Safety certification of Vision-Language-Action (VLA) driving planners under ISO 21448 (SOTIF) rests on an Operational Design Domain (ODD) specification that answers two complementary questions: when does the planner start to fail, and how severely does it fail once it does? We evaluate Alpamayo R1, a 10B-parameter open-weight driving VLA, on 15,968 (clip, attack) pairs. We find a conservative-aggregate gap: an aggregate safe threshold of $σ\leq 50$ under a 15% average displacement error (ADE) budget masks well-sampled scenarios that tolerate the top of the tested grid ($σ= 70$). A Gaussian Mixture Model (GMM) on the changed-explanation subset identifies six discrete severity bands (BIC-optimal $k{=}6$), so two perturbation conditions with the same mean error can differ materially in their share of high-severity (C4/C5) failures. Joining the two analyses on the same corpus surfaces a finding neither yields in isolation: the scenarios with the loosest noise thresholds are not those with the lowest high-severity rate: STOP_SIGNAL concentrates roughly $4\times$ the C4/C5 share of LANE_KEEPING despite tolerating a larger $σ$. A deployable SOTIF ODD specification for driving VLAs therefore requires a two-dimensional safety envelope, not a single aggregate value per hazard.

翻译：根据ISO 21448（SOTIF）标准对视觉-语言-动作（VLA）驾驶规划器进行安全认证，需依据运行设计域（ODD）规范，该规范回答两个互补问题：规划器何时开始失效，以及一旦失效其严重程度如何？我们在15,968个（片段，攻击）对上评估了Alpamayo R1——一个100亿参数的开源驾驶VLA模型。我们发现一个保守聚合缺口：在15%平均位移误差（ADE）预算下，聚合安全阈值σ≤50掩盖了能够容忍测试网格上限（σ=70）的充分采样场景。对变解释子集应用高斯混合模型（GMM）识别出六个离散严重度波段（BIC最优k=6），因此具有相同平均误差的两个扰动条件在高严重度（C4/C5）失效的占比上可能存在实质性差异。将两种分析应用于同一语料库，发现了单一分析无法得出的结论：噪声阈值最宽松的场景并非高严重度率最低的场景——尽管STOP_SIGNAL容忍更大的σ值，其C4/C5占比约为LANE_KEEPING的4倍。因此，驾驶VLA的可部署SOTIF ODD规范需要二维安全包络，而非每个危险场景的单一聚合值。