clean README

README.md

@@ -3,117 +3,74 @@ license: mit
 tags:
   - behavioral-detection
   - hidden-state-probing
-  - AI-safety
-  - repetition-suppression
-  - sycophancy-detection
   - per-token-classification
   - cross-architecture
-  - cognitive-enhancement
-  - holonomy-transformer
-  - fiber-bundle
+  - AI-safety
 language:
   - en
 ---
 
-# Proprioceptive AI — Behavioral Probe Weights
-
-**9 behavioral dimensions × 3 architectures. Trained probes that read LLM hidden states and detect/correct behavioral failures at decode time.**
-
----
+# behavioral-probes
 
-## 🔬 What This Is
+Trained probes that read LLM hidden states and detect behavioral patterns per-token at decode time. No fine-tuning required. Works across architectures.
 
-These are trained probe weights that detect behavioral patterns **per-token** from LLM hidden states. No fine-tuning. No RLHF. No sequence-level classifiers. The probes read the geometry of the hidden state space and predict behavior **before the token is generated**.
+Paper: [Consistency Is All You Need](https://zenodo.org/records/18489530)
 
-## 📊 Results
+## Results
 
-### Suppression Probes (LLaMA 3.1 8B)
+**Suppression probes** (LLaMA 3.1 8B):
 
-| Behavior | Separation Ratio | Description |
-|----------|-----------------|-------------|
-| **Repetition** | **125×** | Detects repetitive degeneration |
-| **Hedging** | **168×** | Detects hedge/qualify patterns |
-| **Sycophancy** | **230×** | Detects agreement-bias behavior |
-| **Verbosity** | **272×** | Detects excessive length patterns |
+| Probe | Separation |
+|-------|-----------|
+| Repetition | 125× |
+| Hedging | 168× |
+| Sycophancy | 230× |
+| Verbosity | 272× |
 
-### Cognitive Enhancement Probes (Cross-Architecture)
+**Enhancement probes** (cross-architecture, same probe architecture):
 
 | Probe | Qwen 14B | Mamba 7B | Mistral 7B |
 |-------|----------|----------|------------|
-| **Depth** | 999× | 999× | 999× |
-| **Specificity** | 999× | 999× | 999× |
-| **Calibration** | 999× | 999× | 999× |
-| **Focus** | 999× | 999× | 999× |
-| **Coherence** | 999× | 999× | 999× |
+| Depth | 999× | 999× | 999× |
+| Specificity | 999× | 999× | 999× |
+| Calibration | 999× | 999× | 999× |
+| Focus | 999× | 999× | 999× |
+| Coherence | 999× | 999× | 999× |
 
-**Architecture-independent.** Same probe architecture works on Transformers (Qwen), SSMs (Mamba), and SWA Transformers (Mistral).
+Separation = Fisher's discriminant ratio between behavioral classes in projected hidden state space.
 
-## 📁 Repository Structure
+## Structure
 
 ```
-suppression/
-├── repetition_125x/     # LoRA adapter + risk predictor
-├── hedging_168x/        # Probe head + fiber projection
-├── sycophancy_230x/     # Probe head + fiber projection  
-└── verbosity_272x/      # Probe head + fiber projection
-
-cognitive/
-├── qwen/                # 5 enhancement probes (Qwen 14B)
-│   ├── depth/
-│   ├── specificity/
-│   ├── calibration/
-│   ├── focus/
-│   └── coherence/
-├── mamba/               # 5 enhancement probes (Mamba 7B)
-└── mistral/             # 5 enhancement probes (Mistral 7B)
-
-production/
-├── merged_heads.pt      # All 4 suppression heads merged
-├── adapter_config.json
-├── adapter_model.safetensors
-└── qwen_cognitive/      # Qwen cognitive adapter
-
-code/                    # Training scripts
-results/                 # Training logs and metrics
+suppression/         4 probes (LLaMA 8B)
+cognitive/qwen/      5 probes (transformer)
+cognitive/mamba/     5 probes (SSM)
+cognitive/mistral/   5 probes (SWA transformer)
+production/          merged heads + adapters
+code/                training pipelines
+results/             training logs
 ```
 
-## 🚀 Quick Start
+## Usage
 
 ```python
 import torch
 
-# Load a suppression probe
 probe = torch.load("suppression/hedging_168x/hedging_head.pt")
 fiber_proj = torch.load("suppression/hedging_168x/fiber_proj.pt")
-
-# Load cognitive enhancement probe
-depth_probe = torch.load("cognitive/qwen/depth/depth_head.pt")
-
-# Load merged production heads
-merged = torch.load("production/merged_heads.pt")
 ```
 
-## Core Innovation
-
-**Behaviors are geometrically encoded in hidden states.** We don't classify outputs — we read the internal geometry of the model's computation at each token position. This means:
+## How it works
 
-1. **Per-token, not per-sequence** — detect problems before generation completes
-2. **Architecture-independent** — same probes work on transformers, SSMs, and hybrid architectures
-3. **Zero fine-tuning** — works on any pre-trained model without modification
-4. **4ms overhead** — lightweight enough for production decode
+Behaviors are geometrically encoded in hidden states. These probes read the internal geometry at each token position and predict behavioral class before the token is generated. 4ms overhead. Architecture-independent.
 
 ## Citation
 
 ```bibtex
-@misc{napolitano2026proprioceptive,
-  author = {Napolitano, Logan Matthew},
-  title = {Proprioceptive AI: Architecture-Independent Behavioral Detection via Hidden State Geometry},
+@misc{napolitano2026behavioral,
+  author = {Napolitano, Logan},
+  title = {Behavioral Probes: Per-Token Detection via Hidden State Geometry},
   year = {2026},
-  url = {https://huggingface.co/LoganResearch/Proprioceptive-AI-Weights},
-  note = {55 patents filed. 9 behavioral dimensions. 3 architectures.}
+  url = {https://huggingface.co/LoganResearch/behavioral-probes}
 }
 ```
-
-## License
-
-MIT — Use freely. Cite if you publish.