This repository contains Anomaly Soul Kit, an open simulation framework for observing the emergence, persistence, and evolutionary inheritance of anomalous behavior in populations of LLM-driven agents. Each agent encodes a numeric internal state — vitality (H) and anomaly intensity (Z) — and expresses that state through LLM-generated text each generation. A detection layer scores each expression against the population across three axes: lexical divergence, structural divergence, and novel vocabulary. Agents whose expressions deviate from the population accumulate anomaly intensity, which feeds back into their fitness and is heritable across generations. The project does not claim these anomalies constitute mind or soul. It provides a reproducible kit for observing whether something — a persistent, evolving deviation — reliably emerges from this process, and what it looks like when it does. --- Update — February 2026 v2 of the anomaly detection layer has been released. Two structural issues identified in early testing have been addressed. First, anomaly score inflation: as the population evolved, an increasing proportion of agents were flagged as anomalous, eventually making the designation meaningless. This has been resolved by replacing absolute scoring with a dynamic baseline — scores are now normalized relative to the population median each generation, making it structurally impossible for the entire population to simultaneously score as anomalous. Second, convergence speed: the original selection pressure caused Z-awakening to saturate too quickly (~90% by generation 50). Scaling has been adjusted to allow slower, more observable divergence dynamics. Two new observational metrics have been added: new_normal_threshold tracks whether what was previously anomalous is becoming the new collective norm, and population_drift measures how much the group as a whole is shifting toward anomalous expression across generations.