Install any skill in seconds. Free to start, no credit card required.
Get Started Free →China standard: Behavioral Safety. # IEEE 2846-2022 自动驾驶系统安全假设
.claude/skills/pangzhenying2025-china-behavioral-safety/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -43% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 80% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 6% | 0% |
| 属性 | 值 | |------|-----| | 标准编号 | IEEE 2846-2022 | | 名称 | A Framework for ADS Safety Assumptions | | 状态 | 已发布(2022年) | | 推荐等级 | P1-强烈推荐 | | 核心内容 | 定义ADS行为安全模型,含RSS(Responsibility-Sensitive Safety)框架 |
IEEE 2846 安全假设体系
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
1. 合理预见行为 (Reasonably Foreseeable Behavior)
└── 其他道路参与者的预期行为范围
2. 安全响应规则 (Safety Response Rules)
└── ADS在各场景下的安全行为要求
3. 事故责任判定 (Blame Attribution)
└── 当碰撞不可避免时的责任归属
4. 安全距离模型 (Safe Distance Model)
├── 纵向安全距离 = f(v_ego, v_front, a_max_brake, reaction_time)
├── 横向安全距离 = f(v_lateral, lane_width)
└── 交叉路口安全模型
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━python# RSS纵向安全距离计算 def rss_longitudinal_safe_distance( v_ego: float, # 自车速度 (m/s) v_front: float, # 前车速度 (m/s) rho: float, # 反应时间 (s), 通常取0.5-1.0s a_max_accel: float, # 最大加速度 (m/s²) a_max_brake: float, # 最大制动减速度 (m/s²) a_min_brake: float, # 前车最小制动减速度 (m/s²) ) -> float: """ d_safe = v_ego * rho + 0.5 * a_max_accel * rho² + (v_ego + rho * a_max_accel)² / (2 * a_max_brake) - v_front² / (2 * a_min_brake) """ v_ego_after_reaction = v_ego + rho * a_max_accel d_ego_reaction = v_ego * rho + 0.5 * a_max_accel * rho**2 d_ego_brake = v_ego_after_reaction**2 / (2 * a_max_brake) d_front_brake = v_front**2 / (2 * a_min_brake) return max(0, d_ego_reaction + d_ego_brake - d_front_brake)
IEEE 2846 中国场景适配考虑
├── 反应时间参数:需考虑中国交通密度
├── 行为假设:电动自行车/三轮车行为模型
├── 交叉路口:中国特有交通规则(右转不停车等)
├── 高速公路:匝道合流行为差异
└── 责任判定:与中国交通法规对齐skills/china-standards/ads-safety/ — ADS安全要求skills/china-standards/ai-safety/ — AI安全标准skills/automotive-e2e-safety-analysis/ — 端到端安全分析| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | fail→pass | 11,062 | 7,463 | -33% | 1 | 1 | 0% | 2,267 | 2,450 | +8% | 0 | 0 | — |
case-18 | fail→pass | 13,176 | 2,487 | -81% | 1 | 1 | 0% | 2,211 | 1,256 | -43% | 0 | 0 | — |
case-01 | pass→fail | 15,455 | 10,943 | -29% | 1 | 1 | 0% | 3,796 | 3,319 | -13% | 0 | 0 | — |
case-02 | pass→pass | 16,697 | 14,454 | -13% | 1 | 1 | 0% | 3,378 | 3,816 | +13% | 0 | 0 | — |
case-03 | fail→pass | 12,747 | 16,543 | +30% | 1 | 1 | 0% | 2,260 | 4,073 | +80% | 0 | 0 | — |
case-04 | fail→pass | 13,079 | 2,161 | -83% | 1 | 1 | 0% | 2,130 | 1,172 | -45% | 0 | 0 | — |
case-05 | fail→pass | 13,007 | 9,606 | -26% | 1 | 1 | 0% | 2,205 | 2,341 | +6% | 0 | 0 | — |
case-06 | pass→pass | 11,313 | 8,806 | -22% | 1 | 1 | 0% | 2,014 | 2,601 | +29% | 0 | 0 | — |
case-07 | pass→pass | 16,853 | 18,915 | +12% | 1 | 1 | 0% | 2,942 | 4,037 | +37% | 0 | 0 | — |
case-08 | pass→pass | 21,894 | 23,573 | +8% | 1 | 1 | 0% | 3,608 | 4,809 | +33% | 0 | 0 | — |
case-09 | pass→pass | 15,407 | 15,246 | -1% | 1 | 1 | 0% | 2,464 | 3,376 | +37% | 0 | 0 | — |
case-10 | pass→pass | 18,056 | 12,430 | -31% | 1 | 1 | 0% | 2,772 | 3,043 | +10% | 0 | 0 | — |
case-11 | pass→pass | 9,162 | 7,414 | -19% | 1 | 1 | 0% | 1,526 | 2,082 | +36% | 0 | 0 | — |
case-13 | pass→pass | 9,960 | 12,780 | +28% | 1 | 1 | 0% | 1,620 | 2,870 | +77% | 0 | 0 | — |
case-14 | pass→pass | 13,444 | 13,702 | +2% | 1 | 1 | 0% | 1,956 | 3,122 | +60% | 0 | 0 | — |
case-15 | pass→pass | 14,096 | 17,215 | +22% | 1 | 1 | 0% | 2,338 | 3,782 | +62% | 0 | 0 | — |
case-16 | pass→pass | 7,262 | 6,161 | -15% | 1 | 1 | 0% | 1,487 | 2,097 | +41% | 0 | 0 | — |
case-17 | pass→pass | 6,739 | 6,038 | -10% | 1 | 1 | 0% | 1,343 | 2,071 | +54% | 0 | 0 | — |
case-19 | fail→pass | 11,433 | 1,951 | -83% | 1 | 1 | 0% | 2,007 | 1,138 | -43% | 0 | 0 | — |
case-20 | pass→pass | 13,979 | 13,007 | -7% | 1 | 1 | 0% | 2,442 | 3,252 | +33% | 0 | 0 | — |
case-21 | pass→pass | 24,956 | 33,106 | +33% | 1 | 1 | 0% | 4,258 | 7,001 | +64% | 0 | 0 | — |
case-22 | pass→pass | 20,425 | 21,728 | +6% | 1 | 1 | 0% | 4,734 | 5,786 | +22% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +23 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.