5.83% Direct Target
UNI-EXP-002 · Result Freeze v1.0
Primary outcome: INCONCLUSIVE
W families showed 7/120 Direct Targets; pooled Controls showed 9/240.
W 词族为 7/120；合并对照词族为 9/240。主结果不确定。
Preregistered primary result
Family-level Direct Target endpoint
3.75% Direct Target
95% CI 0.594–4.075
95% CI −4.12 to +9.57 pp
Two-sided Fisher exact p=.4185. The interval includes RR=1 and also permits RR≥2.0, so the frozen decision rule yields INCONCLUSIVE—not Supported and not Not Supported.
区间既包含 RR=1，也仍允许 RR≥2.0；因此冻结判定规则给出的唯一结论是“不确定”。
Robustness / sensitivity
Secondary analyses do not replace the primary result
Frequency strata
- High: 3/40 vs 7/80; RR=.857; p=1.000
- Middle: 4/40 vs 2/80; RR=4.000; p=.0945
- Lower: 0/40 vs 0/80; RR undefined; p=1.000
Borderline-inclusive
Using codes 1 or 2 only as a sensitivity analysis: W 9/120 vs Controls 12/240; RR=1.500 (95% CI .650–3.461); Fisher p=.3478.
Unavailable under frozen implementation
Repeated random-control resampling, a distinct lemma-level estimate, and V-control analysis were not invented after unblinding because the frozen implementation lacks the required algorithm or annotated data.
No Rescue Rule
No semantic expansion, favorable stratum selection, control replacement, family change, counterexample deletion, or alternative statistic was used to upgrade the result.
Exploratory AI annotation
High machine reproducibility, not human validation
99.17%
Unweighted, four frozen codes
AI-C resolved all disagreements
AI-A, AI-B, and final labels agree on classification
Machine-based exploratory blind annotation — not human double-blind validation.
Limitations
What this result does and does not show
The observed point estimate favors W, but uncertainty is wide and statistical evidence is insufficient under the preregistration. The experiment does not show that W means water, does not establish a letter-origin claim, and does not replace future independent human annotation.
The audit trail retains the contained early AI-B dispatch deviation. It produced no output before interruption, and the usable AI-B run began fresh after AI-A froze; no evidence indicates condition exposure or cross-run contamination.
点估计方向偏向 W，但不确定性很大，预注册标准下证据不足。本实验不证明 W 本身表示水，也不替代未来独立真人标注。