KADLOG R&D · AI Genetics
AI Genetics is a research and experimental program testing whether persistent, configurable determinants of LLM-based AI systems can be organized in a genetics-like cybernetic representation that adds measurable explanatory or predictive value.
Research architecture
The project deliberately separates the empirical study of configurable LLM systems from the mathematical theory of viability-preserving regulation. Their connection is a research object in its own right.
Study persistent instructions, memory, retrieval, tools, agent policies and other controllable determinants under standardized environments, repeated executions and measurable behavioural metrics.
Model an artificial cognitive system through a state space, a viability domain, a cybernetic genome of regulatory operators and the regulatory field generated by their activation.
Formal mathematical foundation
CGF v1.0 defines persistence as a dynamical property of trajectories maintained inside a viability domain by regulatory fields.
Candidate operational mapping
A single output is not a phenotype. In AI Genetics, phenotype refers to a stable or statistically characterized behavioural property observed across controlled repetitions.
CGF Explorer
The Explorer now separates two uses: a numerical viability demonstrator derived from the minimal CGF model, and a controlled-experiment designer for the LLM research track. The second builds a protocol; it does not fabricate behavioural outcomes.
The displayed κH proxy is defined in this demonstrator as the loss of projected terminal viability margin over horizon H after removing one regulator. It is related to the paper’s criticality idea but is not presented as an empirical measurement or as the final CGF definition.
Generated controlled protocol
Experimental research interface. Numerical values in the viability demonstrator are illustrative parameters of a tractable minimal model, not measurements from a production LLM. The current project does not claim biological equivalence with DNA, direct correspondence with neural-network weights, consciousness, sentience or a validated general theory of intelligence.
Falsifiable numerical program
These experiments belong to the minimal mathematical model. Their role is to test viability preservation and the emergence of regulatory criticality before broader claims are made.
ẋ = −Dx
Baseline without genomic regulation: test whether trajectories decay toward the non-viable region.ẋ = −Dx + ΦΓ(x)
Test whether regulatory fields can maintain trajectories inside the viability domain.Γ(−i) = Γ \ {gi}
Remove one regulator at a time and quantify the effect on viability.κi
Test whether a minimal viability-preserving core emerges rather than being imposed a priori.Scientific status
This distinction is part of the research method: the website should evolve only when the project produces a new definition, protocol, result or falsification.
Quadruplet C = (X,V,Γ,ΦΓ), dynamic law, viability domain, persistence attractors, criticality concept and an initial falsifiable numerical program.
Define the candidate Layer-I regulators NDX1, ENR, PRSV, ACC2 and PHY0 through target variables, activation functions, regulatory vector fields and measurable contributions to viability; then execute the numerical program.
Determine which persistent AI determinants, if any, deserve CGF gene/genotype status and whether that representation improves prediction beyond simpler configuration descriptions.
Research progression
Progression is evidence-driven, not calendar-driven.