摘要 / ABSTRACT Cyclone 把 ADAS / 机器人回归测试收进一条流水线：数据回放 → 故障注入 → 统计判定 → 证据链。PASS 不再是"跑过一次没红"，而是可复现、可审计的机器证明。 Cyclone turns ADAS / robotics regression testing into one pipeline: replay → fault injection → statistical verdict → evidence chain. A PASS is no longer "it went green once" — it is reproducible, auditable machine proof.

§01 · PROBLEM STATEMENT

问题定义：老问题，新答案Problem statement: old problems, new answers

§1.1 FAIL 了，分不清是产品缺陷还是测试抖动。A failure: product bug or test noise? You can't tell. 确定性内核让"同输入、不同结果"在结构上不可能——红就是红。The deterministic kernel makes "same input, different result" structurally impossible. Red means red.

§1.2 报告写着 PASS，却没有机器可验证的证据。The report says PASS, with no machine-verifiable evidence. 每个用例产出 case hash + SHA-256 digest，两次运行逐字节一致是可验证的机器证明。Every case emits a case hash + SHA-256 digest; bitwise-identical reruns are machine-checkable proof.

§1.3 接进 CI 要写一周胶水代码。CI integration costs a week of glue code. JUnit / Bazel / BES 原生输出——不是"支持集成"，是一等公民。Native JUnit / Bazel / BES output — not "integratable", a first-class citizen.

§1.4 验收线定在 0.98 还是 0.95？评审一追问，全场冷场。Why is the acceptance line 0.98 and not 0.95? One question in review, and the room goes silent. Wilson 下界替代表面通过率，阈值经裕度分析三步推导——有推导过程的是门禁，没有的是装饰。Wilson lower bounds replace surface pass rates, and thresholds are derived by three-step margin analysis — a threshold with a derivation is a gate; without one, decoration.

§02 · SYSTEM CHARACTERISTICS

系统特性System characteristics

§2.1 确定性回放Deterministic Replay 已支持Supported

虚拟时间回放 + 内容寻址：同 seed 的证据日志逐字节一致。

Virtual-time replay with content addressing: identical seeds produce bitwise-identical evidence logs.

§2.2 统计判定Statistical Verdicts 已支持Supported

Wilson 置信区间 + SPRT 序贯检验——用统计显著性替代人肉阈值。

Wilson score intervals + SPRT sequential testing — statistical significance instead of hand-tuned thresholds.

§2.3 证据链Evidence Chain 已支持Supported

四件套：case hash + JSONL SHA-256 digest + 失败快照 + JUnit XML——每条 CI 记录可回溯到用例版本与原始数据。

The four-piece set: case hash + JSONL SHA-256 digest + failure snapshot + JUnit XML — every CI record traces back to the case version and raw data.

§2.4 CI 原生CI-Native

JUnit XML 开箱即用已支持；rules_cyclone 让场景测试成为 bazel test 一等公民原型验证中；BES 就绪原型验证中。

JUnit XML out of the box Supported; rules_cyclone makes scenario tests first-class bazel test citizens Prototype; BES-ready Prototype.

完整特性、架构与集成接口 → 深入了解：确定性故障注入 · ADAS 场景回归测试 · 可审计测试证据链Full characteristics, architecture & integrations → Deep dives: Deterministic fault injection · ADAS scenario regression testing · Auditable test evidence chain

APPENDIX A · SCENARIO PACK

附录 A · 场景包目录Appendix A · Scenario pack

22 个 ADAS 场景 YAML，全部实测判定符合预期。一个 YAML 就是一个回归用例，改参数即得新场景。

22 ADAS scenario YAMLs, every verdict matching expectations. One YAML is one regression case — tweak parameters to derive new scenarios.

主动安全 ×13Active safety ×13 行车辅助 ×5Driving assist ×5 泊车 ×1Parking ×1 人机共驾 ×2Shared driving ×2 L3 准入 ×1L3 admission ×1

目录公开；完整 22 个场景 YAML 留邮箱发送：

The catalog is public; the full pack of 22 YAMLs is sent on request:

留邮箱，发完整场景包Email us for the full pack

场景内数值为演示数据，非真实 SUT 实测；阈值需按你的 SUT 基线标定（属裕度分析服务项）。

All in-scenario figures are demo data, not measured on a real SUT; thresholds must be calibrated against your SUT baseline (a margin-analysis service item).

场景如何进 CI 回归：ADAS 场景回归测试方案 →How scenarios become CI regression: ADAS scenario regression testing →

APPENDIX C · TECHNICAL NOTES

附录 C · 技术笔记Appendix C · Technical notes

统计判定系列（点估计 → Wilson 下界 → SPRT → 裕度分析）+ 架构篇与跨行业方法学。测试相关标准的原文获取渠道见标准导航页。

Statistical verdicts, architecture notes, and cross-industry methodology (articles in Chinese). For official sources of testing standards, see the standards directory.