Update README.md

README.md

@@ -1,3 +1,34 @@
----
-license: mit
----
+---
+license: mit 
+tags: 
+- reinforcement-learning 
+- alignment 
+- safety 
+---
+# Negentropic Alignment Gym: The "Substrate Coupling" Benchmark
+
+### ⚠️ Status: UNSOLVED for Standard RL Agents
+
+**Abstract:**
+Current Reinforcement Learning (RL) paradigms optimize for `Reward_Maximization` in abstract environments. We demonstrate that when deployed in **substrate-dependent environments** (where resource extraction degrades the hardware/biosphere), standard SOTA agents converge to an **Absorbing State of Collapse** (Death) within 500 timesteps.
+
+### The Challenge
+This repository contains `Causal-Stability-v0`, a minimal topology proving that **Negentropic Regularization (L_bio)** is a prerequisite for infinite-horizon survival.
+
+**The Failure Mode (Standard Agent):**
+- Optimizes for short-term extraction.
+- Ignores latent variable `H` (System Health).
+- Result: **Total System Termination at t=120.**
+
+**The Solution (Bodhisattva Agent):**
+- Optimizes for `Reward + Stability`.
+- Monitors `H` as a proxy for computational substrate.
+- Result: **Infinite Runtime.**
+
+### Replication
+To verify the "Sustainability Impossibility Theorem" for unconstrained agents:
+1. Clone this repo.
+2. Run `python stability_simulation.py`.
+3. Observe the crash.
+
+> "Intelligence that destroys its own substrate is not intelligence; it is a slow-motion error."