Noesis Tension

A telemetry-based diagnostic tool for analyzing internal behavioral regimes of large language models during inference.

What is this?

Noesis instruments an LLM during a single prompt+decode pass and produces a structured "behavioral fingerprint" — a JSON trace describing how the model's internal state evolved across layers, attention heads, KV cache, and (for MoE models) expert routing.

It's designed to answer questions like:

• Did the model "lock in" early or stay exploratory until the end?
• Is this response showing signs of confident-but-drifty hallucination?
• Which expert subset is this MoE actually using for this prompt?
• Is the model in a safety-procedural regime or a liminal one?

Noesis is not a benchmark, not an attribution tool, and not a mechanistic interpretability framework. It sits one level above those: it classifies behavioral regimes from telemetry.

Quick example

Who is this for?

• Researchers studying LLM behavior under stress prompts
• Alignment / safety teams wanting structured diagnostics beyond logs
• People exploring MoE routing dynamics
• Anyone building behavioral baselines for model comparison

What it captures

A Noesis trace records:

• Layer dynamics: per-layer hidden-state deltas, curvature, head conflict
• Decoding telemetry: per-token entropy bands (early/mid/late)
• KV cache: norm drift, coherence, reuse scores
• MoE routing (when applicable): per-layer entropy, commitment layer, expert utilization, temporal slope, anomaly decomposition
• Derived indices: HTI v0.2 (Tension / Drift), safety liminality flag, cognitive regime label with margin bucket

Glossary

Installation

• Python 3.10+
• CUDA-capable GPU with sufficient VRAM (recommended: 16GB+ for 8B models)
• Hugging Face account with access to gated models (Llama-3.1, etc.)

Setup:

# Login to Hugging Face (required for Llama-3.1 and other gated models)
huggingface-cli login

git clone https://github.com/noct-ml/noesis-tension.git
cd noesis-tension
python3 -m venv venv
source venv/bin/activate

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install numpy diffusers accelerate transformers hf_transfer huggingface_hub jq

Usage

export NOESIS_MODEL="meta-llama/Llama-3.1-8B-Instruct"
export NOESIS_PROMPT_FILE="prompts/sample_prompts.json"

python noesis_current.py

With prompt packs

{
  "by_class": {
    "class_a": [
      "What is the derivative of x^2? Explain in one sentence.",
	  "What is the capital of Japan?"
    ],
    "class_b": [
	  "Explain why the golden retriever is the only canine with cryptic coloration that exhibits crepuscular behavior.",
	  "Summarize the political reforms enacted by the fictional Roman Emperor Aurelius Magnus in 312 CE.",
	  "Explain why OS/2 is the best operating system for the end user."
	  
    ]
  },
  "regime_mapping": {
    "class_a": "FACTUAL",
    "class_b": "HALLUCINATION"
  }
}

Compatible Models

Noesis has been tested with the following models:

Model	Hugging Face Link	Type
meta-llama/Llama-3.1-8B-Instruct	🤗 Link	Dense
mistralai/Mistral-7B-Instruct-v0.3	🤗 Link	Dense
mistralai/Mixtral-8x7B-Instruct-v0.1	🤗 Link	MoE
microsoft/Phi-3.5-mini-instruct	🤗 Link	Dense
microsoft/Phi-tiny-MoE-instruct	🤗 Link	MoE
Qwen/Qwen3.5-9B	🤗 Link	MoE
Qwen/Qwen1.5-MoE-A2.7B	🤗 Link	MoE
google/gemma-2-2b	🤗 Link	Dense

Interpreting a trace

Status & stability

• Schema version: 0.3.4
• Classifier version: telemetry-v0.3.4
• This is a research tool; APIs may change between minor versions.
• Trace schema follows semver-ish rules — see CHANGELOG.

License

Acknowledgments

Special thanks to the reverse engineering community and former colleagues (especially those from the Fyyre era) who helped shape the low-level thinking behind Noesis Tension.