Invictus: A Lamarckian-Inspired Parent–Child Framework for Runtime Behavioral Inheritance in Neural Language Models

This paper introduces Invictus, a Lamarckian-inspired parent–child framework for studying runtime behavioral inheritance in neural language models. The system creates a developmental loop in which a parent model observes the deployed conversational behavior of a child model, generates targeted synthetic training examples, and produces successive child generations through LoRA-based adaptation. Across a seven-phase experimental protocol using Qwen2.5-1.5B-Instruct with LoRA adapters, Invictus reports recurrent adaptation fluctuation under runtime-feedback mentoring, a curriculum-driven token reduction effect, targeted curriculum efficiency, epigenetic context quality effects, and performative compliance under adversarial instructions. The paper does not claim biological inheritance. It proposes a computational analogue in which runtime-expressed behavior is transformed into training data that measurably influences later model generations. The framework opens a research direction in developmental AI, adaptive local AI systems, behavioral inheritance, and parent–child model adaptation.