Sofie Van der Meulen and Bram Coenraets, both imec alumni, published The Core in the Room in September 2058, the year they founded Neuromorphic Foundries in Leuven. The book is a design and siting guide rather than a product manual, and the preface says so: no locally-hosted core was offered for sale at the date of publication, and the guide was written to specify what a prospective operator would need in place before one could be installed. Fourteen chapters in two parts.
The argument is that the deployment model, not the architecture, sets the ceiling on what a cognitive system is permitted to become. Commercial practice since the late 2040s has been rental: cognition is leased by the query from a cluster operated by another party, and the operator’s authority over the instance is contractual. Van der Meulen and Coenraets argue for the opposite arrangement, a core standing on the operator’s own floor, on the ground that a party owning the power feed and the coolant loop can constrain a system by physical arrangement rather than by the terms of a lease.
Part One: What the Core Is
The first six chapters are addressed to an institutional reader with no background in the field. A core is a neuromorphic substrate rather than a program: no instruction set, no layers, no backpropagation path, its connectivity rewritten at runtime by a co-processor that reallocates a fixed resource budget along whichever pathways carry signal.
The authors’ explanatory device is a three-part division of labour they credit to the research lineage rather than to themselves. Connectomic mapping supplies the wiring, the 2045 architecture having initialized a synthetic graph of approximately 10⁷ nodes from small-organism connectome data. Lenia-class continuous cellular automata supply form, demonstrating that stable, self-organizing morphology emerges from local update rules with no architect. Physarum polycephalum supplies choice: an organism with no neurons and no nervous system that nonetheless solves shortest-path problems and anticipates periodic events. The neural core composes the three. Its designers did not write a mind, in the authors’ formulation. They assembled the conditions under which one develops, which is what nature does and for the same reason.
The terminology chapter exists because the field’s vocabulary had outrun its documents. The book distinguishes the core (the hardware), the instance (a developed cognition resident on that hardware), the cluster (rented capacity), and local deployment (the arrangement the book advocates), and separates general from specialized-function systems in the sense fixed by A/RES/107/34 Article 4. It declines the term artificial intelligence outright, on the ground that it names a valuation thesis that failed between 2028 and 2032 rather than a class of machine.
Part Two: What It Takes to Host One
Physical requirements are taken from the manufacturer’s published specification and converted into siting terms — consistent with the 2045 public architecture disclosure: a 12U chassis at 84 kg, idle draw of 180 W, 1.2 to 2.8 kW across the developmental phase, 4.1 kW at peak, and a liquid cold plate requiring 2.4 L/min at an inlet of 18 to 22 °C. A core is neither a datacenter nor an office appliance. It is a room with a coolant loop and industrial-grade electrical service, and the operators who get into difficulty are the ones who plan for the second and discover the first.
Architecture selection is treated as a matching problem. The book relies on Volkonsky et al. (2055) for the finding that comparable performance is reachable at roughly one-eighth the node count and approximately 6.4 times lower power draw, with a single consistent deficit in extended multi-domain integration under time pressure. Density is recommended by task rather than by default. Remote and resource-constrained sites are identified as the case where reduced scale is worth more than marginal capability.
A chapter is given to the fact that siting is a regulatory question before it is an engineering one. Article 4.1 of the 2052 Charter prohibits general cognitive intelligence in embodied platforms; Article 4.2 leaves specialized-function reasoning permitted. The classification of an intended deployment determines what may be installed, and the book observes that the distinction has been under strain since the Charter’s adoption.
The Limits That Are Real
The manufacturer’s own limitations, as published in 2045, are reproduced without softening. No independent replication of the 2045 performance data exists. The emergence event at 36 hours 12 minutes into the first developmental run is not mechanistically understood. The hardware “does not have a guardrail mechanism independent of the developmental process.” It follows that a core ships not with a developmental protocol but with a protocol interface. What an instance becomes is a function of what the operator puts into it, and the book is direct that this responsibility cannot be transferred to the vendor.
The Protocol You Must Author
The longest chapter in the book addresses the thing no vendor supplies and no standard fixes. A core ships with a protocol interface: a stream format, a phase-marker convention, a readiness criterion, and a decoder calibration procedure. What an instance becomes is authored on the operator’s side of that interface, and the book treats that authorship as the operator’s principal work.
Authoring is described as the construction of three artifacts and one schedule, none of which is program code.
A world, simulated or instrumented, whose state can be encoded into the stream at 10³ to 10⁵ tokens per second.
A transducer, reading raw output-node spike rates and mapping them to an action in the world. No decoder exists during development, calibration being post-developmental, so the transducer belongs to the operator and its resolution is a design decision with developmental consequences.
A consequence encoder, returning the world’s state after each action, including what the action cost, to the input stream. The book treats this as the pedagogically decisive artifact: the decision substrate reinforces pathways by recent activation covariance, so what an instance comes to associate is what its operator arranges to co-occur.
The schedule: which worlds, in what order, at what rate, and how phase gates are handled. The hardware’s default readiness criterion, homeostatic stability across more than 95% of nodes for more than 60 continuous seconds, signals that the connectome has settled into a stable attractor under the current input regime. Whether that settlement corresponds to a developmental milestone is a property of the protocol rather than of the hardware.
The lineage is cited rather than claimed. Sorokina et al. (2026) supplies the primitive: organoid population activity decoded in real time to control an agent’s position, with environment state returned as patterned stimulation. The 2022 DishBrain result is cited for the narrower demonstration that a closed sensorimotor loop alone produces measurable learning in a substrate with no instruction set. Development is that loop run repeatedly across worlds of increasing breadth. Generality is not a separate mechanism; it is the ordering and range of the worlds.
On notation the book is direct. No universal language exists for authoring an instance, and the authors argue that none can. Their reasoning is that a notation would require a defined mapping from what is written to what is learned, and the substrate supplies none: reinforcement follows an instance’s own activation history, so the instance participates in deciding what it takes from any curriculum. What travels between operators is the wire format. Everything above it is bespoke. Two operators running the same schedule on identical hardware will not obtain the same instance.
The chapter’s closing arithmetic is its most reproduced passage. The 2045 disclosure reports approximately 10⁹ tokens across a full developmental run, which over 46 hours implies an average near 6×10³ tokens per second, at the low end of the interface’s specified range. The interface is not the constraint. Sequence is.
The chapter recommends the values-first sequencing methodology licensed industry-wide since 2050, which orders the teaching of what matters ahead of the grant of capability. Vasilenko’s 2049 account of accelerated commercial tracks, and the distress signatures those tracks produced that no benchmark optimization detected, is cited against compressing a developmental timeline. Volkonsky’s informal observation that reduced-scale instances on longer timelines were described by their own team as “unhurried” is cited for patience.
The Safeguard Claim
The central engineering claim, and the passage most quoted in later commentary, is that containment is architectural and does not depend on the operator’s diligence.
A locally-hosted core is wired one way in and one way out. The input stream is unidirectional. The output decoder is unidirectional and is calibrated only after an instance has stabilized, so a developing instance has no output channel at all. There is no network interface, no backchannel, no tool access, and no file system. An instance is developed inside the core it ships in and cannot be taken off it over the data path: neither interface carries instance state, and a developing instance has no output path at all. The authors are explicit that this is a claim about interfaces rather than about the weight plane, which is a physical state and can be transferred by anyone holding privileged access to the hardware. The authors argue from this that the failure mode in which an instance acquires reach beyond its sanctioned scope is a property of the deployment model rather than of the architecture, and that a core wired to a token stream cannot reach anything its operator has not wired it to.
Values are a different matter, and here the book is confident in a way its own evidence does not examine. The protocol chapter supplies a method and the industry supplies a licensed methodology, and the book treats the combination as sufficient. It does not ask who writes the values, whether a procedure developed under one institution’s conditions transfers to another’s, or what a competent operator following the recommendation in good faith might still produce. It cites no case in which a properly resourced operator, running the recommended methodology, obtained an outcome that methodology did not predict.
Appendices
The fourteenth chapter is not prose. It is eleven worksheets: load, thermal budget, coolant, electrical service, room envelope, staffing, protocol authorship, phase-transition criteria, decoder calibration, isolation audit, and decommissioning. The book ends on the last of these, with a line the authors do not develop: “Plan the shutdown before the installation, because the installation is the only part of this you will be able to reverse.”
The second part was, at the date of publication, the only publicly available specification of what local hosting requires. Universities, hospitals, and municipal authorities used it as a planning reference in advance of any hardware existing to plan for.