# Experiment 001A Round 2 — Public-Safe Summary｜公开安全摘要

## English

**Exploratory AI Result — Tentatively Supported**

**Preregistered Human Annotation: Pending**

The first authorized exploratory AI unblinding found the same modest positive direction in both frozen AI passes:

- **AI-A:** M 76/510 (14.90%) vs Controls 60/510 (11.76%); risk difference **+3.14 percentage points**, 95% CI **−1.05 to +7.33**; one-sided Fisher **p=0.0835**.
- **AI-B:** M 79/510 (15.49%) vs Controls 61/510 (11.96%); risk difference **+3.53 percentage points**, 95% CI **−0.71 to +7.77**; one-sided Fisher **p=0.0608**.

The prespecified evidence-family robustness analysis retained nearly the same positive differences: **+3.42 pp** for AI-A and **+3.47 pp** for AI-B. Both 95% intervals still included zero.

This is a directionally consistent but uncertain exploratory signal. It does not confirm or reject the preregistered human hypothesis. AI-A and AI-B were analyzed separately, and neither pass was selected, merged, or adjudicated after unblinding.

The raw restricted analysis key, row-level identities, joined data, and row-level AI annotations remain private to preserve future human blinding. Experiment 001B was not started.

## 中文

**Exploratory AI Result — Tentatively Supported｜探索性 AI 结果——初步支持**

**Preregistered Human Annotation: Pending｜预注册真人标注：待进行**

第一次经授权的探索性 AI 揭盲在两轮冻结 AI 标注中得到方向一致、幅度较小的正向结果：

- **AI-A：** M 组 76/510（14.90%），对照组 60/510（11.76%）；绝对风险差 **+3.14 个百分点**，95% CI **−1.05 至 +7.33**；单侧 Fisher **p=0.0835**。
- **AI-B：** M 组 79/510（15.49%），对照组 61/510（11.96%）；绝对风险差 **+3.53 个百分点**，95% CI **−0.71 至 +7.77**；单侧 Fisher **p=0.0608**。

预先规定的证据家族去重稳健性分析仍保持近似的正向差异：AI-A 为 **+3.42 个百分点**，AI-B 为 **+3.47 个百分点**；两者的 95% CI 均仍包含 0。

这是一个方向一致但仍不确定的探索性信号，不能确认或否定预注册真人假说。AI-A 与 AI-B 始终分开分析；揭盲后没有挑选、合并或裁决任何一轮结果。

原始受限分析密钥、逐项身份、连接后的数据和逐项 AI 标注继续保持私有，以保护未来真人盲标。Experiment 001B 未启动。