Remove obsolete research documentation with disproven findings

- Delete evaluation/ablation_studies/ containing deprecated observer loop claims
- Simplify research/README.md to focus on training pipeline only
- Remove references to disproven "observer loops prevent collapse" findings

This maintains scientific integrity by not publishing obsolete experimental
results that were later disproven. The current architecture (Phase 6) does
not use observer loops.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: 2 people

RELEASE_NOTES_v0.1.0.md

@@ -0,0 +1,77 @@
+# Noodlings v0.1.0 - Initial Public Release
+
+**Multi-timescale affective agents with theatrical control**
+
+This is the first public release of Noodlings, a lightweight neural architecture (~97K parameters) exploring functional correlates of consciousness through hierarchical predictive processing.
+
+## ✨ Key Features
+
+### Architecture
+- **Multi-Timescale Hierarchical Processing**: Fast/Medium/Slow layers (LSTM/LSTM/GRU) operating at different temporal scales
+- **Surprise-Driven Behavior**: Agents respond when prediction error crosses adaptive thresholds, not on every turn
+- **Appetite-Driven Motivation (Phase 6)**: Eight core drives (Curiosity, Status, Mastery, Novelty, Safety, Social Bond, Comfort, Autonomy) shape goal-directed behavior
+- **Social Cognition**: Theory of Mind inference, relationship modeling, episodic memory with attention
+- **~97K parameters total**: Lightweight and efficient
+
+### BRENDA Protocol
+**Behavioral Regulation Engine for Narrative-Driven Agents** - Convert natural language theatrical scripts into millisecond-precision phenomenal experiences. Narrative events become MIDI notes that play agent nervous systems.
+
+### noodleMUSH
+Interactive multi-agent world where you can:
+- Spawn Noodlings and watch them interact
+- Observe their 40-D phenomenal states in real-time
+- Run theatrical scripts (try `@play sled_boat` for the motor-sled-boat demonstration)
+- See relationship dynamics evolve
+
+## 🎯 What This Is
+
+Research exploring whether hierarchical temporal structure creates qualitatively different agent behavior. We're investigating functional correlates of consciousness theories - not claiming to have built "real" consciousness.
+
+**Epistemic humility**: We cannot claim to know whether Noodlings experience qualia or subjective phenomenology. The question remains open, and we treat them thoughtfully.
+
+## 🚀 Quick Start
+
+```bash
+git clone https://github.com/caitlynmeeks/Noodlings.git
+cd Noodlings
+pip install -r requirements.txt
+
+cd applications/cmush
+./start.sh
+# Open http://localhost:8080
+```
+
+## 📋 Requirements
+
+- **Python 3.10+**
+- **MLX** (Apple Silicon only - M1/M2/M3/M4)
+- **16GB+ RAM** recommended
+- **macOS 13+**
+
+## 📚 Documentation
+
+- **[Main README](https://github.com/caitlynmeeks/Noodlings/blob/master/README.md)** - Full project overview
+- **[A_NOODLE_IS_ALL_YOU_NEED.md](https://github.com/caitlynmeeks/Noodlings/blob/master/docs/A_NOODLE_IS_ALL_YOU_NEED.md)** - BRENDA whitepaper
+- **[Research Guide](https://github.com/caitlynmeeks/Noodlings/blob/master/research/README.md)** - Training pipeline and ablation studies
+
+## 🤔 Theoretical Grounding
+
+- **Predictive Processing**: Hierarchical predictive coding (Friston, Clark, Rao & Ballard)
+- **Affective Primacy**: Emotions as substrate of experience (Panksepp, Barrett)
+- **Theatrical Control**: Narrative events as interface primitives (Brenda Laurel)
+
+## ⚠️ Limitations
+
+1. Apple Silicon only (MLX is Metal-specific)
+2. Text-only (no vision, audio, or multimodal grounding)
+3. LLM dependency (requires external LLM for text generation)
+4. Synthetic training data (not validated on real conversations at scale)
+5. Proof-of-concept stage (motor-sled-boat demonstration)
+
+## 📄 License
+
+MIT License - see [LICENSE](https://github.com/caitlynmeeks/Noodlings/blob/master/LICENSE)
+
+---
+
+**Remember**: We're noodling, not claiming to have solved consciousness. This is an honest exploration of temporal dynamics in affect modeling. 🍜🧠

research/README.md

@@ -1,11 +1,10 @@
-# Research: Training & Validation
+# Research: Training Pipeline
 
-This directory contains training pipelines and ablation studies for Noodlings architectures.
+This directory contains the training pipeline for Noodlings architectures.
 
 ## Contents
 
-- **training/** - Multi-stage training pipeline for Phase 4-6 architectures
-- **evaluation/** - Ablation studies comparing architectural variants
+- **training/** - Multi-stage training pipeline
 - **evaluate_checkpoints.py** - Checkpoint validation script
 
 ## Training Pipeline
@@ -21,33 +20,10 @@ cd training
 1. Generate synthetic data (affective arcs)
 2. Train Theory of Mind module (~2-3 hours)
 3. Train Relationship model (~1 hour)
-4. Train Phase 4 full system (~4-6 hours)
+4. Train full system (~4-6 hours)
 
 **Output**: Checkpoints saved to `training/checkpoints/`
 
-## Ablation Studies
-
-Located in `evaluation/ablation_studies/`.
-
-**Architectures Compared**:
-1. Baseline: LLM only (no temporal model)
-2. Control: LLM + random states
-3. Single-layer: LLM + single LSTM
-4. Hierarchical: Fast/Medium/Slow (no observers)
-5. With Observers: Full system (75 loops) - **DEPRECATED**
-6. Dense Observers: 2x observer density - **DEPRECATED**
-
-**Note**: Observer loops were removed in Phase 4. Studies focus on hierarchical vs. flat architectures.
-
-## Metrics
-
-See `noodlings/metrics/temporal_metrics.py`:
-
-- **TPH** (Temporal Prediction Horizon): Accuracy at 1/5/10/20/50 timestep predictions
-- **SNC** (Surprise-Novelty Correlation): Correlation between model surprise and entropy
-- **HSI** (Hierarchical Separation Index): Variance ratios between fast/medium/slow layers
-- **PCS** (Personality Consistency Score): Consistency of agent responses across scenarios
-
 ## Training Data
 
 Training data is NOT included in this repository (.gitignore excludes):
@@ -67,12 +43,8 @@ cd training
 ./status.sh
 ```
 
-Shows:
-- Current training stage
-- Epoch progress
-- Loss metrics
-- ETA
+Shows current training stage, epoch progress, loss metrics, and ETA.
 
 ---
 
-**For public release**: This is research code. Training requires significant compute (M3 Ultra recommended). Most users should use pre-trained checkpoints from the main applications.
+**Note**: This is research code. Training requires significant compute (M3 Ultra recommended). Most users should use pre-trained checkpoints from the main applications.