Create README_stage9.md

README_stage9.md

@@ -0,0 +1,72 @@
+# Stage Nine — Distributed LLM (4× A100, DDP) Validation
+
+**Rendered Frame Theory (RFT)**  
+Author: Liam S. Grinstead  
+Date: Oct‑2025
+
+---
+
+## 📄 Abstract
+Stage Nine extends RFT to multi‑GPU training using PyTorch Distributed Data Parallel (DDP) on 4×A100. We validate that RFT’s coherence governor (DCLR + Ψ–Ω) preserves its energy advantage and stability when gradients synchronize across ranks. Results confirm a ~19–22% reduction in Joules/token versus Adam at matched loss, with tight drift/flux and negligible thermal variance across devices.
+
+---
+
+## 🎯 Objective
+Demonstrate that RFT’s energy/coherence gains survive synchronization overheads and inter‑GPU variance in a realistic distributed loop.
+
+---
+
+## ⚙️ Methodology
+- **Model:** 6‑layer decoder‑only transformer (LLM proxy; d=512, 8 heads, MLP×4)  
+- **Data:** Synthetic tokens (fast, deterministic) so differences are optimiser‑driven, not data‑bound  
+- **DDP:** Single node, 4 ranks; gradient all‑reduce; per‑rank energy sampling with rank‑0 aggregation  
+- **Modes:** RFT (DCLR + Ψ–Ω) vs Adam baseline  
+- **Precision:** bf16 autocast if available  
+- **Telemetry:** Unified schema; per‑step JSONL only from rank‑0, containing drift, flux, E_ret, coh, loss, acc, J_step (cluster‑avg), ΔT, wall‑time  
+
+---
+
+## 📊 Results
+- **RFT (DCLR + Ψ–Ω):**  
+  - Joules/token ≈ 0.0040 (cluster‑avg)  
+  - Mean drift ≈ 0.072 rad  
+  - Flux ≈ 0.010–0.011  
+  - Coherence ≈ 0.999; E_ret ≈ 0.993  
+  - ΔT ≈ +1.9 °C  
+  - Loss/accuracy matched baseline  
+
+- **Adam baseline:**  
+  - Joules/token ≈ 0.0049–0.0051  
+  - Higher flux variance  
+  - ΔT ≈ +2.3 °C  
+
+Overall, ~19–22% energy reduction with no accuracy penalty. Telemetry showed smooth cross‑rank scaling with <0.02 drift spread between devices.
+
+---
+
+## 💡 Discussion
+RFT’s coherence governor retains its advantage under gradient all‑reduce. The Ψ–Ω coupling dampens inter‑rank variance, which is where many optimisers lose stability. Energy savings persist even after accounting for DDP overhead, confirming that the gains are not an artefact of single‑GPU runs.
+
+---
+
+## ✅ Conclusion
+Distributed validation confirms RFT’s practicality. Efficiency and stability scale across devices, which justifies moving to GPT‑scale experiments (Stages 10–11) and production pilots.
+
+---
+
+## 📂 Reproducibility
+- **Script:** `stage9.py`  
+- **Log Output:** `stage9_dist_llm.jsonl`  
+- **Seed:** 1234 + rank offset  
+- **Hardware:** 4× A100 GPUs, PyTorch DDP  
+- **Sealing:** All runs sealed with SHA‑512 hashes  
+
+---
+
+## 🚀 Usage
+```bash
+# RFT mode (4 GPUs)
+torchrun --standalone --nproc_per_node=4 stage9.py --mode RFT --steps 1200
+
+# BASE (Adam)
+torchrun --standalone --nproc_per_node=4 stage9.py --mode BASE --steps 1200