The pair of eyes we made
When we teach the Viable System Model, System 4↗ usually arrives as a pair of eyes. It looks outward, scans the environment for threats and opportunities, and reports what it finds to the rest of the organization. The picture is compact and can be drawn in a few seconds. It also maps without friction onto a department that already exists under the name of strategy, intelligence or market research. This is why it travels.
We think this is a reduction, and more than a convenient shorthand. The case for this simplification has to rest on evidence: several current teaching formulations set side by side, quoted precisely, with the scope and date of the search recorded so that a reader can check whether we sampled fairly and accurately. We assembled a first set on 24 September 2026: we found 25 independent web texts, four practitioner textbooks and eight pages of our own training material. Thirteen of the 25 web texts describe System 4 chiefly as scanning or looking outward,1 and only three say that it holds a model of the organization itself. Our own training pages follow the scanning pattern on all eight pages checked. It is a convenience sample from one search engine, so the claim describes what we have taught and what we found taught; it does not measure the field.
We write this in the first person intentionally. We did not simply inherit the reduction from a careless secondary literature; as a community we reproduced it. Eyes are easy to draw, easy to teach and easy to assign to an existing unit, and Chapter 9, Why None of This Ever Cost Us Anything, describes the conditions under which a simplification like this persists without correction. Its cost is specific. When sensing is presented as the whole of the function, modeling reads as an optional activity, something a different department performs afterward, and the question of what System 4 is for drops out of the teaching.2 Chapter 1, The Mapping We Teach Is Not the Mapping He Wrote, sets out the general pattern. What follows traces one instance of this phenomenon in detail.
What Beer intended System Four to do
Read against Beer’s own text, System 4 is much more than a set of sensors scanning the environment for future trends, changes and signals: it is the model-building and anticipatory capacity of the viable system. In Brain of the Firm he writes that System Four “must contain a model — some model — of the corporation”, and in The Heart of Enterprise that it “must contain a model of the viable system of which it is the System Four”. That model is where the organization keeps its account of the environment and of its own position in it. A model of this kind does more than record what is currently the case. It forms expectations, checks incoming signals against them, and revises its account fast enough for the revised version to be usable in action; Beer calls this foresight by “rapid-time simulation”. Scanning is one input to that work and does not describe the whole of it.3
The same passages change how the loop between System 3 and System 4 can be read. In the usual reading the loop is a hand-over: System 4 reports what it has seen outside, and System 3, with System 5 arbitrating, turns the report into instruction for the operations. Beer describes something less one-directional. In Brain of the Firm he speaks of a “homeostatic loop between sensation and action” and of a feed-forward loop whose predictions “are compared with real-life outcomes”; in The Heart of Enterprise the intended impact of an outward act “is fed back” to System Four’s model of itself. If those passages carry the weight we put on them, what System 4 sends into the homeostat is a prediction about what the operations will show next, and what System 3 sends back is the check. An instruction issued on the strength of that prediction is a wager, framed so that the next round of operational reporting can settle it. Prediction and report then sit on the same loop rather than at opposite ends of a hierarchy.4
We state this conditionally because the step from Beer’s passages to “the homeostat as a checkable wager” is ours, and we will not claim more for him than the passages warrant. The phrase “prediction channel” appears in none of the eight books we searched. One further temptation needs resisting. The predictive-processing vocabulary of current neuroscience fits this reading closely enough to be seductive. Where we use it, the comparison is ours and says nothing about what Beer wrote or intended.
The anatomy carrying the argument
Beer’s account of regulation drew on the neurophysiology available to him, and part of that neurophysiology is load-bearing rather than illustrative. The item at issue is a reticular-formation model associated with Kilmer, Blum and McCulloch and dated to 1969. We name it because the research brief names it, and because Beer names it himself: the bibliography of the second edition of Brain of the Firm lists, under arousal mechanisms (System Four), W. L. Kilmer, J. Blum and W. S. McCulloch, The Reticular Formation (Air Force Office of Scientific Research, Arlington, Virginia, 1969). What that report actually claims remains to be verified. Until it is, we are naming a target for checking, not citing a source.
We decline the easy argument, that old anatomy invalidates the organizational logic built on it. Chapter 28, What Would Count Against It?, sets out the distinction we rely on here. An analogy can be decorative, and then replacing it costs nothing. Or it can be load-bearing, and then something in the organizational argument moves when the substrate moves. Which of these applies to the reticular-formation model has to be established passage by passage, and not assumed in either direction.
Set against that older substrate is the account described in the research brief: an arousal architecture running from parabrachial nuclei to the basal forebrain, in which monoamines act as gain controls rather than as a single energizing command. The regulatory logic differs, and the difference matters before any organizational consequence is drawn. In the older picture, a command signal switches a system into a state. Gain control adjusts how strongly many distributed processes respond to their own inputs, so that what changes globally is sensitivity rather than instruction. If the organizational inferences we teach depend on the first picture, they need re-deriving under the second.
Action selection: the function the diagram omits
A system that senses and models still faces a problem neither sensing nor modeling solves. Most of the actions available to it are mutually incompatible, and something has to decide which one is expressed. This is action selection, and the five-system diagram does not name it.
The comparative literature describes the vertebrate solution in terms we are borrowing, so we need to state them accurately. The output nuclei of the basal ganglia are tonically inhibitory: at rest, they suppress. The default is therefore suppression of the available actions rather than quiescence, and selection occurs when inhibition is released from one candidate while the others stay held back. Grillner and Stephenson-Jones are the starting point we have been given for the comparative lineage. The claims about tonic inhibition, release from inhibition and homology each need checking against named sources, and should not travel as a package.
Pictured this way, the central regulatory problem changes shape. The question becomes how a field of possible actions is held back, compared and selectively released, where before it was how a command center issues an action. That architecture differs from a pair of eyes wired to a vertical command line, and it is not obvious where the VSM puts it.
The evolutionary depth of this arrangement matters only if it can be sourced. A claim in circulation says the logic has been continuous for approximately 560 million years. If that survives verification, action selection is an ancient organizing problem that any viable system has had to solve, and no late refinement of cognition. We record it here as a claim under verification. If it does not survive, the number comes out; we will not soften it with a hedge.
The model without a forward model
Our textual search of the defined Beer corpus returned a negative result: no function distinct from System 4 appears to perform forward modeling, in the sense of predicting the sensory consequences of the system’s own action. We searched eight Beer books on 24 September 2026 in the OCR text of our calibre copies.5 “Forward model” and “prediction channel” occur in none of them. The function, however, turns up inside System 4 under other names, as foresight by “rapid-time simulation” and as a feed-forward model whose predictions are compared with outcomes. The result stays provisional: OCR can hide hits, and the 1972 edition of Brain of the Firm was not searched.
Taken at face value, the finding leaves two options, and both cost something. Either forward modeling already belongs inside System 4, and then teaching System 4 as environmental scanning has quietly removed one of its central operations from what we pass on. Or it does not belong there, and then the five-system architecture lacks a separately specified function that a system needs in order to anticipate the consequences of its own action. We are in no position to choose between these by analogy with neural architecture, and we do not propose to.
Change in the world, not change of kind
System 4’s anticipation has an outer limit. It equips a system to model changes in its environment while it remains the kind of system whose viability the model exists to preserve. By itself, it does not account for a system becoming a different kind of system.
That boundary is all we claim. The morphogenesis argument belongs to Chapter 7, A Machine That Never Became One, and we do not restate it. We mark where System 4’s competence ends only so that the dependency analysis below is not asked to carry a question it was never framed to answer.
The receipt: what falls with the substrate
Each claim below is listed on its own and put as an open question about what survives if the 1969 substrate is replaced by the modern account. We do not answer the questions here. Answering them is the work this chapter asks for.
- System 4 must sense the environment. Does this follow from the requirements of viable regulation independently of the 1969 anatomy, or does the argument as we teach it route through that anatomy?
- System 4 must maintain a model. Is this secured by Beer’s own argument, by the good-regulator tradition and the internal-model principle examined in Chapter 3, The Law We Quote Is Not the Law Ashby Wrote, or by none of these without further conditions being supplied?
- System 4 must anticipate. What kind of prediction is actually required, and what evidence establishes anticipation as a distinct organizational function rather than a description of competent modeling?
- The System 3–System 4 homeostat carries predictions, not only reports. Does this reading survive replacement of the reticular-formation model, and if so, on what warrant?
- The five-system partition is complete. Where does action selection reside, and does locating it inside System 4 preserve the partition or turn System 4 into an unconstrained remainder that absorbs whatever the other four do not cover?
- Neural organization licenses organizational design. For each claim that survives, is the warrant formal, empirical, analogical, or merely suggestive?
Replacing an obsolete substrate need not abolish System 4. The likelier outcome is a more exact account of modeling, prediction and action selection than the pair of eyes ever gave us, with a clearer statement of which parts of that account rest on argument and which on borrowed anatomy. Part Five takes up what such a reconstruction would require of us, and the Coda, Judy, asks what we owe the model once the work is done.
Notes
- Beer’s name for the function is the “outside-and-then” (Stafford Beer, Diagnosing the System for Organizations (John Wiley & Sons, 1985), 111) and the “development directorate” (Brain of the Firm, 2nd ed. (John Wiley & Sons, 1981), 181; The Heart of Enterprise (John Wiley & Sons, 1979), 235–36) — read. “Environmental scanning” is the term in strategic management (Francis J. Aguilar, Scanning the Business Environment (New York: Macmillan, 1967); H. I. Ansoff, “Strategic Issue Management,” Strategic Management Journal 1, no. 2 (1980): 131–48) and “horizon scanning” the named practice in futures studies (E. Amanatidou et al., “On Concepts and Methods in Horizon Scanning,” Science and Public Policy 39, no. 2 (2012): 208–21; T. Könnölä et al., “Facing the Future: Scanning, Synthesizing and Sense-Making in Horizon Scanning,” Science and Public Policy 39, no. 2 (2012): 222–31) — not read. One text in our sample uses the phrase itself: Dave Hulbert, “Wardley Leadership Strategies,” 12 October 2025. ↩
- The claim is about what the teaching names, not about what the system does. In the system the two are not separable: System Four is “concerned to manage the outside-and-then, but to provide self-awareness for the System-in-focus” (Stafford Beer, Diagnosing the System for Organizations (John Wiley & Sons, 1985), 115), and “it is just this infinite regression of self-images that seems to hold the key to the characteristic self-awareness of viable systems” (ibid., 116) — read. What drops out of the teaching is the name for the modeling, not the modeling. ↩
- Stafford Beer, Brain of the Firm, 2nd ed. (John Wiley & Sons, 1981), 182–83: “the homeostatic loop between sensation and action”; “System Four must contain a model — some model — of the corporation”; a corporate plan “must continuously abort”; and 189, a “feed-forward” loop whose predictions “are compared with real-life outcomes.” Stafford Beer, The Heart of Enterprise (John Wiley & Sons, 1979), 235: “System Four must contain a model of the viable system of which it is the System Four”; and 240, where the “intended impact” of an outward act is fed back to “the System Four model of itself.” Two terms here belong to Beer rather than to current usage. “Homeostatic” is apt for the Three–Four homeostat, in which each system controls the other (see the glossary), but the relation Beer calls a homeostatic loop between sensation and action holds no set point and is not homeostasis in the physiological sense; the modern account of the same relation describes gain control and selective release, not a maintained equilibrium (see the action-selection section below). And the plan that “must continuously abort” is not a method from operational research: Beer gives his reason in the same sentence — “unless that plan is continuously aborting, how shall sanity (the match between the world and the firm’s notion of it) be preserved?” — that is, the plan is held provisionally so that outcomes can overturn it, which is a feedback requirement and not an optimisation. ↩
- The phrase “prediction channel” occurs in none of the eight Beer books we searched (see the note on the search below). The reading of the System 3–System 4 homeostat as a checkable wager is therefore ours, supported by Beer, Brain of the Firm, 182–83 and 189, and Beer, The Heart of Enterprise, 240. The sentence that described “the channel between System 4 and the command channel” was dropped on 12 October 2026, in the rewrite of this section around the System 3–System 4 homeostat: Beer routes System Four’s indications into System Five, not into the command axis (Beer, Brain of the Firm, 2nd ed. (John Wiley & Sons, 1981), 181: “they too are switched into System Five”), and the adaptation link he names is the Three–Four homeostat (Beer, Diagnosing the System for Organizations (John Wiley & Sons, 1985), 120: “The Three-Four homeostat is the organ of ADAPTATION for the enterprise”). ↩
- The search covered the OCR text layer of our calibre copies of Brain of the Firm, 2nd ed. (1981); The Heart of Enterprise (1979); Diagnosing the System for Organizations (1985); Decision and Control and Platform for Change (1994 reprints); Designing Freedom (unpaginated transcript); Think Before You Think (2009); and Beyond Dispute (1994). Terms: “forward model”, “prediction channel”, “feed-forward”, “predict”, “anticipat”, “foresight” and “simulat”. “Forward model” and “prediction channel” have no hits. The function appears in Beer, Brain of the Firm, 183 (foresight by “rapid-time simulation”), 189 and 196 (fast-time simulation), and Beer, Heart of Enterprise, 229 and 240. For the term as defined in motor control (abstract only): R. C. Miall and D. M. Wolpert, “Forward models for physiological motor control,” Neural Networks 9 (1996): 1265–79. ↩