Machine Zygote: Causal Biparental Heredity Before Learning in a Germline--Soma Artificial Agent

Technical Summary: Machine Zygote: Causal Biparental Heredity Before Learning in a Germline–Soma Artificial Agent

Problem Statement
While artificial life and evolutionary robotics have established research directions in developmental encodings, robot reproduction, and inherited controllers, a specific experimental question remains unresolved: Can a newborn artificial agent exhibit measurable biparental heredity before any learning occurs, and can this dependence be isolated causally rather than merely inferred from parent–offspring resemblance? Existing systems often conflate inherited states with post-birth adaptation or rely on generative mappings that do not isolate which parental contribution causes specific phenotypic changes under matched backgrounds. This study addresses the need to separate heredity, development, stochastic variation, and learning in a controlled computational environment.

Methodology
The study introduces Machine Zygote, a computational architecture designed to operationalize a germline–soma separation. The system is defined by the following components:

• Germline Representation: The germline ($G$) is a normalized 102-dimensional vector. It includes 99 autosomal loci and 3 dam-specific loci (analogous to cytoplasmic contributions). Four fixed founders (A–D) are generated from a Sylvester Hadamard construction to ensure distinct but norm-matched parental states.
• Recombination and Zygote Formation: Two parental germlines are independently mutated ($\xi \sim N(0, \sigma^2_{mut})$) and recombined into a zygote ($Z$). The recombination mask, mutation vectors, and developmental noise seeds are stored as a "background" to enable matched-background interventions.
• Developmental Regulation: The zygote parameterizes a developmental regulatory system acting on an initially generic eight-module soma ($U$). The development is governed by a differential equation involving a regulatory matrix ($W$), decay constants, biases, and morphogen sensitivities. The system integrates Euler–Maruyama over a fixed duration.
• Soma and Phenotype Measurement: The resulting soma is frozen immediately after development. A fixed, non-heritable readout maps the frozen regulatory states to an operational controller for a differential-drive body abstraction.
• Experimental Design:
  • H1 (Diallel): A preregistered $4 \times 4$ balanced diallel comprising 640 offspring (40 per cell) to test for factorial parental dependence.
  • H2 (Developmental Necessity): A quasistatic ablation removes recurrence, diffusion, and noise to test if dynamic development is necessary for biparental structure.
  • H3 (Pre-learning): All evaluations occur with a fixed controller; no learning, plasticity, or adaptation is permitted.
  • H4 (Transgression): Tests if recombination produces offspring exceeding parental clone-line bounds.
  • H5 (Causal Intervention): The strongest test, involving 60 matched-background interventions where only one parental germline is substituted while holding the recombination mask, mutation vectors, and noise seed fixed.
• Metrics: Six newborn behavioral traits are measured: mean speed, dominant gait frequency, signed turning bias, perturbation recovery time, oscillator coherence, and maximum exploration radius.

Key Results

1. Biparental Dependence (H1): In the $4 \times 4$ diallel, both dam and sire identities significantly influence five of six newborn traits (speed, gait frequency, turning bias, coherence, and exploration radius) after Holm correction. Recovery time showed weak parental main effects. The modeled parental variance fraction (PDVF) ranged from 36% to 53% for the significant traits.
2. Causal Intervention (H5): In matched-background substitutions, replacing a single parental germline caused phenotype shifts that significantly exceeded the variation induced by same-parent re-mutation or re-development controls for five traits. This confirms that parent identity is not merely correlated with behavior but causally determines it under fixed stochastic conditions.
3. Transgressive Offspring (H4): Recombination produced a statistically significant excess of transgressive offspring for speed (15.0% vs. 1.2% in clones) and gait frequency (17.5% vs. 0%), indicating that recombination can generate novel phenotypes beyond parental bounds.
4. Developmental Dynamics (H2): The hypothesis that recurrent developmental dynamics are necessary for biparental heredity was not supported. A quasistatic ablation (removing dynamics) preserved the mean multivariate phenotype distribution (centroid difference not significant) but altered the parental variance structure. This suggests that while dynamics modulate variance, they are not strictly required to generate the mean biparental phenotype in this architecture.
5. Robustness: The primary factorial effects were recovered across a sweep of 20 parameter points (coupling scale, duration, noise, mutation amplitude).

Significance and Claims
The paper explicitly frames its contribution as methodological and architectural rather than a claim of biological equivalence or physical heredity.

• Causal Biparental Pre-Learning Heredity: The study establishes that in this specific simulation, two parental germline channels jointly determine a newborn behavioral phenotype before any learning occurs. The causal link is demonstrated through matched-background interventions where changing one parent's germline shifts the phenotype.
• Intervention-Centered Framework: The primary value is a reproducible benchmark that separates heredity, development, stochastic variation, and post-birth learning. It provides a formalization for testing causal heredity in artificial agents using a balanced diallel and explicit germline–soma separation.
• Modest Scope: The author clarifies that this is a computational study using a differential-drive abstraction, not a physical robot. The term "Machine Zygote" refers to the computational object formed by recombination, not a claim of biological embryogeny. The study does not establish physical heredity, biological genetics, or autonomous evolution.
• Refutation of Developmental Necessity: A significant negative result is that the specific recurrent dynamics tested are not necessary for the mean biparental phenotype, challenging the assumption that complex ontogeny is a prerequisite for inherited structure in this context.

The paper concludes that while it does not invent artificial embryogeny, it provides a stricter, intervention-centered starting point for future studies on transgenerational memory and the evolution of developmental processes in artificial agents.