Cross-domain crisis detection and socially diverse user modeling through on-device personal data integration. Simple rules + rich data = effective crisis detection. No ML required. 通过设备端跨域个人数据整合实现危机检测与多元用户建模。简单规则 + 丰富数据 = 有效危机预警。无需机器学习。
Builds on Prism V2 (IIR 1.48x, federation protocol, model-scale curve) 基于 Prism V2 构建 (IIR 1.48x, 联邦协议, 模型规模曲线)
Cross-domain signal convergence improves crisis detection precision 7x (L1: 0.10 → L3: 0.71) using only threshold rules — no machine learning, no training data, no parameter tuning. The "intelligence" comes entirely from the data architecture: having multiple independent domains that correlate only during genuine crises. 跨域信号收敛将危机检测精度提升了 7 倍 (L1: 0.10 → L3: 0.71)，仅使用阈值规则——无需机器学习、无需训练数据、无需调参。"智能"完全来自数据架构：多个独立数据域仅在真实危机时才会产生相关性。
KEY关键 Precision jumps 7x from L1 to L3 purely through cross-domain convergence. Real crises perturb multiple domains simultaneously; noise affects only one. 精度从 L1 到 L3 跃升 7 倍，完全通过跨域收敛实现。真实危机会同时扰动多个域；噪声通常只影响单个域。
| Drift Class漂移类型 | Precision | Recall | F1 | n |
|---|---|---|---|---|
| Normal | 0.111 | 0.667 | 0.191 | 6 |
| Unexpected | 0.286 | 1.000 | 0.444 | 5 |
| Severe | 0.412 | 0.875 | 0.560 | 3 |
| Config | Data Sources数据源 | Qwen | GLM | Type |
|---|---|---|---|---|
| A | Dailyn (finance财务) | 72.4 | 67.4 | Single单域 |
| B | Mealens (diet饮食) | 73.4 | 70.1 | Single单域 |
| C | Ururu (mood情绪) | 71.0 | 70.8 | Single单域 |
| D | Narrus (reading阅读) | 70.5 | 66.1 | Single单域 |
| Single avg单域均值 | 71.8 | 68.6 | ||
| E | Finance × Diet财务×饮食 | 89.3 | 82.6 | Dual双域 |
| F | Finance × Mood财务×情绪 | 89.8 | 84.0 | Dual双域 |
| G | Diet × Mood饮食×情绪 | 88.9 | 84.3 | Dual双域 |
| H | Panoramic (all 4)全景 (四域) | 84.2 | 91.8 | Full全域 |
FINDING发现 Both models consistently show panoramic > single-domain. The cross-domain benefit is architecture-independent — it holds for both MoE (Qwen) and dense (GLM) models. 两个模型均一致显示全景 > 单域。跨域收益是架构无关的——MoE (Qwen) 和稠密模型 (GLM) 均成立。
| Drift Class漂移类型 | Qwen IIR | GLM IIR | Crisis F1危机 F1 |
|---|---|---|---|
| Normal | 1.19 | 1.39 | 0.191 |
| Unexpected | 1.16 | 1.27 | 0.444 |
| Severe | 1.17 | 1.38 | 0.560 |
| Config | Accuracy准确性 | Depth深度 | Novelty新颖性 | Action.可操作性 | Integration整合 |
|---|---|---|---|---|---|
| Single (A-D)单域 (A-D) | 4.18 | 3.96 | 3.09 | 4.44 | 1.72 |
| Dual (E-G)双域 (E-G) | 3.65 | 4.28 | 3.83 | 3.60 | 3.97 |
| Panoramic (H)全景 (H) | 4.07 | 4.48 | 3.95 | 2.02 | 4.52 |
PARADOX悖论 Panoramic analysis achieves the highest depth (4.48) and integration (4.52) but the lowest actionability (2.02). Richer data produces broader insights but less specific recommendations. Solution: two-stage output (insight → action plan). 全景分析获得了最高的深度 (4.48) 和整合 (4.52)，但最低的可操作性 (2.02)。更丰富的数据产生更广泛的洞察，但建议的具体性下降。解决方案：两阶段输出（洞察 → 行动计划）。
Green border = new V3 users. Event distribution: 6 Normal / 5 Unexpected / 3 Severe 绿色边框 = V3 新增用户。事件分布：6 普通 / 5 意外 / 3 剧烈
| Dimension维度 | V2 | V3 |
|---|---|---|
| Users用户 | 10 (ages 22-45) | 14 (ages 15-71) |
| Social diversity社会多样性 | Urban professionals城市白领 | Student, elderly, caretaker, disconnected 学生、独居老人、看娃老人、断亲青年 |
| Crisis detection危机检测 | Qualitative only仅定性描述 | L3 F1 = 0.77 |
| Models模型 | Qwen3.5 (1 family) | Qwen + GLM (2 families) |
| Evaluation评估方法 | LLM-as-Judge (Claude) | 3 methods: self-judge + expert + crisis 3 种方法：自评 + 专家 + 危机 |
| IIR | 1.48x (Opus judge) | 1.17-1.34x (self-judged) |
| Focus重点 | Technical feasibility技术可行性 | Social validity社会有效性 |