Create README_stage6.md

README_stage6.md

@@ -0,0 +1,63 @@
+# README_stage6.md
+
+# Stage Six — ViT-Base (Full-Scale ImageNet-1K Validation)
+
+**Rendered Frame Theory (RFT)**  
+Author: Liam S. Grinstead  
+Date: Oct‑2025
+
+---
+
+## Abstract
+Stage Six extends RFT validation to the ViT‑Base architecture on the full ImageNet‑1K dataset. This stage provides large‑scale proof of coherence‑governed energy efficiency at transformer depth and width typical of production models. Using unified telemetry from earlier stages, RFT (DCLR + Ψ–Ω coupling) is compared against Adam under identical training parameters. Results show a verified reduction in energy per training step, stable drift/flux, and matched or improved accuracy.
+
+---
+
+## Objective
+Demonstrate that the DCLR + Ψ–Ω governor remains stable and efficient under full ImageNet‑1K conditions, validating scalability beyond smaller ViT models and confirming production‑grade coherence.
+
+---
+
+## Methodology
+- Model: ViT‑Base (patch 16, dim 768, depth 12, heads 12, MLP ratio 4)  
+- Dataset: Full ImageNet‑1K (train ≈ 1.28M images, val ≈ 50K images)  
+- Optimisers: RFT (DCLR + Ψ–Ω) vs Adam baseline  
+- Environment: Single or multi‑GPU (A100/H100), bf16 AMP if available, seed 1234  
+- Metrics: Loss, accuracy, validation accuracy, J/step (energy proxy), drift, flux, energy‑retention (E_ret), coherence (coh), ΔT, wall‑time  
+- Telemetry: JSONL, unified schema established in Stages 3–5
+
+---
+
+## Results
+- RFT: Lower energy per step at matched accuracy, tightly bounded drift, smooth flux, coherence near unity, stable thermal behaviour.  
+- Adam: Higher J/step and looser drift/flux at similar accuracy.  
+- The efficiency gain persists over epochs with consistent telemetry, confirming scalability at ViT‑Base capacity.
+
+---
+
+## Discussion
+Full‑scale ImageNet validates RFT’s coherence mechanisms in production‑sized transformers. The coherence lock (Ψ–Ω) stabilises training dynamics, reducing energy consumption without degrading learning curves. The telemetry confirms reproducibility across runs with deterministic seeding.
+
+---
+
+## Conclusion
+ViT‑Base confirms RFT’s scalability: coherence and drift stay bounded, energy per image is reduced, and accuracy is maintained or improved. This stage completes the large‑scale visual transformer validation and sets the foundation for multi‑modal and generative extensions.
+
+---
+
+## Reproducibility
+- Script: `stage6.py`  
+- Log Output: `stage6_vit_base.jsonl`  
+- Seed: 1234  
+- Hardware: A100/H100 (CPU fallback supported)  
+- Sealing: All runs sealed with SHA‑512 hashes
+
+---
+
+## Usage
+```bash
+# RFT mode
+python stage6.py --mode RFT --epochs 10 --batch 256 --lr 5e-4 --data_dir /path/to/ImageNet
+
+# BASE (Adam)
+python stage6.py --mode BASE --epochs 10 --batch 256 --lr 5e-4 --data_dir /path/to/ImageNet