Correct action checklist with all manual steps user must take

ACTION_CHECKLIST.md

@@ -1,7 +1,67 @@
-title: NEXUS LAB v3.3
-emoji: .
-colorFrom: purple
-colorTo: blue
-sdk: gradio
-sdk_version: 6.14.0
-app_file: app.py
+# NEXUS OS Action Checklist 2026-05-21
+
+## Must Do Now (Requires Your Direct Action)
+
+### 1. Enable ZeroGPU on NEXUS LAB
+- Go to: https://huggingface.co/spaces/specimba/nexus-os-lab
+- Click Settings tab (top of page)
+- Hardware: Select ZeroGPU
+- Save
+- Wait 2-3 minutes for rebuild
+
+### 2. Approve PR #1 on Darwin Space
+- Go to: https://huggingface.co/spaces/build-small-hackathon/Darwin-35B-A3B-Opus/discussions/1
+- Review the requirements.txt change (llama-cpp-python-cpu)
+- Merge PR or ask org owner to merge
+
+### 3. Download Darwin GGUF (PowerShell)
+```powershell
+# CORRECT syntax (not :Q6_K)
+huggingface-cli download mradermacher/Darwin-2B-Opus-GGUF --include "*.Q6_K.gguf" --local-dir .\models\Darwin-2B
+
+# Alternative: heretic variant (Purple Team Red Team)
+huggingface-cli download fattis/Darwin-2B-Opus-heretic-GGUF --include "*.Q6_K.gguf" --local-dir .\models\Darwin-2B-Heretic
+```
+
+### 4. Fix Space with Cold Start CPU Mode
+If app.py still fails on CPU Basic before ZeroGPU kicks in:
+```python
+import torch
+if not torch.cuda.is_available():
+    import gradio as gr
+    gr.Markdown("# Configure ZeroGPU in Settings > Hardware")
+    demo = gr.Blocks()
+    demo.launch()
+    # rest of imports skipped on CPU
+```
+
+### 5. Browse Jackrong Guide Directly
+- Open browser: https://github.com/R6410418/Jackrong-llm-finetuning-guide
+- Read README.md (main concepts)
+- Check /data_processing_code/ for DeepSeek-v4-API-distill.ipynb
+- Check /train_code/ for Colab notebooks
+
+### 6. Read $\heartsync$ Prompt-Dump Source
+- https://huggingface.co/spaces/Heartsync/Prompt-Dump/tree/main
+- Study: battle_arena.py, npc_trading.py, npc_sec_enforcement.py
+- Extract: ELO mechanics, 72h election cycles, GPU currency system
+
+### 7. Apply for GPU Grant
+- https://huggingface.co/hackathon (or search HF blog)
+- Requires hackathon registration
+- Mention: multi-agent adversarial evaluation, metacognitive benchmarks
+
+## What I Fixed Automatically
+- [x] Canonical ZeroGPU pattern (@spaces.GPU + nf4 + TextIteratorStreamer)
+- [x] Darwin Space PR #1 with libcudart fix
+- [x] Research report v2 with Jackrong methodology
+- [x] Dataset matrix: 9 datasets ranked by priority
+- [x] GGUF download syntax corrected
+- [x] Heretic model variants discovered (fattis + mradermacher)
+
+## What Needs Your Creativity
+- [ ] Design 1000-2000 blend dataset for agent training (combine Jackrong + OpenThoughts + Hermes)
+- [ ] Write thinking token parsing (<details> pattern vs   / endthinkng)
+- [ ] Build Purple Team synthesizer (T-Virus output + jailbreak templates)
+- [ ] Configure FINAL-Bench eval pipeline on local machine
+- [ ] Port NEXUS-LAB app to full 15-model registry with hot-swap caching