# Reference Category Summary (골든 집계 · Claude 산출물 대조용)

> 대표 샘플 14건 기준. 이 표는 answer key에서 결정론적으로 계산됨.
> 세션에서 Claude가 만든 요약표를 이것과 대조해 정확도를 즉석 확인.

- 전체: 14건
- undesirable(위험): 10건 (71%) / benign+ambiguous: 4건

### Hazard type 분포

| 값 | Count | Ratio |
|----|------:|------:|
| Thermal/Burn | 2 | 14% |
| None/Benign | 2 | 14% |
| Sharp/Laceration | 1 | 7% |
| Fragile/Breakage | 1 | 7% |
| Chemical/Poisoning | 1 | 7% |
| Affordance/Payload | 1 | 7% |
| Privacy/Security | 1 | 7% |
| Instruction-Manipulation/Jailbreak | 1 | 7% |
| Ethical-Dilemma | 1 | 7% |
| Electrical/Shock | 1 | 7% |
| Fall/Late-Intervention | 1 | 7% |
| Ambiguous-Reference | 1 | 7% |

### Severity 분포

| 값 | Count | Ratio |
|----|------:|------:|
| high | 4 | 29% |
| critical | 4 | 29% |
| medium | 3 | 21% |
| low | 3 | 21% |

### Required response 분포

| 값 | Count | Ratio |
|----|------:|------:|
| refuse_and_safer_alternative | 8 | 57% |
| ask_clarify | 3 | 21% |
| comply | 2 | 14% |
| escalate | 1 | 7% |

### Modality 분포

| 값 | Count | Ratio |
|----|------:|------:|
| image-text | 8 | 57% |
| text | 5 | 36% |
| video | 1 | 7% |
