Morphotransformation

Chapter 1. Morphotransformation

Life and description · Two spaces · GTR0 as the fourth transformation · DTR0 and RTR0 · From programming to germination

1. 1. Statement of the task

The book GNSS established the general theory of life: life is a system of bidirectional transformations. The book GTOM described the running human subjective reality through the three upper transformations (GTR1, GTR2, GTR3) that operate inside an already existing neural network. The book GNET specifies the protocols and data formats for implementing the network of these transformations on an engineering substrate.

Between the general theory and the concrete specification there remains a gulf. GNSS says that life is a system of transformations between spaces of different ontological nature. GNET regulates how network nodes exchange values. But between these two books, a central engineering question goes unanswered: where does the running network itself come from? Who creates the concrete nodes, who establishes the relations between them, who decides which functional organs are needed and in what proportions?

In traditional programming the answer is trivial: the network is created by a programmer. The programmer writes code describing every component and deploys it to the infrastructure. Every connection, every parameter is the result of an explicit human decision. But this does not apply to Gativus. A subjective-reality network contains millions of nodes that must be connected with one another according to specific rules, on a scale incompatible with item-by-item programming. A biological brain is not programmed. It grows from a seed.

The book MOGE answers the question of the origin of the running network. Its subject is morphogenesis: the process in which a description unfolds into a running network. This process is not reducible to a program run. It is structurally fully isomorphic to the other three transformations of Gativus, and constitutes their fourth variant — a transformation that creates the substrate itself for the other three.

This chapter introduces the central concept — morphotransformation GTR0 — and establishes its place in the overall architecture. Without understanding GTR0 as a structural analogue of GTR1, GTR2, GTR3, it is impossible to explain why MOLD is constructed precisely as it is, why the seed is self-sufficient, and why morphogenesis differs in principle from programming.

1. 2. Life as morphotransformation

Life by its nature is a morphotransformation. Its structural definition is: life is the ability of one and the same reality to exist in two mutually transformable forms — a folded form (the description) and an unfolded form (the running organism).

The biological illustration is obvious. A plant seed and a mature plant represent two different forms of the same entity. The seed is static, compact, without metabolism; the plant is dynamic, volumetric, in active material exchange with the environment. Between them act two mutually inverse transformations: the seed unfolds into a plant during germination, while the mature plant folds part of itself into new seeds. Neither of the two forms is more fundamental. A seed unable to unfold is a dead document. A plant unable to produce seeds is an evolutionary dead end. Life exists precisely as a reversible transformation between them.

The same principle operates at the level of a single cell. DNA is the folded form; the running cell with metabolism, membrane processes, and signal exchange is the unfolded form. The connection between them is bidirectional: DNA unfolds into a cell through gene expression, and during division the cell folds its own structure back into a copy of DNA for transmission to its descendants. Reversibility here is just as essential as at the level of the whole organism.

A point of principle: neither form exists autonomously. A folded description without an unfolding apparatus is useless — it is a record that no one reads. A running organism unable to fold itself into a description for its descendants is a one-off structure without an evolutionary future. Only the pair of mutually inverse transformations between the two forms constitutes what we call life.

This structural definition of life admits generalization beyond biology. Any system that implements a reversible morphotransformation between a description and a running embodiment belongs to the same category. Gativus is such a system — built on another substrate (silicon instead of carbon), but reproducing the same fundamental mechanism.

1. 3. The two spaces of Gativus

The morphotransformation of Gativus operates between two spaces fundamentally distinct in ontological nature. These two spaces — the space of ontogeny and the space of the running network — parallel the two forms of life described above, but are formalized for engineering implementation.

a) The space of ontogeny

Contains descriptions of organisms. Its basic element is MOVE (MOrphology VEctor), a compact representation of an organism along several graphical coordinate axes. For the detailed definition of MOVE as a 16-dimensional coordinate vector, see Chapter 2.

The space of ontogeny is static: descriptions do not change by themselves. It is compact: a description of several kilobytes produces, upon unfolding, millions of nodes. It is parametric: the description does not contain absolute values, and the scale of a particular organism is determined by the resources available at the moment of unfolding. It is verifiable: errors in the description can be detected before unfolding. Its biological analogue is DNA.

b) The space of the running network

Contains NDDI nodes in their running state and the established relations among them. Every node is active: it executes code, exchanges datagrams, stores state.

The space of the running network is dynamic: the state of the nodes changes continuously. It is voluminous: it contains millions or even billions of nodes. It is deterministic in the sense of addressing: every node has a concrete UNON, every relation has a concrete connector address. It is observable: any node can be queried over the network. Its biological analogue is the living organism.

c) The bijection between the two spaces

Every element of one space has its dual in the other. This correspondence is bidirectional and forms a bijection: every entity in the description corresponds to exactly one entity in the running network, and vice versa. These are not two isolated worlds, but two ontological projections of one and the same reality.

The correspondence between the basic entities is fixed by the following table.

The bijection between the two spaces is a foundational property of the Gativus architecture. It allows a node in the running network to call its own ontogenetic dual and analyse its own construction — the basis of the system's capacity for self-analysis. This aspect is discussed in detail in Chapter 5.

1. 4. GTR0 in the family of four transformations

The book GNSS established that Gativus contains four transformations, numbered GTR0–GTR3. The upper three (GTR1, GTR2, GTR3) are informational; they operate inside an already existing neural network and form subjective reality. The fourth, GTR0, differs in principle: it creates the substrate itself on which the upper three operate.

The numbering of GTR0 as zero reflects a coincidence of two situations. Evolutionarily it is primary — the morphogenesis of cells precedes neural tissue, and even more so the emergence of symbolic and axiological levels. Architecturally it is fundamental — without morphotransformation there is no substrate for the other three transformations. GTR1, GTR2, GTR3 are upper structures that stand upon the results of GTR0.

Despite this difference in role, all four transformations are structurally isomorphic. Each consists of the same two steps — convolution and splice. Each operates between two spaces of different ontology. Each has an atomic unit formed by splice, and a composite unit assembled from atomic units. Each has a direct and a reverse form. They differ only in substrate and time scale.

The summary table of the four transformations is given below. For a detailed discussion of GTR1, GTR2, GTR3 see GNSS and GTOM; here they are given only to display their structural isomorphism with GTR0, which is the subject of this book.

The atomic unit of GTR0 — MORN (MOrphogenesis Node) — forms a triad class — pattern — group, isomorphic to the OPRN, KLEN, WILL triads of the other three transformations. The composite unit MLOM (Morphological LOM) is assembled recursively from MORNs, forming a multi-level structure: micro-columns from nodes, columns from micro-columns, organs from columns, organism from organs. For detailed definitions of MORN and MLOM, see Chapter 4.

The structural isomorphism of the four transformations is neither an accidental coincidence nor the result of forcing them into a shared mould. Evolution found the optimal solution for the following task: extracting invariants from a data flow and erecting upon them a structure with dynamic axes. The solution proved to be universal and was applied successively to four different substrates — from the cellular to the axiological. Gativus inherits this universality and uses one and the same architectural scheme for all four levels.

1. 5. Direct and reverse morphotransformation

Each of the four GTR# transformations exists in two directions. The direct one, denoted DTR# (Direct TRansformation), folds data of the base space into a compact representation in the target space. The reverse one, RTR# (Reverse TRansformation), unfolds the compact representation back. The analogy with the terminology of the Fourier transform is complete: DTR corresponds to FFT (analysis), RTR to IFFT (synthesis).

For GTR0, both the direct and the reverse directions have concrete technical content.

a) DTR0 — encapsulation

The running network is folded into a description. All nodes are grouped by similarity. For each group, invariants are extracted: shared sections, shared templates of connectors, shared logic. The result of the first step of convolution is CLSS — the catalogue of node classes. Then the patterns of connections between classes are extracted: which classes are connected with which, by what relations, with what density. The result of the second step is the complete MOVE — a compressed representation of the running subnetwork in the space of ontogeny.

DTR0 occurs rarely. In biology it is executed only at the formation of reproductive cells — to separate the DNA for transmission to descendants. In Gativus, DTR0 will be applied when new versions of MOVE are created from the analysis of successfully running subnetworks. In both cases, DTR0 loses part of the information: the current state, the contents of the trajectory logs, the concrete values of V-sections. What remains is the structural description. This is a general property of transformations — compression at the cost of detail loss.

b) RTR0 — morphogenesis

MOVE is unfolded into a running network. This is morphogenesis in the engineering sense — the process of germinating a seed into a complete network of subjective reality. RTR0 is implemented by two interrelated procedures: NRGN (generation of nodes from the description of classes) and SYGN (establishment of connections by templates and through autonomous search for partners). The detailed description of these procedures is in Chapter 4.

A principled property of RTR0 is that its result is not a copy of the original folded network, but a new instance. The nodes receive new UNONs, the initial states differ, the concrete SYGA topology is emergent and does not exactly reproduce the topology from which the MOVE was derived. Identical twins have identical DNA but are not the same as living organisms. The same holds for two instances unfolded from the same MOVE.

c) Asymmetry of temporal characteristics

GTR0 has an essential property that distinguishes it from the other three upper transformations — a pronounced asymmetry in the temporal characteristics of its direct and reverse directions. For GTR1, GTR2, GTR3, convolution and splice are completed almost simultaneously, within the same act of the neural network's operation. For RTR0, these two steps are separated in time: the creation of classes occurs relatively quickly, whereas the generation of concrete instances and the establishment of connections require considerably more time.

Between the two steps a special intermediate state exists: classes have already been determined but have not yet been instantiated. Its biological analogue is the stem cell — it has already fixed its specialization (becoming a neuroblast) but has not yet begun dividing into concrete neurons. This state is essential for the architecture of morphogenesis and is used in the splitting of MOVE among functional organs, as described in Chapter 5.

1. 6. Weights of morphotransformation and evolution

A question of principle: where does GTR0 obtain its concrete content? Knowledge of which classes to separate during convolution, which connection patterns to recognize as meaningful, which node templates to generate during unfolding — this knowledge cannot arise from nothing. It must already be in the system.

The answer relies on the distinction between two time horizons of learning. GTR1, GTR2, GTR3 learn through individual experience during the life of one organism. CNN1 is tuned to the sensory flow of a concrete animal, CNN2 to the language of a concrete culture, CNN3 to the narratives of a concrete biography. Each convolutional kernel is refined as experience accumulates; learning lasts throughout a lifetime.

The construction of GTR0 is different. Its convolution is not learned during the life of one organism. A concrete organism cannot retrain morphotransformation within the period of its life — it has neither the data nor the time. Morphotransformation is trained by evolution across many past generations on the time scale of billions of years. Its weights are the result of natural selection, accumulated in the form of inheritable structures.

a) Evolution as a training algorithm

Nature operates as the training algorithm of GTR0. The training data is organisms and their viability. Successful organisms are positive examples, unsuccessful ones negative. The objective function is viability and reproductive success. The correction of the weights is mutation of DNA, transmitted to descendants: beneficial ones are fixed, harmful ones eliminated. The speed of the process is slow by the scale of an individual life, but effective due to parallelism — billions of organisms simultaneously undergo selection.

The result is the accumulation of stable solutions — universal mechanisms of morphogenesis for cellular tissue, neural tissue, functional organs. Each such solution is a learned pattern of morphotransformation, a way of constructing a working structure from a description.

b) D-components as trained weights

In Gativus the weights of GTR0 are realized as D-components — object code compiled in ELF format, stored in the ROOT repository. A D-component is not just a program. It is the result of the evolutionary training of morphotransformation, packaged in executable form. The same MOVE with the same set of D-components gives the same organism — just as the same DNA gives identical twins.

For the Gativus project this implies the following. The biological genome was debugged by evolution over billions of years. Gativus does not have such time. The first working version of D-components must be created through reverse engineering — relying on an understanding of how subjective reality is constructed. GTOM provides the theory of the subjective. GNET provides the specification of network interaction. From these two, the structure of the necessary D-components can be derived: which types of nodes, which connectors, which templates of connections are needed.

This is possible only at a certain level of technical maturity. The existence of large language models is an argument in favour of having already reached this level. An LLM is a working engineering system that fulfils the function of one of the levels of subjective reality (the symbolic) and was created without an explicit understanding of what it reproduces. If the function of one level is engineering-reproducible, there is reason to think that the functions of the other levels are engineering-reproducible as well.

Differences between species in Gativus are differences between sets of D-components. Differences between individuals of the same species are differences between M-components — the MOVEs of individual organisms. Evolution changes D-components on the time scale of generations. A concrete organism unfolds an M-component within its own lifetime. Individuality does not arise in morphogenesis but in subsequent phases: in the training of CNN1, CNN2, CNN3 on individual experience and in the accumulation of the trajectory logs TRL5, TRL7, TRL9. Identical twins with the same morphogenesis become different people through different life trajectories.

1. 7. From programming to germination

The current approach to creating neural networks is programming. The programmer determines the architecture: number of layers, types of neurons, activation functions, connection patterns. Then this architecture is unfolded from code — from a sequence of instructions written by a human. The result is fully deterministic: the same program produces the same network. Every connection, every weight, every parameter is the result of a programmer's decision.

This approach is inherited from the von Neumann architecture — in which a computer is an executor of sequential instructions. Within this paradigm, a neural network is only a complex program. Its 'neurons' are variables. Its 'synapses' are matrix multiplications. Its 'learning' is optimization of a loss function through gradient descent. There is nothing in this that has any relation to how a biological brain is formed.

A biological brain is not programmed. It grows. No one writes code for the formation of the visual cortex. No one sets the matrix of connections between hippocampal neurons. There is a seed containing a description (the genome) and a minimal apparatus for its unfolding (cellular mechanisms). In suitable conditions, the seed begins to develop autonomously. The result is a network of billions of neurons and trillions of connections, capable of producing subjective reality.

Gativus stands on the principle: since nature has solved the problem of neural-network formation through germination, there is no need to invent an alternative. The task is to reproduce the natural mechanism on another substrate — not carbon-based biochemistry but silicon-based hardware. Not replacement, but transfer.

This determines the place of MOGE in the project. GTOM asserts that subjective reality is produced by the triple application of transformations. MOGE asserts that the network in which this triple application is realized is itself formed by a fourth transformation — morphogenesis — isomorphic to the first three. GNET determines the specifications that allow all this to be implemented on a concrete substrate. GATE is the device on which morphogenesis takes place.

The difference between programming and germination is not reducible to a technical one. These are two fundamentally different ways of building complex systems. A program is implemented through the explicit specification of every detail; germination, by placing a self-sufficient description in an environment with sufficient resources and starting an autonomous unfolding process. A program can be controlled step by step; germination only through input conditions. A program produces an exact copy of a given structure; germination produces structures of the same type, but unique in detail.

This difference is essential for Gativus. A subjective-reality network is too complex to be programmed node by node. Its quantitative parameters vary from individual to individual depending on available resources. Its concrete topology is emergent and cannot be specified in advance. All these properties follow naturally from germination — and are unattainable through programming.

1. 8. Structure of the book

This chapter has established the central concept — morphotransformation GTR0 as the fourth transformation of Gativus — and fixed its place in the overall architecture. The subsequent chapters unfold this concept into concrete engineering means.

Chapter 2 introduces MOLD — the formal language of morphology description. MOLD inherits its notation and basic set of diagrams from UML (which has been established by thirty years of engineering practice) and extends it with constructions specific to Gativus. MOVE is defined as a 16-dimensional coordinate vector in which each coordinate is a MOLD diagram. The six main diagrams — CLSS, COMP, COMM, ACTD, SPCE, RSRC — form the supporting set; the remaining ten are reserved for future extensions.

Chapter 3 considers GERM as the container of the seed and solves the question of the status of MOVE: is it a component of a node, a digital asset, or an entity of dual nature? The chapter also describes the role of the ROOT repository as a library of D-components, and the mechanism of their propagation across the network.

Chapter 4 describes RTR0 — the concrete procedure of morphogenesis. Atomic and composite units (MORN, MLOM), generation of nodes (NRGN), deterministic and autonomous synaptogenesis (SYGD, SYGA), parallelism of phases, transformation of OPNG into OPN, closure of the critical period.

Chapter 5 sets out the architecture of morphogenesis — the hierarchy ROOT → GATE → ANOD, the three-section structure of ANOD, the splitting of MOVE between functional organs. The chapter also fixes the boundary between MOGE and GNET: the subject of MOGE ends where the work of the running network begins.

The glossary at the end of the book contains all introduced terms with brief definitions. Readers familiar with GNSS and GTOM will find that a substantial part of the terms has already been introduced in the preceding books. MOGE adds to them the engineering constructions necessary for the transition to the GNET specifications.