Prism

Cross-Domain Insight Emergence 跨域洞察涌现

Same model, same users, only the data input changes. Full panoramic integration (Config H) consistently outperforms any single-domain or dual-domain configuration. 相同模型，相同用户，只改变数据输入。全景整合（配置 H）在所有用户上持续优于任何单域或双域配置。

IIR per User (Full Panoramic / Single-Domain Average) 各用户 IIR（全景整合 / 单域均值）

IIR ≥ 1.0 indicates cross-domain integration adds value. Target: ≥ 1.5x IIR ≥ 1.0 表示跨域整合产生增量价值。目标：≥ 1.5x

Diminishing Returns Beyond 9B 9B 以上收益递减

The jump from 2B to 9B (+14.9) is nearly 3x the gain from 9B to 35B (+5.6). A 9B model captures most of the insight quality, validating the "medium model" thesis. 从 2B 到 9B 的提升 (+14.9) 是 9B 到 35B (+5.6) 的近 3 倍。9B 模型已捕获大部分洞察质量，验证了"中等模型"命题。

Real-Time Inference on Consumer Hardware 消费级硬件上的实时推理

All models run locally on Apple Silicon via MLX. No cloud, no API calls, no data leaves the device. 所有模型通过 MLX 在 Apple Silicon 上本地运行。无云端、无 API 调用、数据不离开设备。

Three-Tier Device Topology 三层设备拓扑

Vertical apps collect data naturally. The home server federates summaries — never raw data — across devices to produce panoramic insights. 垂直应用自然采集数据。家庭服务器联邦式聚合摘要 — 从不接触原始数据 — 跨设备生成全景洞察。

Zero Data Leakage by Design 架构层面零数据泄漏

Raw data never leaves the originating device. Only compressed summaries traverse the local network. The federation protocol achieves 125.5x compression while preserving insight quality. 原始数据从不离开产生它的设备。只有压缩摘要在局域网内传输。联邦协议实现 125.5x 压缩比，同时保持洞察质量。

LLM-as-Judge Scoring LLM-as-Judge 评分体系

Each model-generated insight is scored by Claude Opus 4.6 (Anthropic) on four dimensions, each 0–25 points, for a total of 0–100. The judge receives the user's full data context and the generated insight, producing scores with detailed justifications. 每条模型生成的洞察由 Claude Opus 4.6（Anthropic）在四个维度上评分，每维度 0–25 分，满分 100。评审模型接收用户完整数据上下文和生成的洞察，给出带有详细理由的评分。

Methodology reference: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (Zheng et al., NeurIPS 2023) 方法论参考：Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena（Zheng et al., NeurIPS 2023）

Early Warning for Life Crises 生活危机的预警系统

Cross-domain data integration enables detection of cascading life crises that single-domain analysis cannot see — from financial stress spiraling into health emergencies, to burnout escalating into medical errors. 跨域数据整合使系统能够检测到单域分析无法发现的连锁生活危机 — 从经济压力演变为健康紧急事件，到过度疲劳升级为医疗事故。

Privacy is essential: users are more likely to maintain honest records when they trust that no external party will see their data. 隐私至关重要：当用户确信没有外部方能看到数据时，他们更愿意保持真实的记录。

10 Users Spanning Income & Life Stages 覆盖不同收入与人生阶段的 10 位用户

Each user has 90 days of synthetic data across 4 apps with an injected life event to simulate temporal drift. 每位用户拥有 4 个应用、90 天的合成数据，并注入生活事件以模拟时序漂移。