Pebble 的“温柔”背后，是一套认真的工程与临床方法，以及对前沿研究的持续跟读。这一页讲讲：我们怎么做、正在探索什么，以及我们站在哪些研究的肩膀上。Behind Pebble's gentleness is serious engineering and clinical method — and a habit of reading the frontier closely. Here's how we work, what we're exploring, and the research we build on.
以下每一项，都对应一个明确的临床或工程判断，而不是功能堆叠——并且都已经在产品里运行。Each one reflects a deliberate clinical or engineering decision — not a pile of features — and each is live in the product today.
每一轮对话都由个案概念化驱动——先理解“此刻正在发生什么”，再决定怎么回应，而不是逐句应答。Every turn is driven by a case formulation — we first understand what's happening right now, then choose how to respond, rather than replying line by line.
先判断你此刻更需要被倾听、想要方法，还是在求助，再动态决定回应的方式与边界。We first sense whether you need to be heard, want a method, or are reaching for help — then route the response accordingly.
信件、文字、语音共用同一套人格、记忆与安全底线；从打字切到说话，对话不断线。Letters, text, and voice share one persona, memory, and safety floor — switch from typing to talking without losing the thread.
分层危机识别 + 输出守卫 + 隐私优先。安全是贯穿系统的架构，不是事后加的过滤器。Layered crisis detection, output guarding, privacy-first — safety runs through the architecture, it isn't a filter bolted on after.
每一次措辞改动都在真实模型上评测、由独立评委盲评、过回归门。质量与安全，我们都不做取舍。Every change in wording is tested on real models, blind-reviewed by independent judges, and gated by regression tests. We trade off neither quality nor safety.
记得你上次聊到哪里，并追踪状态随时间的变化——让陪伴有连续性，而不是每次从零开始。Remembers where you left off and tracks how your state shifts over time — so support has continuity, not a cold start each visit.
情绪往往藏在“怎么说”里，而不只在“说了什么”。Pebble 的语音理解在语义之外引入副语言韵律建模——从语速、停顿、音高与能量的起伏中，提取文字无法承载的情绪信号，让语音陪伴从“能对话”走向“听得懂”。该能力目前处于灰度内测，并正探索向心理健康、公益热线、对话质检等机构合作方开放私有化部署与接口评估。Emotion often lives in how something is said, not only in the words. Beyond semantics, Pebble's voice understanding adds paralinguistic prosody modeling — reading the emotional signal in pace, pauses, pitch and energy that text alone can't carry, so voice companionship moves from speaking to understanding. The capability is in limited gray-scale beta, with private, on-premise deployment and API evaluation opening to partners across mental health, helplines, and conversation quality assurance.
语音韵律感知 · 灰度内测中Voice-prosody sensing · limited beta这些是我们正在攻的方向——还没全部进产品，但每一项都朝着“更懂人、更负责任”的目标走。These are the problems we're actively working on — not all in the product yet, but each one aimed at care that understands more, and is more responsible.
这里列出对我们的临床与工程思考有影响的部分公开文献——从认知行为疗法的循证基础，到大模型心理治疗的最新架构与评测方法。A selection of the public literature that shapes our clinical and engineering thinking — from the evidence base of CBT to the newest architectures and evaluation methods for LLM-assisted care.