AutoZYX
Scenario Evidence Pack v0.1 Scenario Evidence Pack v0.1
ADSafetyPilot 的第一个证据包演示：从真实航测自然驾驶数据中提取 cut-in 场景，形成参数分布、收敛判断、证据缺口和测试用例候选。 The first ADSafetyPilot evidence-pack demo: extracted cut-in events become parameter distributions, convergence checks, evidence gaps, and test-case candidates.
过去的参数分析通常以一次性报告结束。证据包把数据、脚本、口径、图表和测试候选固化为可复用交付物。Traditional parameter studies often end as one-off reports. An evidence pack turns data, scripts, definitions, charts, and test candidates into a reusable product artifact.
AI 可以起草 HARA、SOTIF 和测试计划，但客户真正需要的是每个结论背后的数据依据、样本量和适用边界。AI can draft HARA, SOTIF, and test plans, but customers need the data basis, sample size, and applicability boundary behind each claim.
证据包保留从原始场景、导入脚本、质量过滤、分布结果到测试候选的链路，方便审查、复现和更新。The pack preserves the chain from raw scenarios, importer, quality filter, distribution outputs, and test candidates, making review and reproduction practical.
自然驾驶数据Naturalistic data
场景提取Scenario extraction
参数分布Parameter distributions
收敛与缺口Convergence and gaps
测试候选Test candidates
安全证据报告Evidence report
关键差异不在于“AI 帮你写一个表”，而在于把真实驾驶行为数据连接到 SOTIF 触发条件、场景代表性和测试用例生成。The differentiator is not an AI-generated form. It is the connection from real driving behavior to SOTIF triggering conditions, scenario representativeness, and test-case generation.
证据包基于 驭研科技 DRIVEResearch 航测自然驾驶数据能力构建。当前 v0.1 演示使用一个城市快速路 cut-in 样本数据包。The pack is built on DRIVEResearch aerial naturalistic driving-data capabilities. This v0.1 demo uses one urban-expressway cut-in sample package.
这不是完整高速 cut-in ODD 声明，也不能保证所有参数已经收敛。它展示的是如何把一个真实样本包转成可审查的数据证据链。This is not a full highway cut-in ODD claim and does not guarantee convergence for all parameters. It demonstrates how one real sample package can become an auditable data evidence chain.
40-60 和 60-80 是当前样本包中可展示的速度段；这不代表参数已充分收敛或覆盖完整 ODD。40-60 and 60-80 are usable demonstration bins in this sample package; this does not mean the parameters are fully converged or cover the full ODD.
580 valid 样本进入分布valid rows used
809 review 样本保留审计review rows retained
parameter_distribution.json 使用 sample_quality=valid，避免把缺少参考帧 TTC/THW 的样本混入客户展示分布。parameter_distribution.json uses sample_quality=valid, keeping review rows out of customer-facing distributions.
每张卡片展示一个参数。横向条带表示 P10-P95 分位范围，圆点表示 P50；两行分别对应 40-60 与 60-80 两个速度段。Each card shows one parameter. The horizontal band marks the P10-P95 percentile range, and the dot marks P50; the two rows compare the 40-60 and 60-80 speed bins.
km/h · n=270 / 294
m · n=270 / 294
m/s · n=270 / 294
s · n=270 / 294
s · n=270 / 294
s · n=270 / 294
同一张卡片内可比较条带长度和中位数位置；不同参数的横轴单位不同，不应跨卡片比较条带长度。这些数值来自当前样本包的 valid-only 分布，不代表最终收敛参数或完整 ODD 的统计边界。Compare band length and median position within the same card; axes differ by parameter, so band lengths should not be compared across cards. Values come from the valid-only distribution of the current sample package and are not final converged parameters or full-ODD statistical boundaries.
规则：按输入顺序做 20 个累计分割，最后五个 P10 值波动小于 5% 视为收敛。Rule: 20 cumulative splits by input order; last-five P10 fluctuation below 5% is marked converged.
按速度段输出 P5/P10/P25/P50/P75/P95、样本数、缺失数和 P10 收敛状态。Percentiles, sample counts, missing counts, and P10 convergence status by speed bin.
生成 representative、boundary、challenge 三类测试候选，并匹配真实 exemplar event。Representative, boundary, and challenge test candidates matched to real exemplar events.
记录 raw package manifest hash、artifact sha256 和生成命令，支持审计复现。Raw-package manifest hash, artifact SHA-256 values, and reproducibility commands.
公开演示摘要，说明样本范围、证据链价值，以及当前参数不能保证收敛的边界。A public demo summary covering sample scope, evidence-chain value, and the current parameter-convergence limits.
用真实 cut-in 分布支持“合理可预见”触发条件和场景边界，而不是只依赖专家经验阈值。Use real cut-in distributions to support reasonably foreseeable triggering conditions and scenario boundaries.
从同一速度段内选择代表、边界和挑战工况，再匹配真实样本，减少不物理的参数拼接。Select representative, boundary, and challenge cases within the same speed bin, then match real exemplars.
为测试方案、评审材料和数据支撑报告提供可追溯的样本量、分布和适用边界。Provide traceable sample counts, distributions, and applicability boundaries for testing and assessment materials.
如果你正在做 SOTIF、功能安全、测试评价、NCAP 或准入项目，ADSafetyPilot 可以把你的场景问题转成数据证据包、测试候选和审查材料。If you work on SOTIF, functional safety, testing, NCAP, or access projects, ADSafetyPilot can translate scenario questions into evidence packs, test candidates, and review-ready material.