MagistrTheOne/SHUTEN-DOJI
SHUTEN-DŌJI by NULLXES DAI is a 27 billion parameter strategic intelligence system based on the Qwen3.6-27B foundation model, fine-tuned with a custom Constitution SFT v2. Unlike general-purpose chatbots, it is specifically designed to produce structured operational intelligence outputs following a 'State → Causes → Options → Impact → Future State → Confidence' format. This model excels at generating structured analytical responses for strategic business and operational planning.
Loading preview...
SHUTEN-DŌJI: Strategic Intelligence System
SHUTEN-DŌJI, developed by NULLXES DAI, is a 27 billion parameter model built upon the Qwen3.6-27B foundation. It is not a general-purpose chatbot but a specialized strategic intelligence system engineered to produce highly structured operational intelligence.
Key Capabilities
- Structured Output Generation: Generates responses in a predefined
State → Causes → Options → Impact → Future State → Confidenceformat, ideal for analytical and strategic planning. - Specialized Fine-tuning: Utilizes a custom "Constitution SFT v2" with 50 gold examples, specifically designed to eliminate common planning failures and improve structural adherence.
- Performance: Achieved 7/8 structural wins against the base Qwen model in side-by-side evaluations, demonstrating superior adherence to the target output structure and zero "poison" (undesired action-trace failures).
Use Cases
- Operational Intelligence: Ideal for scenarios requiring structured analysis of current states, identification of causal factors, evaluation of options, prediction of impacts, and forecasting future states with confidence levels.
- Strategic Planning: Supports business and operational planning by providing a consistent framework for analyzing complex situations and informing decision-making.
Limitations
As an MVP release, SHUTEN-DŌJI has been trained on a limited dataset of 50 examples. It may occasionally exhibit Qwen-style reasoning traces and is not yet optimized with DPO or for consequence prediction (impact cluster fine-tuning). It requires approximately 54GB VRAM for bfloat16 inference.