Jneopallium – Biologically-grounded Java framework for natural neuron networks (safety-first autonomous AI)

Research Awesome paper
1

Hi everyone,I’m Dmytro Rakovskyi, independent researcher. For the last few months I’ve been developing Jneopallium — an open-source Java framework for modeling biologically realistic neuron networks at customizable levels of detail.Core features:

  • Typed signals with explicit fast/slow loop frequencies (bioelectric + neuromodulatory timescales)

  • Multi-receptor neurons (each neuron can implement multiple interfaces with dedicated processors)

  • Full autonomous-AI architecture with 28 neuron classes, including:

    • Harm discriminator (consequence-model safety gate with asymmetric caution learning and hard ethical invariants)

    • Loop-prevention subsystem (detects and gently breaks runaway cycles without permanent damage)

    • Embodiment, affect, curiosity, glia, sleep, and working memory modules

  • Optional non-blocking LLM integration with strict verification

  • Designed for real-world deployment (JVM + planned FPGA/gRPC backend)

The goal is to build safe, interpretable, biologically-plausible autonomous systems for robotics, BCI, industrial control, and clinical decision support.I’m looking for:

  • People interested in using Jneopallium in their projects

  • Collaborators (code, testing, new modules, hardware integration)

  • Feedback from the SNN / neuromorphic / AI-safety / embodied-AI communities

Repo: https://github.com/rakovpublic/jneopallium would love to hear your thoughts or discuss possible collaboration. Even simple feedback on architecture or use-cases is very welcome.Thank you!
Dmytro Rakovskyi

2

This is a fascinating architecture, especially the implementation of the Harm Discriminator and Anti-looping subsystem. In my current projects on Linux-based environments, I’ve found that preventing infinite recursive loops in autonomous decision-making is one of the most critical challenges when deploying 8b models.

Your approach to multi-receptor neurons with dedicated processors is a brilliant way to handle neuromodulatory scales. I’m particularly interested in how your non-blocking LLM integration maintains strict verification. In my experience, ensuring safety invariants while keeping orchestration efficient is where most frameworks struggle.

The planned FPGA/gRPC backend sounds promising for industrial and clinical control. I’d be interested to see how your “asymmetric caution learning” performs when synchronized with real-world medical or legal datasets, where the cost of a false negative is extremely high.

Impressive work on Jneopallium!