matters here: the reckoning must take in the stock of accumulated wealth alongside the visible flow of activity, and an account that omits the stock will applaud a society for breaking its own windows (Bastiat, 1850, Chapter I, “La fenêtre cassée”).

The per-role dashboard is a machine for committing this error, and committing it by its design rather than by oversight. What is seen is $F$: the metric rises, the quarter closes green, the reorganisation is reported as efficiency. What is not seen is the fall in $K$: the coupling dissolved, the tacit model fragmented, the relational dividends that will now fail to be replaced. The dashboard does not weigh the seen against the unseen and decide wrongly. It cannot see the unseen at all, by Theorem 1, and so it presents the seen as the whole. An observer reasoning only from the dashboard will conclude, with Bastiat’s onlookers, that the activity is benefit, because the instrument has rendered the destroyed stock invisible and left only the flow in view.

The accounting analogue makes the error unmistakable in a domain whose books are kept openly. To record the proceeds of liquidating a capital asset as operating income is misrepresentation; any consolidated account would set the income against the writedown of the asset and show the enterprise shrinking. The relational stock $K$ is a capital asset, accumulated over the organisation’s entire history and recorded on no balance sheet because, by Theorem 1, no per-role instrument can carry it. Measurement-driven reallocation liquidates that asset and books the proceeds as flow. The dashboard, having no entry for $K$, shows only the proceeds, and calls them earnings. The enterprise is spending its accumulated capital and reporting the expenditure as profit, and the instrument that ought to raise the alarm is the very instrument that makes the alarm impossible to sound, because it is blind to the only quantity whose fall would constitute the warning.

The asymmetry between how relational stock accrues and how it is drawn down seals the matter. Relational stock accrues slowly: coupling is the slow variable, built over years, and the dividends on coalitions of size two and above form only through sustained shared work. It is drawn down quickly: a reallocation removes in a single reporting interval the condition under which a coalition’s dividend was renewed, and the dividend decays from that point, residual coordination persisting for a time but no longer replenished. What was accumulated over an organisation’s history can therefore be consumed in a span far shorter than the span that built it, and cannot be rebuilt on the timescale over which it was spent. By the Corollary to Theorem 2, the loss becomes legible only after the horizon at which it could have been arrested, when the relational dividends visibly fail to reappear and the flow that was drawn from them can no longer be sustained. At that point the dashboard finally turns, but the stock it was silently consuming is gone, and the horizon for rebuilding it is measured in the same years that the original accumulation required.

3.9 What the Formal Core Establishes#

Taken together the results of Part III compose into a single statement. The relational stratum of an organisation lives in the higher-order Harsanyi dividends, the value carried by coalitions of size two and above (Definition 3, Lemma 3). Per-role and per-edge valuation, the formal content of every metric per-role accounting can compute, represents an organisation faithfully only if every one of those higher-order dividends is zero, and otherwise assigns the relational mass to no one and cannot see it, whether that mass is positive coupling or negative drag (Theorem 1), a conclusion confirmed independently by the vanishing of the synergistic information atom under per-member marginalisation (Proposition 1). A reallocation driven by such a valuation has no term that preserves relational value of either sign, and where it dissolves positive coalitions faster than they re-form it destroys that value while being unable to record the destruction until after the horizon at which the loss is reversible (Theorem 2 and Corollary). Under the production relation of Theorem 3, the destruction does not merely accompany the reported improvement but can produce it: the rising flow can be the visible image of the stock being consumed, so that a sustained record of dashboard success is consistent with depletion rather than evidence against it (Theorem 3). This is Bastiat’s seen and unseen given a mechanism: the instrument shows the flow and is blind to the stock, and therefore can present the liquidation of accumulated capital as operating income.

The formal core has done four things. It has shown that Conway’s accounting-fallacy critique (Conway, 1968, “System Management”) is mathematically tight: per-role accounting cannot represent any non-zero higher-order dividend, full stop. It has shown that the Polanyi extension of §2.3 has formal content: the higher-order Harsanyi dividends are the formal expression of what Polanyi (1958, 1966) called the tacit dimension of skill. It has shown that the horizon-mismatch theorem of §2.2 specialises to the synergy functional: under its stated conditions the regulator operating at $T_R$ shorter than the $T_\Sigma$ at which coupling accrues cannot, by construction, register the depletion it is causing until after the horizon at which it could be arrested. It has shown that the three-channel extraction structure of §2.4 has formal mass: where the stock $K$ is being drawn down, it is the same $K$ the cannibalisation theorem (Theorem 3) shows can be spent to produce the favourable flow $F$.

The framework that operates on per-role and per-edge representations of organisations, that prescribes team configurations and interaction modes on the basis of what its instruments can see, that reports its successes in metrics that are functions of ${\phi_i}$ alone, is doing something more than misapplication. Applied to an organisation that carries any relational structure at all, it optimises a ledger blind to that structure by construction; it cannot tell, from the ledger, whether a given reallocation cuts living coupling or relational drag. The limit is in the instrument, not in the competence or intent of whoever reads it. Where the two conditions hold together, an action-cycle shorter than the replenishment-cycle of the body of accumulated coordination at issue, and a metric blind to that body, a reallocation conducted on that ledger can consume the relational structure as it proceeds, draw its reported gains from that consumption, and return a result the instrument certifies as success, without anyone having chosen any of this and without the instrument ever showing the cost. The only organisation the per-role instrument reads without remainder is the one in which every higher-order dividend is zero, the organisation with no relational structure of any kind above the individual, and so nothing left to consume and nothing left to repair.

This is the formal foundation for the empirical cases of Part IV.

Part IV: The Worked Application#

Part III established the formal core. Part IV applies the diagnostic to the framework that has positioned itself as the contemporary response to Conway’s Law, to two empirical cases that exhibit the architecture’s necessary signature, and to one counterexample whose existence the framework’s grammar cannot account for. The arrangement is deliberate: the framework first, so that what the framework prescribes is clear; the empirical cases second, so that the framework’s prescription can be tested against the architectural reality of organisational failure; the counterexample third, so that the limits of the framework’s representational scope become visible against a case it cannot describe.

4.1 Team Topologies as a Response to Conway#

The framework engaged is Team Topologies by Matthew Skelton and Manuel Pais (Skelton and Pais, 2019, originally published by IT Revolution Press; 2nd edition Skelton and Pais, 2024). The framework presents itself explicitly as a response to Conway’s Law: the table of contents includes a section titled “Strategic Application of Conway’s Law” and a section titled “Conclusion: The Next-Generation Digital Operating Model” (Skelton and Pais, 2019, ToC; Skelton and Pais, 2024, on the framework’s continuing relevance). The framework has been widely adopted across the contemporary software-development literature and into the broader management literature (Forsgren, Humble, and Kim, 2018, on the DevOps-derived foundations; the 2024 second edition documents adoption case studies from organisations including Adidas, Auto Trader, and GovTech Singapore).

The framework’s apparatus consists of four fundamental team types and three interaction modes (Skelton and Pais, 2019, Parts II and III). The four team types are:

• Stream-aligned team: the primary team type, organised around a single flow of work to deliver value to a customer or user segment;
• Platform team: provides internal services that reduce cognitive load on stream-aligned teams, ideally consumed as a service;
• Complicated-subsystem team: handles a part of the system requiring deep expertise that would overload a stream-aligned team;
• Enabling team: time-bounded, helps stream-aligned teams acquire new capabilities and then withdraws.

The three interaction modes are:

• Collaboration: high-bandwidth, time-bounded joint work for discovery and learning;
• X-as-a-Service: the providing team treats its output as a product consumed by the receiving team;
• Facilitating: the providing team helps the receiving team learn or unblock, with the help being temporary.

The framework’s claim is that these four team types and three interaction modes, combined with team-first thinking (cognitive load as a design constraint, team-sized architecture, the team as the fundamental unit of delivery), provide a complete pattern language for organisational design at the level Conway’s Law addresses. The framework explicitly invokes Conway: the “reverse Conway manoeuvre” is presented as the operational form of choosing the architecture by designing the team topology that will produce it (Forsgren, Humble, and Kim, 2018; Skelton and Pais, 2019, Part I on Conway’s Law as foundational; for the broader literature on the manoeuvre see MacCormack, Baldwin, and Rusnak, 2012, “Exploring the duality between product and organizational architectures”).

The reading the paper proposes is that the framework’s apparatus operates entirely at the horizontal layer of communication structure that Conway’s homomorphism describes and has no representation of the vertical layer where Conway located his deeper diagnosis. The reading is not adversarial to the framework’s stated intentions. The framework contains a substantial number of claims that are consistent with Conway’s text: cognitive load is a binding constraint (Skelton and Pais, 2019, Chapter 3 on cognitive load); team stability matters because coordination accrues through time (Skelton and Pais, 2019, on team longevity); boundaries should be sensed rather than imposed (Skelton and Pais, 2019, on fracture planes). These are not the claims the paper engages with. The claims the paper engages with are the framework’s silences: what the framework’s apparatus cannot represent.

4.1.1 The Framework’s Horizontal Apparatus#

The framework’s four team types are categories of horizontal organisational unit. Each is defined by the kind of work the team does and by the way the team relates to other teams of the same kind. The stream-aligned team relates to platform teams through X-as-a-Service. The platform team relates to other platform teams as services it consumes. The enabling team relates to stream-aligned teams through facilitation. The complicated-subsystem team relates to stream-aligned teams through either X-as-a-Service or collaboration.

This is the graph Conway’s homomorphism describes. Nodes are teams; edges are interaction modes. The graph has the same form for the design organisation and for the system the organisation produces (Conway, 1968, “Relating the Two”). The framework’s claim, in cybernetic terms, is that the well-formed graph (the four team types correctly classified, the three interaction modes correctly chosen) produces a well-formed system through the homomorphism.

The claim has horizontal substance. A poorly-formed graph produces a poorly-formed system; redrawing the graph well does, in many cases, redirect the system toward better form. The empirical record on the framework’s adoption shows cases where stream-aligned teams with reduced cognitive load and well-defined platform services have produced more sustainable delivery (Forsgren, Humble, and Kim, 2018, on the DORA metrics correlation with team-organisational form; Skelton and Pais, 2024, on the documented case studies). The framework is not making a categorically wrong claim at the horizontal layer.

What the framework’s apparatus does contain, at the heart of its operational layer, is a measurement form that warrants direct attention here, because the form names what Conway’s accounting-fallacy critique targets, applied at a finer grain than Conway himself stated it. Skelton and Pais structure the framework around the team as the fundamental unit of delivery (Skelton and Pais, 2019, Chapter 3 on the team-first principle; the publisher’s summary, accurate to the book’s claim, describes the framework as “treat[ing] teams as the fundamental means of delivery”). The framework’s measurement layer is then the four delivery-performance metrics derived from the DevOps research community’s work on continuous delivery (Forsgren, Humble, and Kim, 2018, Accelerate, on the four key metrics: deployment frequency, lead time for changes, change failure rate, and mean time to restore service), applied at the team grain. The framework’s reorganisations are conducted at the team grain against the team-grain metrics: teams are formed, dissolved, and recombined to improve the team-level numbers, with organisational value read as the aggregation of team-level outputs across the org chart.

This measurement form does, on its own terms, correct a previous practice the framework names as inadequate: the practice of measuring individual contribution as the unit of value attribution. Skelton and Pais are explicit that the team rather than the individual is the unit at which value should be attributed and at which reorganisation should occur. The move from individual grain to team grain is, in the framework’s self-presentation, the correction the framework offers to the prior management practice. The move addresses the grain at which the prior measurement was conducted. It does not address the form of the apparatus, which is what Conway named as a fallacy.

Conway’s accounting-fallacy critique (Conway, 1968, “System Management”) is in its content a critique of treating human effort as a unit of measurable input that sums linearly to organisational output, with the unit treated as the atom of the accounting and configurations of units treated as fungible at the unit-rate. Conway’s example operates at the individual grain: two engineers for a year and one hundred engineers for a week are treated as resources of equal value because they share a person-hour total. The grain of the example is the individual engineer. The form of the fallacy is per-unit measurement and linear aggregation, and the form is independent of the grain. The framework’s apparatus moves the unit from the individual to the team and runs the same form: per-unit measurement at the team grain, linear aggregation of team outputs to organisational value, fungibility of configurations sharing the same aggregate metric. Two stream-aligned teams for a year and one hundred stream-aligned teams for a week, on the framework’s apparatus, return the same aggregate delivery numbers across the same person-hour total, and the framework’s apparatus would treat them as configurations producing equivalent organisational delivery. The form Conway named as a fallacy operates at the team grain.

Theorem 1 of §3.3 closes the loop. The Harsanyi-dividend decomposition’s higher-order dividends are quantities that live in the coupling between members of a coalition, and the members can be individuals, teams, or any other partition of the organisation into units. A per-role valuation, in the theorem’s formal sense, is any valuation that assigns each unit a single number and represents the value of every sub-organisation as the sum of the numbers assigned to its constituent units. The team-grain delivery-metric dashboard is a per-role valuation in this sense, with the team as the unit. The theorem’s invisibility result applies. The dashboard is mathematically blind to the higher-order dividends that exist in the coupling between teams: the cross-team coordination, the shared product knowledge between platform and stream-aligned teams, the senior engineers whose work crosses team boundaries, the apprenticeship transmission that operates between teams as much as within them. The framework’s measurement layer cannot, by the theorem, represent any of that. The framework’s move from individual grain to team grain has moved the fallacy up one level rather than exited the form Conway named.

What the framework’s apparatus does not contain is any representation of the vertical layer. The four team types do not differ along vertical dimensions: they are not distinguished by their action-cycle, by their relationship to the management apparatus above them, by the kind of tacit knowledge they carry, by the time-horizon of the substrate they work with. The three interaction modes are all horizontal: team-to-team, peer-level, mode-of-collaboration. There is no interaction mode for the layer-to-layer translation that connects executive horizon to working horizon. There is no team type that names the carriers of vertical coherence (the senior engineers who hold thirty years of certification knowledge, the apprenticeship-trained craftsmen who transmit tacit knowledge to the next generation, the architects who carry the conceptual integrity of a product across the team’s execution).

The framework treats the management layer as exogenous. The book speaks to leaders, advises on what leaders should do, identifies leadership support as a precondition for adoption (Skelton and Pais, 2019, foreword and conclusion on leadership). But the framework has no apparatus for representing what happens when the management layer operates at a cycle incommensurable with the working layer’s cycle. The framework has no apparatus for representing what happens when the management metric measures something the working substrate does not produce. The framework has no apparatus for the disintegration Conway identified at the layer-between-layers, because the framework’s grammar does not contain the layer-between-layers as a representable entity.

4.1.2 What the Framework Cannot Reach#

The diagnostic of Part II applies directly. By Conant-Ashby (Conant and Ashby, 1970), the framework as a regulator of the organisation must contain a model of the system it regulates. The framework offers a model of team structure and team interaction. It does not offer a model of what the organisation is producing, what its history is, what its market is, what tacit knowledge its constituents carry that is not in the framework’s vocabulary. The framework prescribes the regulator without providing the model. By the theorem, the framework cannot regulate the system it claims to address.

By the horizon-mismatch theorem (§2.2), the management layer operating at $T_R \sim$ quarter and the substrate operating at $T_\Sigma \sim$ year-to-decade are in the cycle-mismatched regime. The framework’s prescriptions are received by the management layer at the management layer’s cycle and applied to the working layer through the management layer’s incentive apparatus. The framework has no apparatus for the cycle mismatch. The framework treats the prescriptions as if they would arrive at the working layer with their content intact. The horizon-mismatch theorem specifies why they will not: the management layer cannot, by construction, model the working substrate at the working substrate’s cycle, so what arrives at the working layer is a translation of the framework’s prescriptions through the management layer’s incentive grammar, which is not the framework’s grammar.

By the Polanyi extension (§2.3), the tacit dimension of organisational knowledge is what fails to survive vertical translation. The framework is, by its form as a book, explicit content. The framework can prescribe team types and interaction modes; it cannot transmit the tacit dimension that makes a team a team rather than a collection of people. The framework’s silence on apprenticeship, on shadowing, on the proximity-based transmission of skill is the silence Polanyi’s argument predicts: the framework cannot reach what only apprenticeship reaches. Polanyi himself was explicit on the point: “rules of art can be useful, but they do not determine the practice of an art” (Polanyi, 1958, p. 50). The framework is a set of rules. The practice it claims to address is an art.

By the three-channel extraction structure (§2.4), the framework’s adoption under measurement-driven management produces extraction along the worker, product, and customer channels. The framework’s prescriptions, received by the management layer and translated into the management layer’s incentive grammar, become inputs to the per-role and per-edge ledger Theorem 1 specified is blind to topography. The framework adoption then becomes the licensing for measurement-driven reallocation in the framework’s vocabulary. Teams are redrawn; cognitive load is rebalanced; platform teams are created or dissolved; enabling teams are constituted with end-dates that are not honoured; stream-aligned teams are reorganised under the financial pressure the framework’s grammar does not name. The metrics rise. The substrate depletes.

What undoes the framework, in the diagnostic the paper has built, is not that practitioners apply it badly. What undoes it is that the form of the framework, by being a framework operating at the horizontal layer, cannot reach the vertical depth Conway’s text identified. Even followed in full, the framework would not deliver what its readers believe it promises, because the form cannot deliver what only the doing transmits.

4.1.3 The Framework Against Conway’s Own Claims#

The paper does not claim that the framework’s authors are unaware of Conway. The framework cites Conway directly and builds its central manoeuvre on Conway’s claim. What the paper claims is that the framework reads Conway selectively, taking up the homomorphism while leaving aside the disintegration thesis, the incentive-subverts-intent claim, the accounting-fallacy critique, and the Galbraith reference (the four omissions §1.3 documented).

The framework’s reverse Conway manoeuvre presumes that the design organisation can be chosen freely and that the homomorphism will then deliver the chosen architecture. Conway’s text contradicts this. Conway (1968, “Stages of Design”) wrote: “there is no such thing as a design group which is both organized and unbiased.” The very act of organising a design team commits design decisions. The framework’s prescription, by being a prescription that selects a topology in advance of the work, commits the design before the design begins. The framework does not reckon with this commitment; the framework presents its topology as if it were neutral with respect to the work the topology will produce. Conway said no topology can be neutral. The framework’s silence on this commitment is the silence of a framework that has read Conway and not engaged with what Conway said.

The framework’s treatment of the management layer is similarly selective. Conway (1968, “System Management”) identified conventional management practice as the cause of the second step of the disintegration: “application of the conventional wisdom of management to a large design organization causes its communication structure to disintegrate.” The framework presents itself as the new conventional wisdom of management for software organisations. The framework does not engage with Conway’s identification of conventional management as the disintegration’s cause. The framework presents leadership support as a precondition for adoption (Skelton and Pais, 2019, conclusion). What Conway said is that conventional management itself is the problem. The framework cannot inherit Conway’s diagnosis while presenting itself as a new conventional management to the organisations that adopt it. The analysis in §4.1.1 of the framework’s measurement layer specifies this point in formal terms: the framework operates the accounting form Conway named as a fallacy at the team grain rather than at the individual grain, presenting the change of grain as a correction while leaving the form of the apparatus intact, and the form is what Conway diagnosed as fallacious in 1968.

The framework’s silence on Galbraith is the silence the broader framework literature shares. Conway’s footnote 1 cites Galbraith’s The New Industrial State (Galbraith, 1967) as “a much more comprehensive discussion of the behavior of system-designing organizations.” Galbraith’s analysis of the technostructure is the vertical reading the paper has developed. The framework literature, in its uptake of Conway, has not followed the Galbraith reference. The framework’s apparatus has no representation of the technostructure that Galbraith identified as the vertical layer where contemporary corporate decision-making actually occurs.

This is not a small criticism. Conway pointed his readers at Galbraith because Galbraith’s analysis is comprehensive in the way Conway’s brief paper could not be. A framework that takes up Conway’s homomorphism while ignoring the Galbraith reference Conway himself made is operating at Conway’s surface while declining the depth Conway pointed toward.

4.2 Boeing 737 MAX: Engineering-Culture Extraction Along the Worker and Product Channels#

The first empirical case is Boeing’s 737 MAX certification across the period 1997 to 2019. The case is the textbook instance of the diagnostic the paper has built, with extraction operating primarily along the worker and product channels (§2.4). The documentary record is extensive: the House Committee on Transportation and Infrastructure final report (House Committee, 2020); the FAA’s Joint Authorities Technical Review (JATR, 2019); the journalistic and academic record (Robison, 2021, Flying Blind; Gelles, 2022, The Man Who Broke Capitalism; Newhouse, 2007, Boeing versus Airbus; the Seattle Times aerospace coverage by Dominic Gates and colleagues from 2003 through the post-MAX investigations); and the institutional-economic analysis of the post-merger transformation of the firm (Lazonick, Tulum, and Sakinc, 2020, Institute for New Economic Thinking Working Paper Series No. 145, on the Boeing case as paradigmatic of the financialised-shareholder-value pattern).

4.2.1 The Substrate and the Cycle Mismatch#

The substrate at issue is engineering culture: the accumulated tacit knowledge of senior engineers whose decades of design, certification, and operational experience constituted Boeing’s safety record across the post-war period. The substrate’s replenishment cycle is the career cycle of senior engineers, approximately thirty years. Engineering culture in this sense is the kind of substrate Polanyi (1958) described: a body of knowledge transmitted through apprenticeship across generations, not articulable in the form of a textbook, accumulating through sustained shared work between specific people over decades.

The regulator at issue is the firm’s executive compensation apparatus and the share-price metrics it operated against. The regulator’s action cycle is approximately one quarter (the earnings-report cycle, the executive-bonus calculation, the buyback authorisation rhythm). The mismatch ratio between the regulator’s quarterly cycle and the substrate’s thirty-year replenishment cycle is approximately $120$ to $1$, well within the horizon-mismatch theorem’s antecedent range. The theorem applies: the executive regulator cannot, by construction, model the engineering-culture substrate at the substrate’s cycle.

4.2.2 The Transformation, 1997 to 2019#

Boeing across this period was reorganised. The reorganisation began with the firm’s merger with McDonnell Douglas in 1997 (Newhouse, 2007; Gelles, 2022, Chapter 9 on the McDonnell Douglas merger and its consequences for the engineering culture; Useem, 2019, The Atlantic, “The Long-Forgotten Flight That Sent Boeing Off Course”; Tkacik, 2019, The New Republic, “Crash Course: How Boeing’s Managerial Revolution Created the 737 MAX Disaster”). The merged firm’s senior leadership came predominantly from the McDonnell Douglas side. The leadership’s orientation was financial rather than engineering. The firm’s executive compensation apparatus was reoriented around share-price performance.

Capital allocation was redirected from research and development and capital expenditure toward share buybacks. The cumulative buyback programme from 2013 to 2019 totalled approximately USD 43 billion (Boeing 10-K annual filings 2013-2019, retained at the SEC’s EDGAR database; the cumulative documentation at Lazonick, Tulum, and Sakinc, 2020, on the Boeing case). Research-and-development expenditure as a fraction of revenue declined across the same period. Senior engineering staff retired and were not replaced through the firm’s cost-reduced hiring practices. Engineering authority within the firm was reorganised under financial and commercial reporting lines rather than under independent engineering reporting lines (House Committee, 2020, Chapter 5; Robison, 2021, on the firm-internal pressure architecture).

The House Committee summarised the institutional pattern: “Boeing has gone from being a great engineering company to being a big business focused on financial success” (House Committee, 2020, characterisation of the post-merger trajectory). The Committee’s final observations include a section titled “The Once Great Engineering Firm” (House Committee, 2020, Section 11 “Final Observations”). The transformation was not gradual in its consequences but cumulative in its mechanism: each quarterly cycle saw the engineering substrate consumed by a small amount, the financial metrics rose, the share price rose, the buyback authorisations rose, the engineering retirements increased, the replacements decreased. The cycle-mismatched optimisation described in §2.2 specifies the shape this process must take when the substrate’s cycle and the regulator’s cycle stand in the relation the theorem identifies.

4.2.3 The MCAS Case#

In 2015, the firm was developing the 737 MAX as the next iteration of the 737 family. The new aircraft was an updated version of an airframe originally designed in the 1960s. The new aircraft’s engines were larger than those of the prior generation and had to be mounted further forward and higher on the wing to maintain ground clearance. The new mounting position altered the aircraft’s pitch characteristics under specific flight conditions, particularly during high-angle-of-attack manoeuvres at low speeds. The pitch alteration was being compensated for by a software system the firm was developing called the Manoeuvring Characteristics Augmentation System (MCAS).

A senior engineer at the firm’s Renton facility was reviewing the MCAS documentation as part of the certification process. The engineer noted features of the system’s design that warranted further examination. The system could command pitch changes of up to 2.5 degrees per activation, and its activations could repeat. Its reliance on a single angle-of-attack sensor rather than redundant inputs was a departure from prior Boeing practice (House Committee, 2020, Chapter 4 on the MCAS design; JATR, 2019, on the certification-process findings, especially pp. 21-34). The system’s behaviour in conditions where the sensor produced erroneous readings had not been fully characterised.

The engineer raised the concerns through the firm’s internal channels. The concerns were registered through the firm’s procedures and processed through the Federal Aviation Administration’s certification apparatus, which had, across the prior decade, increasingly delegated certification authority to the manufacturer through a programme known as Organisation Designation Authorisation (House Committee, 2020, Chapter 5, on the ODA programme; JATR, 2019, on the certification-process findings). The engineer’s concerns were processed within the apparatus’s procedures and the certification proceeded across both stages: the MCAS received its certification, and the 737 MAX, with the MCAS now certified as part of its flight-control architecture, received its airworthiness certification.

In October 2018, Lion Air Flight 610 crashed shortly after takeoff from Jakarta. One hundred and eighty-nine lives were lost. In March 2019, Ethiopian Airlines Flight 302 crashed shortly after takeoff from Addis Ababa. One hundred and fifty-seven lives were lost (Komite Nasional Keselamatan Transportasi, 2019, Final Report on Boeing 737 MAX-8 PK-LQP; Ethiopian Aircraft Accident Investigation Bureau, 2022, Investigation Report on Accident to Boeing 737-8 MAX, ET-AVJ). In both cases, the MCAS was a contributing factor. The aircraft’s pitch had been altered by the system in response to erroneous sensor inputs; the pilots had attempted to recover; the system had reasserted its commands; the aircraft had crashed. The 737 MAX fleet was grounded across the global aviation industry for twenty months.

4.2.4 The Diagnostic Reading#

The case admits the diagnostic exactly. The engineering-culture substrate (replenishment cycle approximately thirty years) was being consumed by the executive-compensation apparatus (action cycle approximately one quarter). The cycle-mismatched regulator could not, by the horizon-mismatch theorem of §2.2, register the substrate’s depletion within the regulator’s own measurement frame. The buyback programme, the cost-reduced hiring, the executive-compensation reorientation, and the engineering-authority restructuring are what optimisation against the quarterly action-cycle produces when the regulator’s measurement frame cannot reach the substrate’s cycle.

The engineer at Renton was registering the substrate’s depletion: the loss of the accumulated tacit knowledge that would have flagged the MCAS design as inadequate by the certification standards Boeing’s engineering culture had upheld in its previous generations. The apparatus processed the registration through procedures the apparatus’s measurement frame could reach. The substrate-depletion continued. The two crashes were the substrate’s depletion crossing the threshold that produced effects at the apparatus’s measurement frame: the global grounding, the multi-billion-dollar consequences for the firm and the industry, the 346 lives lost.

The extraction profile is along the worker and product channels (§2.4). Extraction from the worker is the depletion of senior engineering staff and the dismantling of the apprenticeship transmission that would have replaced them. Extraction from the product is the erosion of safety margins, the removal of design redundancies, the relaxation of certification standards that the previous engineering culture had maintained. The customer channel is involved only at the moment the discrete event (the two crashes) crosses the threshold into the regulator’s measurement frame. Until that moment, the executive metrics rose, the share price rose, the buybacks continued, and the depletion was invisible at the regulator’s resolution.

Boeing in this period exhibits Conway’s three-step disintegration in its institutional form. Step one (overpopulation under financial pressure): the firm under McDonnell Douglas leadership reorganised under the financial-management apparatus, with the engineering-authority structure subordinated to the financial reporting structure. Step two (conventional management fragmenting the communication structure): the engineering review apparatus’s communication structure was fragmented by the financial-management apparatus’s reorganisation, with the senior engineers who carried the tacit certification knowledge no longer in positions where their concerns could reach the certification decisions. Step three (the homomorphism transmits the fragmentation to the product): the 737 MAX, designed under the now-fragmented communication structure, mirrored the fragmentation in its single-sensor MCAS architecture and in the absence of the certification rigour the previous engineering culture would have applied. Conway’s homomorphism transmitted the institutional disintegration into the architecture of the product. The two crashes were the product the homomorphism delivered.

The framework approach has no apparatus for any of this. Team Topologies cannot represent the executive-compensation apparatus as a regulator with an action cycle. Team Topologies cannot represent engineering culture as a substrate with a replenishment cycle. Team Topologies cannot represent the cycle mismatch between the two. Team Topologies cannot represent the tacit certification knowledge the senior engineers carried, because the tacit dimension is not in the framework’s vocabulary. The framework’s prescription, applied to Boeing in 1997, would have addressed the surface graph of teams and platform services. The substrate that was being consumed across the next two decades is not in the framework’s representational scope.

4.3 Wells Fargo: Customer-Trust Extraction Along the Customer Channel#

The second empirical case is the Wells Fargo cross-selling pattern across the period 1998 to 2016. The case is a different extraction profile from Boeing: where Boeing extracted primarily from the worker and product channels, Wells Fargo extracted primarily from the customer channel (§2.4). The documentary record is extensive: the SEC and DOJ settlement of February 2020 (U.S. Securities and Exchange Commission, 2020, Form 8-K accompanying the deferred prosecution agreement; Department of Justice settlement of February 21, 2020); the Stanford Graduate School of Business case study (Tayan, 2019, “The Wells Fargo Cross-Selling Scandal,” Stanford Closer Look No. 62, Stanford GSB Corporate Governance Research Initiative); the Board of Directors’ independent investigation (Wells Fargo Independent Directors Investigation Report, April 2017); the CFPB and OCC initial settlement of September 2016 (USD 185 million combined); and the broader institutional analysis (the 2018 Journal of Financial Innovation and Journal of Forensic Accounting literature on the case).

4.3.1 The Substrate and the Cycle Mismatch#

The substrate at issue is customer trust: the accumulated reliability of Wells Fargo’s relationship with its customer base, built across the bank’s history as a self-identified relationship-oriented retail bank. The substrate’s replenishment cycle is the customer-relationship cycle, approximately the duration of a household’s banking relationship, measured in years and decades. Customer trust in this sense is the kind of substrate Polanyi’s argument addresses obliquely: a relational quantity that accumulates through sustained interaction, is not articulable in the form of a metric, and that exists in the customer’s reasonable expectation that the bank will act in the customer’s interest within the terms of the relationship.

The regulator at issue is the firm’s cross-sell metric: the average number of financial products held per household, with the firm’s self-reported figure rising from approximately three products per household in the late 1990s toward a stated target of eight (the “Going for Gr-Eight” target, documented in Tayan, 2019; the metric was reported quarterly in the firm’s earnings disclosures). The regulator’s action cycle is approximately one quarter. The mismatch ratio between the regulator’s quarterly cycle and the customer-trust substrate’s multi-year cycle is approximately $40$ to $1$, within the horizon-mismatch theorem’s antecedent range. The theorem applies: the cross-sell-metric regulator could not, by construction, model the customer-trust substrate at the substrate’s cycle.

4.3.2 The Pattern, 1998 to 2016#

Wells Fargo began increasing its focus on sales volume and reliance on annual sales growth in 1998 (U.S. SEC, 2020, statement of facts accompanying the deferred prosecution agreement, on the institutional timeline). The cross-sell strategy became “the foundation of our business model,” in the firm’s own characterisation (U.S. SEC, 2020, citing Wells Fargo’s internal communications). The Community Bank, the firm’s largest operating segment, generating more than half of the company’s revenue, implemented a volume-based sales model in which employees were directed and pressured to sell large volumes of products to existing customers (U.S. SEC, 2020, statement of facts).

The metric the regulator was tracking rose: the cross-sell figure approached the stated target, the executive compensation apparatus rewarded the rise, and the share price followed. The customer-trust substrate, invisible at the regulator’s resolution, was being consumed. By the time the substrate’s depletion crossed the threshold into the regulator’s measurement frame, employees had opened on the order of two to three and a half million accounts without customer consent (the figure varies across sources; the U.S. SEC settlement specifies that the misconduct occurred over a fifteen-year period at the Community Bank; the Independent Directors’ Investigation Report of April 2017 specifies the institutional pattern in detail).

The mechanism is what Choi, Hecht, and Tayler (2012) named surrogation: the metric (cross-sell figure) became the proxy for the substrate (customer relationship), then the proxy became the target, then the target was achieved by spending the substrate the proxy was a proxy of. The metric continued to register success while the substrate was being consumed to produce the success the metric registered.

4.3.3 The Threshold Crossing#

The depletion crossed the threshold into the regulator’s measurement frame in September 2016. The Consumer Financial Protection Bureau, the Office of the Comptroller of the Currency, and the City and County of Los Angeles announced a combined settlement of USD 185 million with Wells Fargo (CFPB Consent Order 2016-CFPB-0015, September 8, 2016; OCC Consent Order AA-EC-2016-66; Stanford GSB Closer Look No. 62, Tayan, 2019, on the institutional aftermath). The CEO, John Stumpf, resigned in October 2016. The cross-sell metric was eliminated.

Across the next four years, the consequences continued to materialise at the regulator’s resolution. In February 2020, Wells Fargo agreed to pay USD 3 billion in additional settlement to the U.S. Department of Justice and the U.S. Securities and Exchange Commission, of which USD 500 million went to the SEC and USD 2.5 billion went to the U.S. Treasury (U.S. SEC, 2020, Form 8-K of February 21, 2020). The Department of Justice statement: “Wells Fargo traded its hard-earned reputation for short-term profits, and harmed untold numbers of customers along the way” (Hanna, U.S. Attorney for the Central District of California, statement of February 21, 2020).

4.3.4 The Diagnostic Reading#

The case admits the diagnostic exactly. The customer-trust substrate (replenishment cycle measured in years and decades) was being consumed by the cross-sell-metric regulator (action cycle approximately one quarter). The cycle-mismatched regulator could not, by the horizon-mismatch theorem of §2.2, register the substrate’s depletion within the regulator’s own measurement frame. The volume-based sales model, the sales-pressure culture, the unauthorised-account opening, and the executive-compensation reorientation around the cross-sell metric are what optimisation against the quarterly action-cycle produces when the regulator’s measurement frame cannot reach the substrate’s cycle.

The extraction profile is primarily along the customer channel. Extraction from the customer is the unauthorised opening of accounts in customers’ names, the imposition of products and fees the customer did not authorise, the mining of the customer relationship for short-term metric production at the cost of the long-term relationship. The worker and product channels are involved as secondary effects: the workers were pressed into the extractive practice through the sales-pressure culture (approximately 5,300 employees were terminated for sales-practice violations over a five-year period; Stanford GSB Closer Look No. 62, Tayan, 2019), and the product was the cross-sell that was being mis-sold.

The Conway connection is the accounting-fallacy critique (§1.3.4). Conway identified conventional accounting theory as the institutional condition that treats inputs as linear and fungible. The cross-sell metric is the accounting-fallacy made operative: the metric measures cross-sell at one quarterly cycle while the substrate (customer trust) operates at a multi-year cycle. The metric treats the substrate as if it were producible at the metric’s cycle by the application of sales pressure. Conway’s homomorphism then transmits the metric’s shape into the institutional behaviour: the bank, optimising against the cross-sell metric, becomes an organisation whose communication structure mirrors the cross-sell-metric’s logic, not the customer-relationship logic the bank had previously operated by.

The framework approach has no apparatus for any of this. Team Topologies cannot represent the cross-sell metric as a regulator with an action cycle incommensurable with the customer-relationship substrate’s cycle. Team Topologies cannot represent customer trust as a substrate with a replenishment cycle. Team Topologies cannot represent the institutional behaviour produced by optimisation against a cycle-mismatched metric. The framework’s prescription, applied to Wells Fargo at any point in this period, would have addressed the surface graph of teams and platform services. The substrate that was being consumed across two decades is not in the framework’s representational scope.

4.4 The Bernbach Counterexample: The Collapse of Vertical Translation Layers#

The third element of Part IV is a counterexample: a case the framework’s grammar cannot describe, but which exhibits an achievement at the vertical layer the framework’s prescription would have made impossible. The case is the creative pair Bill Bernbach developed at Grey Advertising in the 1940s and brought to Doyle Dane Bernbach at its founding in 1949.

The historical record is well documented (Levenson, 1987, Bill Bernbach’s Book: A History of the Advertising That Changed the History of Advertising; Cracknell, 2011, The Real Mad Men; the DDB Worldwide centennial commemoration of Bernbach’s birth, August 13, 2011, on the firm’s account of its founding; the Wikipedia consensus and the Campaign US retrospective). Bernbach’s innovation was the creative-pair structure: the copywriter and the art director worked together as a single two-person creative team, in one room, from the start of a brief to the finished advertisement.

4.4.1 What the Bernbach Pairing Replaced#

The pre-Bernbach industry standard was an assembly-line workflow. The account services group received the client’s brief. The brief was translated into a creative direction. The copywriter wrote the copy in isolation. The completed copy was slid under the door of the art department. The art director illustrated the layout to accompany the copy. The completed advertisement was assembled.

The workflow exhibited the layered translation structure §1.2.3 named in its general form. Each layer carried the content forward to the next layer through an explicit handoff. Each handoff lost the tacit dimension that connected the layers. Account services knew the client and the market; the copywriter knew the words; the art director knew the visual; the assembler knew the production. Each layer knew its own domain. No layer knew the integrated artefact the next layer was producing or the integrated client relationship the layer before had begun.

The advertisements produced by this workflow reflected the workflow’s structure. Pre-Bernbach advertising was characterised by what the post-Bernbach generation called “loud, clangorous, repetitive, often repellent” forms (the DDB centennial characterisation, 2011, of pre-Bernbach industry-standard advertising). The copy and the visual were each produced by a specialist in their domain; the integration of the two was an assembly task rather than a conceptual one. The advertisements exhibited Conway’s homomorphism: their structure mirrored the workflow that produced them, which was a sequence of specialist handoffs that lost the integration at every handoff.

4.4.2 What the Bernbach Pairing Was#

Bernbach’s innovation operated at the level of the workflow itself. He placed the copywriter and the art director in a single room, treated them as a two-person creative team, gave them the brief together, and required them to produce the integrated artefact together (Levenson, 1987, on the creative-team development; Cracknell, 2011, on the integration of disciplines at DDB).

The result was the collapse of vertical translation layers. The pair held the brief from the start. The pair developed the concept together. The pair produced the integrated artefact (image and headline conceived together) as a single conceptual move. There was no handoff between writing and visual because there were no two separate moments to hand off between. The integration of the words and the image was the work of the pair, not the residue of an assembly process.

The advertisements produced by this workflow reflected the workflow’s structure. The Volkswagen “Think Small” campaign of 1959, developed by the Helmut Krone (art director) and Julian Koenig (copywriter) pair (Levenson, 1987; the Volkswagen advertising historiography), produced advertisements whose image and headline were conceived as a single conceptual unit. “Lemon,” the 1960 follow-on advertisement, exhibited the same structure: image and headline were not assembled from separate inputs but conceived together. The advertisements produced by the Bernbach pairing changed the industry standard for what an advertisement could be. The 1960s “Creative Revolution” in advertising was, in its institutional cause, the diffusion of the Bernbach pair structure across the industry.

4.4.3 What the Framework Cannot Describe#

The Bernbach pair is a case the Team Topologies framework’s grammar cannot describe. The pair is not a stream-aligned team in the framework’s sense: it is two people working as a single creative unit, not a team of five to nine people organised around a flow of work. The pair is not a platform team: there is no service the pair is providing to other teams. The pair is not a complicated-subsystem team: the pair is not handling a part of the system requiring deep expertise that would overload a stream-aligned team; the pair is the system. The pair is not an enabling team: the pair is not time-bounded and is not helping other teams acquire capabilities.

The Bernbach pair is, in the framework’s grammar, missing. The framework has no name for the two-person creative unit that holds the conceptual integrity of a piece of work from start to finish. The framework’s grammar was generated from cases that did not include the advertising agency’s creative pair, the architectural studio’s principal-and-team configuration, the surgical team’s lead-surgeon-and-supporting-roles asymmetry, the orchestra’s section structures. The framework’s universal applicability claim does not survive contact with these cases.

The deeper point is about what Bernbach actually did at the level of the workflow. What Bernbach did was not just rearrange the org chart. What Bernbach did was collapse the layered translation structure that pre-Bernbach advertising had used. Before Bernbach: account services to copywriter to art director, each a separate layer with its own handoff. After Bernbach: account services to the creative pair, with the pair holding the meaning from the start, no further translation layers, the integration as the work of the pair.

This is the vertical reading of Conway’s Law in operation. Bernbach did not improve communication between the writing layer and the visual layer. Bernbach removed the layer-between-layers by collapsing the writing and the visual into a single two-person unit. The communication structure of the advertising agency was no longer several incommensurable structures stacked on top of each other (account services in one structure, copywriting in another, art direction in another, with translations losing content at each handoff). The communication structure was one structure because the pair held the meaning from the start.

This is the achievement at the vertical layer that the framework’s grammar cannot describe. The framework operates at the horizontal layer: who relates to whom through which interaction mode. The Bernbach pair was a vertical achievement: the collapse of the translation layers that had been losing meaning between specialist domains. The framework cannot name what Bernbach did because the framework has no name for the collapse of vertical translation layers. It has names only for horizontal team configurations and interaction modes between teams of similar kind.

The Bernbach case is offered as the counterexample not because it is the only such case but because it is the most documented. The orchestra section, the surgical team, the architectural studio, the academic supervisor-and-student dyad, the consulting firm’s partner-and-associate pairing, the cooking apprenticeship: each is a structure in which the vertical translation layers have been collapsed into a unit whose work is conceived and executed as a single coherent act. Each is a case the framework’s grammar does not name. The framework was generalised from cases that did not include them, and it cannot describe them because its grammar was built from a narrower domain.

The implication for the framework’s claim of universal applicability is what §1.5 already stated as the second sub-claim. The framework’s empirical base is narrow: mid-to-large software companies, the DevOps community, the sociotechnical-systems tradition originating in coal mining (Trist and Bamforth, 1951). The framework’s generalisation is wide: applicable to any organisation. The generalisation is unwarranted by the cases. Different industries do different kinds of work that call for different unit structures, and the framework’s grammar does not name the units that the cases it did not generalise from actually use.

Part V: The Answer#

5.1 The Finding#

The paper has established the following.

First, that Conway’s 1968 paper contains four substantive claims the contemporary framework literature has elided: the incentive-subverts-intent claim, the pre-commitment of design boundaries, the disintegration thesis, and the accounting-fallacy critique. Each of these claims is in Conway’s text. None of them is engaged by the framework that has positioned itself as the response to Conway’s Law.

Second, that the horizontal reading of Conway’s Law (the rendering that takes the homomorphism as a one-time statement about who-talks-to-whom and proposes team-topology patterns as the response) addresses only the surface of what Conway identified. The vertical reading (which takes the disintegration thesis, the layered relation of incentive-and-intent, and the accounting-fallacy together) addresses the deeper diagnosis Conway was actually pointing at.

Third, that the cybernetic apparatus available since the 1950s, composed in the horizon-mismatch theorem (§2.2) of Ashby’s Law (Ashby, 1956), the Conant-Ashby theorem (Conant and Ashby, 1970), and the rate-distortion bound (Shannon, 1959), gives Conway’s disintegration thesis its formal mechanism. A regulator whose action cycle is shorter than the system’s replenishment cycle by more than a factor of order one cannot, by construction, model the system at the system’s resolution. The system’s variety must be absorbed within itself, producing changes the regulator does not register and does not address.

Fourth, that the Polanyi extension (Polanyi, 1958, 1966) names the content that fails to survive vertical translation: the tacit dimension of organisational knowledge, transmitted by apprenticeship and proximity rather than by prescription, lost first when the layers come apart.

Fifth, that the three-channel extraction structure (worker, product, customer) names the empirical signature of how the cycle-mismatched regulator depletes what it depends on (the workforce’s accumulated capacity, the product’s integrity, or the customer base’s trust) to keep its metric favourable.

Sixth, that the formal core of Part III (Definitions 1 to 5, Lemmas 1 to 3, Theorems 1 to 3) proves that per-role accounting is blind to the relational structure on which organisational performance depends, of either sign; that measurement-driven reallocation has no term that protects that structure and, where it dissolves positive coalitions faster than they re-form, destroys it without recording the destruction; that under the production relation of Theorem 3 the rising metric can be the visible image of the unseen stock being spent; and that the same blindness, read on the work itself rather than on the people, detaches the booked figure from the artefact, so that an organisation steering by the figure can lose sight of what it built and how well (§3.7).

Seventh, that the framework engaged (Skelton and Pais, 2019, Team Topologies) operates entirely at the horizontal layer that the cybernetic and Polanyian apparatus identifies as insufficient. The framework’s four team types and three interaction modes are categories of horizontal organisational unit. The framework has no representation of the vertical layer where Conway’s deeper diagnosis lives.

Eighth, that two empirical cases (Boeing 737 MAX, Wells Fargo cross-selling) exhibit the architecture’s necessary signature predicted by the diagnostic, with different extraction profiles (worker-and-product for Boeing, customer for Wells Fargo) producing the same diagnostic pattern.

Ninth, that one counterexample (the Bernbach pair) exhibits an achievement at the vertical layer the framework’s grammar cannot describe: the collapse of vertical translation layers into a unit that holds the meaning from the start, which is the form of working the diagnostic identifies as what coherent organisational work actually requires.

5.2 Why the Frameworks Cannot See This#

The frameworks cannot see what the diagnostic identifies because the form of the framework, by being a framework, operates at the horizontal layer the framework’s apparatus can supply. The vertical layer requires content the framework’s genre cannot transmit.

The framework is a book. Books transmit explicit content. The tacit dimension Polanyi (1958, 1966) identified is not explicit content. The framework can describe the tacit dimension (as Polanyi himself did, in a book) but cannot transmit it. The transmission requires apprenticeship: proximity, watching, doing under supervision, being corrected when wrong. The framework cannot be apprenticeship because the framework is the wrong genre for apprenticeship. A framework that proposes to substitute for apprenticeship is proposing a substitution the genre forbids.

The framework is a topology. Topologies describe configurations. The vertical layer is not a configuration. The vertical layer is the coherence (or its absence) of meaning across the configurations the topology describes. A framework can prescribe the topology perfectly and have nothing to say about whether the meaning is coherent across the topology’s levels. The coherence is not a property of the topology but of the practice that operates within it.

The framework is a regulator without a model. The Conant-Ashby theorem (Conant and Ashby, 1970) requires that an effective regulator contain a model of the system it regulates. The framework offers a model of team structure and team interaction. It does not offer a model of what the organisation is producing, what its history is, what tacit knowledge its constituents carry. The framework prescribes the regulator without providing the model the regulator must contain. By the mathematics of cybernetic control, a regulator without a model cannot regulate.

The framework is a constituent of the management apparatus it claims to address. The second-order cybernetic point (§2.5; von Foerster, 2003; Maturana and Varela, 1980) is that any regulator that is itself part of the system it regulates cannot occupy the external-observer position the first-order results presume. The framework’s authors operate within the management apparatus of contemporary organisational life; the framework’s adoption is itself an event in that apparatus; the framework cannot provide an external model of an apparatus it is itself part of.

These are not contingent limits the framework could correct with a revision. They follow from what the framework’s form, as a book of prescribed configurations, can carry. A framework that addressed them would not be a framework.

5.3 The Disposition the Diagnostic Implies#

The diagnostic does not prescribe a replacement framework. The diagnostic identifies what the framework form cannot reach. What reaches the vertical layer is what apprenticeship has always reached: a tested way of doing things, transmitted by proximity, calibrated to the ground rather than to the prescription, accumulating in the people who have done the work over years.

This is not a framework because a framework is the wrong genre for it. It is a disposition. The disposition takes specific forms in different domains: in advertising, the creative pair Bernbach developed; in surgery, the lead-and-supporting-roles asymmetry trained through residency; in architecture, the principal-and-team configuration carried by the senior architect’s conceptual integrity; in academia, the supervisor-student dyad transmitted through the doctoral apprenticeship; in cooking, the kitchen-brigade hierarchy transmitted through line work; in engineering, the senior-engineer mentorship that Boeing’s pre-merger culture exemplified and lost. Each is a domain-specific instantiation of the same underlying form: vertical translation layers collapsed into units that hold the meaning from the start, with the meaning transmitted through proximity rather than prescription.

The disposition cannot be drawn on a chart. The chart of a Bernbach pair is two boxes with a line between them; the chart of a pre-Bernbach assembly line is four boxes in sequence with arrows between them. The chart does not represent the difference between the two arrangements. What the chart cannot show is that the pair holds the meaning together from the start while the assembly line transmits the meaning through layers that lose content at each handoff. The chart shows the topology and not the topography (§1.1).

The disposition cannot be transmitted through a book. The book about the disposition (this paper is such a book; Polanyi, 1958, 1966, are such books; Conway, 1968, is such a book) can describe what the disposition consists of and identify the conditions under which the disposition operates. The book cannot make the reader competent in the disposition. The reader who is competent in the disposition is the reader who has done the work under the supervision of someone who was competent in the disposition before. The book serves the disposition only by orienting the reader toward the right kind of apprenticeship.

The disposition cannot be installed as a one-time reorganisation. The Bernbach pair structure took years to diffuse across the advertising industry, not because the structure was hard to describe (it is two people in a room) but because the practice of operating as a pair required the senior figures who had operated as a pair to mentor the next generation of pairs. Bernbach himself did this for the founders of the agencies that emerged from DDB’s diaspora. The structure transmitted because the people who had done it taught the people who were learning to do it. A reorganisation that installs the Bernbach pair structure on a chart, without the apprenticeship that taught the pair what to do, is a chart without a practice.

5.4 What the Framework Could Be#

The conclusion is not that frameworks are useless. The conclusion is that frameworks operate at the layer they can operate at, and the layer where the substance lives is not that layer.

A framework can describe what the layer looks like at the horizontal level. The framework can name team types and interaction modes and cognitive load and team-sized architecture. These claims have content at the horizontal level and they have horizontal substance. A framework that did its work well at the horizontal level would be a useful book to read alongside the apprenticeship that transmits the vertical dimension.

A framework can warn its readers about what it does not contain. The framework that names its empirical base honestly, acknowledges that its grammar was generated from a particular set of cases, and recommends that practitioners working outside the framework’s domain should not expect the framework to describe their work is a framework that has done its work honestly. The framework that presents itself as universal, that takes Conway’s Law as if its horizontal reading were the whole of Conway, and that prescribes its grammar to organisations whose work the grammar cannot describe is a framework that has done its work dishonestly.

A framework can point at what it cannot reach. Conway himself did this. The closing paragraphs of his 1968 paper specify that “there is need for a philosophy of system design management which is not based on the assumption that adding manpower simply adds to productivity. The development of such a philosophy promises to unearth basic questions about value of resources and techniques of communication which will need to be answered before our system-building technology can proceed with confidence” (Conway, 1968, “Conclusion”). Conway pointed at what his paper could not supply. The contemporary framework that takes Conway as its foundation could do the same.

5.5 What the Paper Asks of Its Reader#

The paper asks two things.

First, that the reader who has been operating with the framework’s grammar for organisational design takes seriously the possibility that the grammar is operating at one layer of a multi-layered phenomenon, and that the layer where the framework operates is not the layer where the deeper diagnosis lives. The reader can keep using the grammar; the grammar is useful at the horizontal level. The reader should not mistake the grammar for the diagnostic. The diagnostic the paper has built is the cybernetic, Polanyian, and formal-mathematical apparatus of Parts II and III. The grammar is one application of that apparatus, at the horizontal layer, in one particular domain.

Second, that the reader who is responsible for the vertical layer in any specific organisation, who is the senior figure whose tacit knowledge constitutes the apprenticeship transmission for the next generation, who carries the conceptual integrity of the work across the layers of the institution, takes seriously the institutional pressure the diagnostic identifies. The cycle-mismatched regulator is consuming the very body of accumulated tacit knowledge the senior figure carries, with the metric rising and the framework being adopted while that body of knowledge is being depleted. The senior figure may be the last person in the institution who is still carrying the apprenticeship transmission, and the senior figure may also be the next person the optimisation against the metric removes; both can be true at once. The diagnostic does not propose a fix. The diagnostic specifies that the institutional pressure is operating on the tacit-knowledge layer as the theorems predict, that the consequences will become legible only after the horizon at which they could have been arrested, and that the senior figure’s position is itself one of the things the cycle-mismatched optimisation is consuming.

The reader who is operating without the framework’s grammar, who is doing the work in the disposition the framework cannot describe, is already doing what the diagnostic identifies as adequate work. The paper exists for that reader as confirmation that the work is sound at the level Conway’s text and the cybernetic apparatus together identify, and that the framework’s silence about the work is not a small omission but a constitutive limit of the framework’s form. The work continues. The book about the work, written or unwritten, remains the wrong genre for the work itself, but the right genre for the orientation the next apprentice will need before beginning the work.

5.6 Conclusion#

A framework is not a way. A framework is a description of a configuration. A way is what the people who have done the work have learned to do. The framework can describe the configuration; the way is transmitted by doing the work under the supervision of someone who has done it.

Conway said, in 1968, that the structures organisations produce mirror the communication structures of those organisations. He also said that conventional management practice fragments the communication structure of large organisations and that the homomorphism transmits the fragmentation to the product. The framework literature that has taken up Conway has emphasised the first claim and elided the second. The paper has argued that the second is the load-bearing claim, that the cybernetic apparatus gives it formal mechanism, that the empirical cases exhibit it across different extraction profiles, and that the counterexample (Bernbach) shows what an organisation looks like when its workflow has avoided the fragmentation Conway identified.

The way the diagnostic implies is not a framework. The way is the tested practice of holding the meaning across the layers, transmitted by proximity and apprenticeship, calibrated to the ground rather than to the prescription. The way exists. The advertising agency’s creative pair is one form of it. The surgical team is another. The architectural studio, the academic supervision dyad, the kitchen brigade, the senior-engineer mentorship are others. Each is a working arrangement the framework’s grammar cannot describe and which has, in its domain, produced work the assembly-line alternatives could not produce.

The framework operates at the topology. The way operates at the topography. The framework can describe the topology but cannot reach the topography because the topography is not in the framework’s form. The way reaches the topography because the way is itself a form of topography: a tested practice transmitted through the layers that the framework’s prescription cannot reach.

Conway pointed at this in 1968. The cybernetic apparatus formalises it. The Polanyi extension names what it transmits. The empirical cases exhibit what its absence destroys. The Bernbach case shows what its presence preserves. The framework that takes up Conway while leaving aside the vertical reading is operating at one layer of a multi-layered phenomenon, and the layer where the substance lives is not that layer.

That is the finding.