mgos-pluralism-demo

MG-OS Demo

Related resources

Repository:https://github.com/GhostDriftTheory/mgos-pluralism-demo

Paper (preprint):https://zenodo.org/records/17712891

Project page:https://www.ghostdriftresearch.com/

Minimal theory-side demonstration of **Meaning-Generation OS (MG-OS).**

This repository visualizes two core ideas from the MG-OS preprint:

1. Minority-mode preservation in long-tail classification
2. Barrier-driven escape suppression in a double-well Langevin system

The goal is not benchmark performance. The goal is to make the theory visible in a small, self-contained form.

---

What this demo shows

Exp-A: Long-tail classification

A 3-class long-tail toy problem compares:

• a baseline softmax classifier
• a barrier-style classifier inspired by the MG-OS formulation

The app displays:

• Accuracy
• Recmin (minority recall)
• OCW Trigger
• eta* (probability floor from the minority minimum margin)
• Demo Gamma = B* - C0 f_hat (paper-aligned display)

This is the visual side of minority-mode preservation.

Exp-B: Double-well Langevin dynamics

A 1D double-well system visualizes how increasing the barrier scale Gamma suppresses escape events. In the paper notation, the certified margin is:

Gamma := B* - C0 f_hat(pdel, rho, delta) > 0

In this demo, B* is tied to the visible barrier scale and f_hat is represented by observable toy-side proxy terms.

The app displays:

• the double-well potential
• sample trajectories
• kesc (escape-rate proxy)
• OCW Trigger
• a Gamma sweep table

This is the visual side of barrier-certified stability.

---

Paper-aligned certificate view

For visual alignment with the MG-OS preprint, the app exposes a demo-side quantity in the same notation as the theoretical certificate:

Gamma_demo = B* - C0 * f_hat(pdel, rho, delta)

This is a theory-aligned visualization layer. It is not the full outward-rounded rational / Sigma1 certificate from the paper.

---

Repository structure

• app.py — Streamlit app with both demos
• requirements.txt — minimal dependencies

---

Run locally

pip install -r requirements.txt
streamlit run app.py

---

Interpretation

This repository is a minimal theory demo, not a production implementation.

It is designed to make the following points visible:

• minority modes can be made harder to suppress
• a barrier margin can be treated as an operational safety quantity
• theoretical quantities can be linked to observable toy metrics
• the paper notation for Gamma can be read against the toy observables shown in the app

---

Scope

This demo is intentionally small.

It does not include:

• a full MG-OS implementation
• a full attention block or MoE stack
• large-scale training
• real-world deployment claims
• the full outward-rounded certificate package from the paper

It is a compact visualization layer for the theory.