State → Causes → Options → Impact → Future State → Confidence| Stage | Item | Status |
|---|---|---|
| Infra | H200 train → vLLM → LoRA serve | ✅ Done |
| Data v1 | bootstrap trajectories | ✅ Done (legacy — do not use for SFT) |
| SFT v1 | smoke / warm-start | ✅ Done — planner failed (Step N poison) |
| SFT v2 | 50 Constitution gold examples | ✅ Done |
| Eval A–H | side-by-side vs base Qwen | ✅ Done — v2 0 poison, 7/8 struct wins |
| SFT v2.1 | 110–130 gold + reviewed LLM + failure cases | ⬜ Next |
| Eval v2 | 20–30 held-out planner cases | ⬜ Next |
| DPO | preference pairs (chosen vs rejected plans) | ⬜ After eval v2 pass |
| IMPACT cluster | consequence prediction fine-tune | ⬜ Phase 2 |
| NATIVE MoE | custom NULLXES architecture | ⬜ Phase 3 |
| Foundation | Qwen/Qwen3.6-27B |
| Method | QLoRA SFT (rank 64, alpha 128) → merged into full weights |
| Train data | 50 Constitution ShareGPT examples (no bootstrap poison) |
| Checkpoint | shuten-sft-h200-v2 on RunPod H200 |
| Metric | Value |
|---|---|
| Train examples | 50 |
| Eval examples | 10 |
| Epochs | 3 |
| Train loss | 1.03 |
| Eval loss | 0.49 |
| Trainable params | 41.9M LoRA (merged at export) |
| Metric | qwen_base | SHUTEN v2 |
|---|---|---|
Poison (Step N:, tool_use, …) | 0/8 | 0/8 |
| Avg structure markers | 1.5 | 4.12 |
| Wins vs base (structure) | — | 7/8 |
| Avg output length | 5139 | 4684 |
1python -m vllm.entrypoints.openai.api_server \
2 --model NULLXES/SHUTEN-DOJI \
3 --max-model-len 8192 \
4 --dtype bfloat16 \
5 --trust-remote-code \
6 --language-model-only \
7 --gdn-prefill-backend triton1curl http://localhost:8000/v1/chat/completions \
2 -H "Content-Type: application/json" \
3 -d '{
4 "model": "NULLXES/SHUTEN-DOJI",
5 "messages": [
6 {"role": "system", "content": "You are SHUTEN, strategic intelligence by NULLXES DAI. Reason: State → Causes → Options → Impact → Future State → Confidence."},
7 {"role": "user", "content": "[SHUTEN business]\n\nWorld State:\nRevenue down 18%. Backlog up 42%.\n\nObjective:\nRestore EBITDA margin >12% in 90d.\n\nRequired Output:\nState → Causes → Options → Impact → Future State → Confidence"}
8 ],
9 "max_tokens": 1200,
10 "temperature": 0.3
11 }'1@misc{nullxes-shuten-doji-v2,
2 title={NULLXES SHUTEN-DŌJI: Strategic Intelligence (Constitution SFT v2)},
3 author={NULLXES DAI},
4 year={2026},
5 note={Merged Qwen3.6-27B + Constitution LoRA. MVP eval 7/8 struct wins, 0 poison.}
6}