# Stage Four — ViT-Tiny (ImageNet Subset) Validation

**Rendered Frame Theory (RFT)**  
Author: Liam S. Grinstead  
Date: Oct‑2025

---

## Abstract
Stage Four evaluates RFT on a transformer vision model (ViT‑Tiny), extending beyond CNN baselines. Using ImageNet‑1K subsets (or synthetic stand‑ins), we compare RFT (DCLR + Ψ–Ω coupling) against Adam. Results show lower energy per step and tighter drift/flux stability at matched accuracy, demonstrating that coherence‑aware optimisation benefits attention layers as well as convolutions.

---

## Objective
Verify that RFT’s efficiency and coherence gains persist in transformer architectures by benchmarking ViT‑Tiny on an ImageNet‑like workload with identical hyperparameters across modes and unified telemetry.

---

## Methodology
- **Model:** ViT‑Tiny (patch 16, dim 192, depth 12, heads 3) with CLS token  
- **Data:** ImageNet‑subset via ImageFolder (train/val), or synthetic images/labels if no `--data_dir` supplied  
- **Setup:** PyTorch 2.x, Python 3.10, A100/H100 (BF16 autocast if available), seed 1234  
- **Metrics:** Loss, accuracy, J/step (NVML if present; proxy otherwise), drift, flux, energy‑retention (E_ret), coherence (coh), ΔT  
- **Parity:** Same batch, learning rate, and steps across RFT and BASE  
- **Orbital Coupler:** Ψ–Ω drift/flux synchronisation each iteration  
- **Optimisers:** DCLR (RFT) vs Adam (BASE)

---

## Results
- **RFT (DCLR + Ψ–Ω):** Lower energy‑per‑step at comparable loss/accuracy; drift remains tightly bounded, flux stable; coherence near unity throughout the run.  
- **Baseline (Adam):** Higher J/step and looser drift/flux behaviour at matched settings.  
- **Synthetic data fallback:** Reproduces the same qualitative efficiency pattern, indicating optimiser–telemetry interaction is the source of gains, not dataset artefacts.

---

## Discussion
ViT‑Tiny confirms that RFT’s coherence mechanisms generalise to attention‑based architectures. The reduction in J/step with stable drift/flux mirrors Stage One (CNN behaviour) and Stage Two (multi‑agent dynamical stability), indicating architecture‑agnostic benefits driven by coherence‑aware optimisation.

---

## Conclusion
RFT scales cleanly to transformers: energy efficiency improves without compromising learning signals, and coherence remains robust across iterations. This stage prepares for deeper vision models (ViT‑Small, ViT‑Base) and multi‑modal settings.

---

## Reproducibility
- **Script:** `stage4.py`  
- **Log Output:** `stage4_vit_tiny.jsonl`  
- **Seed:** 1234  
- **Hardware:** A100/H100 (CPU fallback supported)  
- **Sealing:** All runs are sealed with SHA‑512 hashes

---

## Usage
- **RFT mode:**  
  `python stage4.py --mode RFT --steps 1000 --batch 256 --lr 5e-4 --data_dir /path/to/imagenet_subset`
- **BASE mode (Adam):**  
  `python stage4.py --mode BASE --steps 1000 --batch 256 --lr 5e-4 --data_dir /path/to/imagenet_subset`
- **Synthetic fallback:** Omit `--data_dir` to auto‑generate data.