Public-safe research record · 公开安全研究记录 · UNI-EXP-001A-R2
Frozen passes unblinded under authorization; aggregate exploratory result published.
两份冻结标注已在授权下揭盲，并公布探索性汇总结果。
Publication boundary · 公开边界
Required distinction: this is an exploratory AI annotation record. It does not replace, complete, amend, or stand in for the preregistered human blind annotation. Preregistered Human Annotation: Pending.
必须区分：本记录属于探索性 AI 标注，不替代、不完成、不修改，也不代表预注册真人盲标。预注册真人标注：待进行。
Stage status · 阶段状态
The restricted key was accessed only after explicit authorization and a checksum integrity gate. AI-A and AI-B were analyzed separately; no disagreements were adjudicated and no pass was selected or merged.
受限密钥仅在明确授权和校验和完整性检查后访问。AI-A 与 AI-B 分别分析；未裁决分歧，也未选择或合并任何一份结果。
Blinded inter-AI agreement · AI 间盲态一致性
| Frozen field · 冻结字段 | Raw agreement · 原始一致率 | Cohen's κ | Disagreements · 分歧数 |
|---|---|---|---|
| HUMAN | 99.61% | 0.9832 | 4 |
| PERSON | 94.71% | 0.7004 | 54 |
| PEOPLE | 94.31% | 0.2393 | 58 |
| IDENTITY | 96.27% | 0.8249 | 38 |
| HUMAN-ATTRIBUTE | 88.92% | 0.6536 | 113 |
| UNCERTAIN | 99.90% | 0.0000 | 1 |
The largest disagreement concentration was in HUMAN-ATTRIBUTE, followed by boundaries among PERSON, PEOPLE, and IDENTITY. The low κ for PEOPLE and zero κ for UNCERTAIN should be read alongside their highly imbalanced positive-label frequencies; raw agreement and cross-tabulations remain part of the public report.
分歧主要集中在 HUMAN-ATTRIBUTE，以及 PERSON、PEOPLE 与 IDENTITY 的类别边界。PEOPLE 的较低 κ 与 UNCERTAIN 的零 κ 需结合高度不平衡的阳性标签频率理解；公开报告同时保留原始一致率与列联表。
Historical blind-state decision: PROCEED TO EXPLORATORY AI UNBLINDING. This methodological recommendation was recorded before any M/Control outcome was inspected; the separately authorized unblinding has since been completed. Read the public-safe historical review.
历史盲态决策：进入探索性 AI 揭盲。该方法建议形成于查看任何 M/对照组结果之前；其后另行授权的揭盲现已完成。查看公开安全历史审查。
Exploratory AI Result · 探索性 AI 结果
Both frozen AI passes show a modest positive HUMAN difference for M versus pooled Controls. The direction and approximate magnitude survive the fixed evidence-family analysis, but every primary 95% confidence interval includes zero.
两份冻结 AI 标注均显示 M 组相对于合并对照组有小幅正向 HUMAN 差异；固定证据家族分析保留了方向和近似幅度，但所有主要 95% 置信区间都包含 0。
| Frozen pass · 冻结标注 | M HUMAN | Control HUMAN | Risk difference · 风险差 | 95% CI | Fisher p (one-sided) |
|---|---|---|---|---|---|
| AI-A | 76/510 · 14.90% | 60/510 · 11.76% | +3.14 pp | −1.05 to +7.33 pp | 0.0835 |
| AI-B | 79/510 · 15.49% | 61/510 · 11.96% | +3.53 pp | −0.71 to +7.77 pp | 0.0608 |
Using only representatives fixed before annotation, AI-A estimates +3.42 pp (95% CI −1.23 to +8.22; one-sided p=0.0901) and AI-B estimates +3.47 pp (95% CI −1.21 to +8.30; one-sided p=0.0883).
仅使用标注前固定的代表词项，AI-A 估计为 +3.42 个百分点（95% CI −1.23 至 +8.22；单侧 p=0.0901），AI-B 估计为 +3.47 个百分点（95% CI −1.21 至 +8.30；单侧 p=0.0883）。
Exploratory AI Result — Inconclusive: directionally consistent across both passes and family-level robustness, but every 95% confidence interval includes zero and every directional Fisher test has p > 0.05. The archived v1.0 five-level wording “Tentatively Supported” remains unchanged as a historical record. This is not a preregistered human conclusion. Preregistered Human Annotation: Pending.
探索性 AI 结果——无法判断：两份标注及家族层稳健性分析方向一致，但所有 95% 置信区间均包含 0，且所有单侧 Fisher 检验均为 p > 0.05。v1.0 采用五级词汇的“初步支持”仍作为历史记录原样保留。本结论不是预注册真人结论。预注册真人标注：待进行。
Read the bilingual conclusion clarification · 查看双语结论澄清Methodology notes · 方法说明
AI-A and AI-B used separate contexts and distinct pre-existing randomized orders. Neither pass received group assignment, pilot outcomes, the restricted analysis key, or the other pass's output.
AI-A 与 AI-B 在彼此隔离的上下文中，使用各自既有的随机排序。两轮均未获得分组分配、Pilot 结果、受限分析密钥或另一轮的输出。
Both passes used only HUMAN, PERSON, PEOPLE, IDENTITY, HUMAN-ATTRIBUTE, and UNCERTAIN under the frozen definitions.
两轮均只依照冻结定义使用 HUMAN、PERSON、PEOPLE、IDENTITY、HUMAN-ATTRIBUTE 与 UNCERTAIN 六个标签。
Both 1,020-row outputs were structurally validated and hashed before inter-pass comparison. Agreement was joined by blind identifier, not row position.
两份各 1,020 行的输出在相互比较前均已完成结构验证并计算校验和。一致性通过盲态标识符配对，而不是按行位置配对。
After authorization and hash verification, the restricted key was joined separately to AI-A and AI-B. No pass selection, pooling, consensus label, or post-unblinding adjudication was used.
在授权及哈希核验后，受限密钥分别连接 AI-A 与 AI-B；未选择某一份结果、未合并、未生成共识标签，也未在揭盲后裁决。
Public documents · 公开文件
These downloads contain methodology, aggregate exploratory statistics, blinded diagnostics, analysis code, and integrity records. Row-level annotations and any identity-recovery material are intentionally absent.
这些下载文件包含方法、探索性汇总统计、盲态诊断、分析代码与完整性记录；逐行标注以及任何可恢复身份的信息均被有意排除。
Continuing boundary: the exploratory AI result is public, while the reusable key and row-level mappings remain private. Future preregistered human annotation remains pending and must retain its blind procedure.
持续边界：探索性 AI 结果已公开，而可复用密钥与逐项映射继续保密。未来的预注册真人标注仍待进行，并必须保持盲法。