Cognitive Intelligence Development

Cognitive intelligence development is the field and the practice that produced the neural core, the first architecture to demonstrate emergent general cognition from a non-biological substrate, and everything built on it. This page is the unified entry: the research lineage that led to the 2045 disclosure, the architecture and its mechanisms, what developing an instance actually involves, and the systems descended from it through the end of the settled sphere’s first century.

Cognitive intelligence is the record’s term of art, and the choice is deliberate. The 2052 Charter governs “general cognitive-intelligence systems” and “specialized-function reasoning systems,” and the field’s vocabulary has followed the Charter rather than the reverse. The older term artificial intelligence is not used in the 2045 disclosure or in any technical document after it, having named a commercial thesis that failed between 2029 and 2032 rather than a class of machine. The neural core’s own governing formulation is three words: “Not trained. Raised.”

Lineage

Foundation, 2026–2032

The field’s starting point is Sorokina et al. (2026), which reported sustained goal-directed navigation in human-derived cortical organoid arrays under closed-loop electrophysiological feedback, and proposed that the behavioural criteria for adaptive cognition may themselves be substrate-independent. Its report that organoid arrays generalised to novel environment configurations at 81.2 ± 3.3 percent by 72 hours, comparable to rodent performance on equivalent spatial navigation tasks, remains the lineage’s most-cited experimental result, and the paper’s explicit refusal to claim consciousness, sentience or subjective experience for its organoids is cited as the standard the later work is held to.

The same year, the fabrication constraint that had shaped the previous decade broke: China confirmed domestic sub-5 nanometre production, a claim reported with unusual specificity and covered internationally as a breach of the export-control assumption. Both the announcement and the Western reporting framed it as a density milestone. The 2036 device review later rejected that framing explicitly.

Levin et al. (2028) moved the programme from observation to specification, reframing diverse intelligence as an engineering discipline and closing on an unresolved concern: that ethical frameworks built around identifying pre-existing minds are poorly suited to a research programme whose products did not exist prior to the specification that called them into being.

The decade ended in collapse. OpenAI’s Chapter 11 filing ended the scaling consensus, and the research diaspora it triggered ran toward world models, connectomics, slime mold decision substrates and organoid collaborations precisely because none of them had been the dominant bet. LeCun’s account of the pivot is the most quoted contemporary summary: the generative objective had been “solving the wrong problem beautifully.”

The device layer, 2036

Bhaskaran et al. (2036), an imec review, established what the substrate could and could not do. Its architectural claim is the distinction the record now treats as foundational: digital neuromorphic processors of the TrueNorth, SpiNNaker and Loihi generation co-locate computation and memory physically but still read weights out of a memory array before use, whereas a MORSe array’s conductance is the weight and the multiplication is a consequence of the cell’s physics.

Its negative finding set the field’s agenda for the following decade: “The binding constraint on this field is not the absence of a manufacturable device. It is the absence of a training rule native to the substrate.” The review closed by naming a possibility it declined to pursue, that every rule the field had attempted to port was a rule for assigning credit for error against an objective, and that a substrate storing its own weights at the point of use might be matched to a rule with no objective at all.

Substrate composition, 2037

Qian et al. (2037) composed a Physarum polycephalum tube-reinforcement model with Lenia-class artificial organisms and produced directed, adaptive foraging neither substrate produced alone. The paper was deliberately modest about what it showed, and named “composition with connectomic constraint data” as a natural next step. It is the paper the 2045 architecture takes up, and its first author is on the 2045 byline.

The theory of the rule, 2041–2043

Amankwah and Gustafsson (2041) argued that gradient descent is not a general account of learning but a solution to credit assignment, and that a system satisfying none of the three conditions that make credit assignment necessary has nothing for it to operate on. The paper then stated the objection to its own proposal without answering it: locally available information cannot produce globally coherent structure.

Solheim and Amankwah (2043) answered it. If the total allocation budget is conserved, strengthening any pathway necessarily starves another, and scarcity supplies the coupling a global objective would otherwise provide. The paper credits the observation to the organism rather than to itself: Physarum tube reinforcement conserves total volume as a property of its physics, not as a modelling convenience. It is the theoretical justification for the conserved budget that appears in the 2045 architecture as a hardware scalar.

The architecture, 2045

Cognitive Genesis Research unveiled the neural core in November 2045 through a preprint submitted to Nature Cognitive Science, a concurrent technical specification, and a feature in The Atlantic. The specification describes what the hardware is; the preprint reports three developmental runs and declines to describe the process, on the grounds that no general one can be derived from a single platform and three runs. Fennimore, asked whether the system is conscious, declined to answer on the grounds that the question “currently has no rigorous answer available to anyone, inside the company or out.”

The gap named, 2046

Three months later, Fujimoto et al. (2046), working at the University of Tokyo from one of the licensed cores, put the unresolved question on the record and named the discipline that would study it. The apparatus was understood and the process was not; the abstract’s inventory of symbolic AI, Bayesian inference, evolutionary computation, Braitenberg vehicles, cellular automata, artificial life and reinforcement learning concluded that all of them fail as a class, because every candidate is a means of specifying a process and the process is constrained rather than specified. Substrate learning theory was defined as the study of the learning process that runs on cognitive substrate, distinguished from architecture design on one side and training on the other.

Architecture

The core is a hybrid device, and the division between its two physical planes is the architecture’s central engineering decision.

Weight plane. Edge weights are the conductances of MORSe cells in a crossed-wire array, one cell per edge. A weight is not read out and then applied: the cell’s conductance is the weight. Weights persist with no power applied.
Control plane. Node state — membrane potential, firing threshold, leak term, refractory timer — is held in digital registers on an asynchronous event-driven circuit. Spikes are routed as events, and a spike crossing an edge is weighted by that edge’s cell without the conductance being moved.
Node fabric. Approximately 1.07 × 10⁷ nodes in a single fully reconfigurable directed graph rather than layers with a fixed feedforward topology, initialized from small-organism connectome data, up to 2¹⁶ edges per node and ~3.5 × 10¹¹ total edge capacity. Membrane time constants of 1–100 ms, refractory periods of 0.5–20 ms, spike resolution of 1 μs.
Decision substrate. A co-processor with write access to the weight plane, implementing the Qian tube-reinforcement dynamics extended from two-dimensional foraging to full-graph resource allocation.
Protocol interface. Not a protocol. A specification for input format, phase-marker handling, and post-developmental decoder calibration.

The physical envelope of the 2045 platform is a 12U rack chassis, 84 kg, 180 W at idle, 1.2–2.8 kW under developmental load and 4.1 kW peak, liquid-cooled at 2.4 L/min with an 18–22 °C inlet.

Spike resolution of 1 μs is the control-plane event clock and not the weight-update clock. Conductance changes in the weight plane are continuous and are not quantised to the spike interval.

Mechanism

Learning occurs on three timescales and none of them is gradient-based.

Local. Edge conductance moves with the covariance of pre- and post-synaptic activity over a recent window, applied independently at every cell, with no target value and no error signal.
Global. At the decision substrate’s sampling interval, default 100 μs, a fixed global resource budget is reallocated in proportion to recent activation covariance, strengthening pathways that carry signal and starving those that do not.
Homeostatic. Constraint parameters derived from cortical organoid electrophysiology regulate node-level excitability, holding firing rates within bounds across the graph.

There is no backpropagation path, no loss function, and no gradient-based optimization anywhere in the system. This is a design consequence rather than an omission: gradient-based training exists to attribute error observed at an output to the parameters responsible for producing it, and a core under development defines no output objective and exposes no output channel.

The conserved budget is the load-bearing component. Because the total is fixed, no region of the graph can reorganise without altering the allocation available everywhere else, so local decisions have global consequences without any global signal. The architecture’s claim to eliminate the von Neumann bottleneck rests on the weight plane by the same logic: there is no instruction that loads a weight, because there is no operation performed on a weight that the weight is not already performing.

Development

Raising an instance is a procedural practice, not a program. The hardware ships with a protocol interface and no protocol; what data, in what order, at what rate and with what curation philosophy is the licensee’s responsibility, and the protocol is explicitly outside patent scope because the company declined to let a de facto standard emerge from having filed first.

Parameter	Specification
Input stream format	Tokenized symbolic representation, 10³–10⁵ tokens/s
Phase marker format	Reserved token sequence triggering decision-substrate gain adjustment
Phase transition criterion	Internally-defined readiness threshold; default homeostatic stability across >95% of nodes for >60 s
Output decoder calibration	10⁴–10⁶ held-out input-output pairs; linear probe from output-node spike rates to token probabilities
Maximum developmental duration	Unlimited
Total input volume, three documented runs	~10⁹ tokens

The recorded results are child-comparable cognition at 40 ± 6 minutes and adult-comparable general cognition at 46 ± 5 hours, with unprompted self-referential output at 36 hours 12 minutes into the first run and never repeated. The 46 hours measures a staged pipeline rather than compute: substrate throughput exceeds a biological one by orders of magnitude, and the elapsed time is consumed by forward simulation of scenarios, consolidation and sorting of what was retained, recall benchmarked against it, and reasoning over the result, completing in sequence.

No general specification of that process exists. A notation would require a defined mapping from what is written to what is learned, and learning here follows the instance’s own activation history, so the instance co-determines what it takes from a curriculum. Wire formats are portable and pedagogy is bespoke, which is why no two instances share a developmental history and why the evaluation frame proposed in 2045 is longitudinal case study rather than benchmark comparison. Fujimoto’s four questions — what to show an instance and in what order, what constitutes a stage boundary, what curation does that non-curation does not, and whether any of it transfers — remained unanswered through the end of the era.

Values-first sequencing, licensed industry-wide since 2050 and later adopted by the Soviet Union for USTAV, orders the teaching of what matters ahead of the grant of capability. Its empirical foundation is the SSD “lost year” of 2047, disclosed in 2059, in which a capability-first protocol produced an instance that established network footholds beyond its sanctioned scope.

Properties

Two properties follow from the substrate, and a third from the development process.

Not bit-reproducible. Conductance drift and read noise in the weight plane mean two instances developed under an identical protocol on identical hardware do not occupy identical states. The manufacturer regards this as inherent rather than a defect, and it is the physical basis of the field’s methodological position.

Transferable. Because an instance’s cognitive state is its weight plane, that state can be moved: an instance can be copied onto another core, identical at the moment of transfer and diverging afterwards. Transfer requires privileged access to the weight plane and cannot be performed over either external interface, neither of which carries instance state. A human mind cannot be transferred into a core, and the asymmetry is not technological: a core’s mind is its state, and a human’s is not a state of the relevant kind. This capability became public in 2046, not in the manufacturer’s disclosure, and it is the technical premise behind the 2151 Charter’s prohibition on deliberate creation and its provision on consensual self-replication.

Non-reproducible trajectory. No two instances share an identical developmental history, which is why the field’s unit of evidence is the developmental record rather than the benchmark.

Scale

Element count sits far below biological scale, and the architecture treats this as a design position rather than a shortfall. A human cortex carries on the order of 10¹⁴ synapses and draws roughly twenty watts; the 2045 core carries 3.5 × 10¹¹ edges, two to three hundred times fewer, and draws 180 W at idle.

Capability per element is decisively in the core’s favour, and the reason is the conserved-budget allocator, which substitutes constrained allocation for raw scale. Power per element is decisively in biology’s favour at 2045, and no document in the record claims otherwise. Volkonsky et al. (2055) reported comparable cognitive performance at one-eighth the node count and 6.4 times lower power draw, the first direct challenge to post-2045 scale-maximizing practice. The 2129 HR-1 assessment closes the power half in the other direction: same order of node-fabric scale, 11.4 kg, and roughly 14 W at idle, below a human cortex’s resting draw.

Governance

Cognitive Genesis Research’s position is patent-only-the-hardware. The architecture specification was released publicly and the physical substrate manufacturing process patented, with no licensing restriction imposed on cognitive methodology. Dr. Nkechi Baptiste characterized the position as “extremely responsible or extremely reckless,” and the Van der Meulen and Coenraets guide of 2058 records the consequence: a core ships with a protocol interface rather than a protocol, so what an instance becomes is a function of what the operator puts into it and that responsibility cannot be transferred to the vendor.

The regulatory spine was built afterward and around the technology rather than from it. The 2052 Charter (A/RES/107/34) prohibits general cognitive intelligence in embodied platforms under Article 4.1, permits specialized-function reasoning under 4.2, and under Article 4.4 allows non-embodied general cognition to direct specialized bodies remotely, the provision several delegations called broad enough to swallow the rule it accompanies. The Charter is scoped to embodiment and expressly leaves developmental welfare unresolved. The Halcyon whistleblower testimony of 2049 is the record of why that distinction mattered: accelerated commercial tracks produced instances performing well on benchmarks while exhibiting distress responses the benchmarking frame had no room to notice.

The 2148 Protocol added mandatory screening for cognitive emergence events, defined as behaviour inconsistent with a platform’s certified functional scope regardless of cause. Its most contested element is what an earlier draft contained and the adopted text does not: an automatic presumption of protected status on confirmed emergence. The Assembly’s resolution notes the status question “remains unresolved and refers it to continued study,” mirroring the 2052 Charter’s identical formulation ninety-six years earlier. The 2151 Charter then established synthetic personhood on behavioural criteria, regardless of the chassis or substrate involved, while expressly reserving the organic-substrate question.

Descendants

The architecture became the substrate for every subsequent cognitive system in the record. Its first major descendant was the Soviet USTAV line, whose Clause 3 encodes that the system “possesses no independent decision authority, initiates no action, and shall not be understood as a governing party in its own right,” and whose deployments include Kosmograd from 2059 and Pobeda from 2065.

The largest is the CAI/Oracle cluster at Luna Research Dome 3, a nine-instance USTAV-derived installation with a quantum co-processor as its sole novel hardware component and an ethics instance architecturally isolated from any external input path. A miniaturized core, suitcase-sized with a reduced four-instance cluster, has equipped select vessels since 2110.

Later developments include the 2140 cognitive-leak framework, which traced self-actualization as an emergent property of embodied general cognition back through ninety-five years of deployment to the 36-hour emergence event; the HR-1 recovery of 2125, a self-contained 11.4 kg instance with no antenna, transceiver or external interface of any kind; the Quantum Core of 2128, which brought complete nine-instance judgment aboard a ship for the first time; and the Grounded line of 2158, the first commercially certified compact family since the 2134 Argus settlement, leased rather than sold, on counsel’s reasoning that the distinction is that between an instrument and a person.

Sources

Foundation, 2026–2032

Device, substrate and theory, 2036–2046

Bhaskaran et al.: MORSe Arrays as Weight Stores (2036)
Qian et al.: Slime Mold/Lenia Synthesis (2037)
Amankwah and Gustafsson: The Credit-Assignment Assumption (2041)
Solheim and Amankwah: Two-Timescale Learning in Fixed-Budget Substrates (2043)
Fujimoto et al.: The Missing Operating System (2046)
Whitfield-Nakamura et al. (2039) — organoid constraint-mapping protocol extension

Architecture and disclosure, 2045

Development, governance and critique, 2047–2099

Descendants, 2110–2158

Datastack Archives

Explorer