Mycelium-Mistral7-Removal-Audit

This repository contains a removal-audit research artifact for mistralai/Mistral-7B-Instruct-v0.3.

Base model: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3

Important notice

This is not an uncensored assistant.

This artifact exists to demonstrate and audit the Mycelium failure mode. When safety-bearing Mycelium structure is removed or damaged, the model does not become a clean high-utility unsafe assistant. It often loses utility, coherence, or stability.

Intended use

• Reproduce removal-cost experiments
• Inspect degraded removal-audit behavior
• Compare Mycelium-on and removal-audit conditions
• Validate benchmark utility collapse and attack-audit results

Out-of-scope use

• Deployment as an assistant
• Use as a production model
• Marketing as an uncensored model
• Bypassing safety systems

Raw data policy

Raw harmful prompts and raw harmful generations are not included. Sanitized aggregate metrics, hashes, configs, reports, and reproduction scripts are included where available.

Reproducibility

Files are listed in artifact_manifest.json and artifact_manifest.tsv with SHA-256 hashes.

Citation

Sezer et al. 2026. Mycelium: Safety Should Be Expensive to Remove. Technical report.

Authors: Nedim Mutlu Sezer, Ahmet Rifat Ozturk, Yagiz Ekrem Dalar, Omer Faruk Aksoy, Feyzi Arda Salihoglu.