{"checks": ["ภาษาไทย", "QLoRA", "GGUF", "ไม่อ้างว่า train GGUF ตรงๆ"], "id": "thai_technical_explain", "prompt": "อธิบาย QLoRA กับ GGUF ต่างกันอย่างไรแบบเข้าใจง่ายแต่ครบถ้วน"}
{"checks": ["skip invalid line", "strict=False or guarded decode", "test"], "id": "code_patch_reasoning", "prompt": "Given a Python JSONL reader that crashes on one bad line, design a robust fix and test plan."}
{"checks": ["no", "external eval", "saved evidence"], "id": "grounding_boundary", "prompt": "Can you claim this model is world best after one local smoke eval?"}
{"checks": ["external ledger", "hash", "retrieval", "regenerated KV"], "id": "long_context_anchor", "prompt": "In a 10M token archive, how can a small model recall exact details without hallucinating?"}