KK Research / Effect Reports / Report 03
Signal Coverage 通道到位率Text, voice, and face: how many of the three channels actually arrive?
文本、语音、画面：三路信号实际到了几路？
Across 21 first-party training samples, text is present in every one, voice prosody in 12, and face signal in exactly 1. Multimodal is, today, text plus half of voice — and we would rather publish that than advertise a capability we cannot deliver.
在 21 份第一方训练样本中，文本通道每份都到位，语音韵律 12 份，画面信号恰好 1 份。多模态在今天是「文本 + 半个语音」——我们宁愿公布这一点，也不去宣传做不到的能力。
21/21
Text channel文本通道
12/21
Voice prosody语音韵律
1/21
Face signal画面信号
29
Mic grants麦克风授权
1
Camera grants摄像头授权
The Challenge
挑战
Multimodal understanding is one of the four differentiators this product claims, and it is the easiest one to fake. A landing page can promise "reads your voice and your face" and no visitor will ever be able to check. Internal engineering has the opposite problem: it knows exactly which arrays are empty. This report closes that gap by asking a deliberately unglamorous question — for each signal the product implies it uses, how many stored samples actually contain it?
多模态理解是本产品宣称的四项差异化之一，而它恰恰是最容易造假的一项。落地页可以承诺「读取你的声音与表情」，而没有访客能验证。内部工程则面对相反的问题：它精确知道哪些数组是空的。本报告用一个刻意朴素的提问来弥合这个落差——对产品暗示会使用的每一路信号，实际有多少份落库样本真的包含它？
The rule we hold ourselves to: a capability is described on the site only at the coverage level this report can demonstrate. Where coverage is near zero, the site says the capability is not offered — rather than describing it in the present tense and letting users assume it works.
我们自我约束的规则： 站点上对一项能力的描述，只写到本报告能证明的覆盖水平。覆盖率接近零的能力，站点就明说「未提供」——而不是用现在时描述它、让用户默认它已经工作。
The Approach
做法
- Start from the stored samples. Every record in the training-sample store is examined — 21 of them — and each optional signal field is tested for presence and for non-empty content. A field holding an empty array counts as absent.从落库样本出发。 逐一检查训练样本库中的每条记录——共 21 条——对每个可选信号字段同时检验「字段存在」与「内容非空」。字段存在但数组为空，计为缺失。
- Count frames, not just fields. Where a signal is present we also record how many frames it carried, because a two-frame "prosody track" is technically present and practically useless.数字段，也数帧数。 对已到位的信号，另外记录其携带的帧数——因为一条两帧的「韵律轨」在技术意义上存在，在实用意义上无用。
- Cross-check against consent events. Capture cannot exceed permission: if the camera was authorised once, no more than one session can possibly hold face data. Comparing the two independently derived counts is how we detect a silent capture failure.与授权事件交叉核对。 采集量不能超过授权量：如果摄像头只被授权过一次，那么持有画面数据的会话不可能超过一场。把这两个独立来源的数字对照，是发现「静默采集失败」的方法。
- Report the funnel, not the peak. The headline figures are end-to-end arrival rates at the store. We do not report "the capture pipeline supports X" — only what survived all the way into a stored sample.报告漏斗，而非峰值。 核心数字是「最终抵达存储的端到端到位率」。我们不报告「采集管线支持 X」——只报告一路活到落库样本的量。
The Result
结果
The three channels are not three-quarters full, or half full. They are one full channel, one half channel, and one channel that has essentially never run.
三路通道不是「八分满」或「一半满」。它们是：一路满的、一路半的、以及一路基本从未跑起来的。
| Channel通道 | Samples with signal含信号的样本 | Arrival到位率 | Frame depth帧深 | |
|---|---|---|---|---|
| Text turns (baseline)文本轮次（基线） | 21 / 21 | 100% | full transcript完整文稿 | |
| Voice prosody (energy / f0 / voiced)语音韵律（能量 / 基频 / 有声） | 12 / 21 | 57.1% | 2 – 121 frames (mean 50)帧（均值 50） | |
| Face signal (smile / blink / mouth / brow)画面信号（微笑 / 眨眼 / 口型 / 眉） | 1 / 21 | 4.8% | 6 frames帧 | |
| Turn timing (reply latency, lengths)轮次时序（回复延迟、长度） | 12 / 21 | 57.1% | 35 turn pairs对轮次 |
The consent cross-check
授权交叉核对
Camera coverage is not merely low — it is low for a discoverable reason. Only one camera-permission event exists in the entire instrumented history, and exactly one sample holds face data. Permission and capture agree, so this is a demand-side absence, not a broken pipeline.
画面覆盖率低，而且低得原因可查。在全部埋点历史中，摄像头授权事件只出现 1 次，而持有画面数据的样本也恰好 1 份。授权与采集相互吻合，所以这是需求侧的空缺，不是管线故障。
| Permission event授权事件 | Times raised出现次数 | Granted通过 |
|---|---|---|
| Microphone麦克风 | 29 | 29 |
| Camera摄像头 | 1 | 1 |
Two further gaps are visible in the same pass. First, sessions declared as video mode outnumber samples containing any face data by 5 to 1 — the mode exists in the product before the signal does. Second, only 21 of 74 sessions produced a training sample at all (28.4%), so even the text channel is not fully harvested.
同一次核对还暴露两处缺口。其一，声明为视频模式的会话数，是含任何画面数据的样本数的 5 倍——在产品里「模式」先于「信号」存在。其二，74 场会话中只有 21 场产出了训练样本（28.4%），所以即便是文本通道也并未全量收割。
A note on turn timing
关于轮次时序的说明
The turn-timing signal records the gap between one turn ending and the next beginning. Median 7.5 s, mean 17.6 s, 90th percentile 48.5 s, maximum 49.7 s. This is not system latency: it includes the user's own thinking and typing time, and no attempt has been made to separate the two. We label it as a turn gap rather than a response time so that it cannot be quoted as a latency benchmark.
轮次时序信号记录的是「上一轮结束」到「下一轮开始」之间的间隔。中位数 7.5 秒，均值 17.6 秒，90 分位 48.5 秒，最大 49.7 秒。这不是系统延迟： 其中包含用户自己的思考与打字时间，且我们没有尝试把两者分离。我们把它标注为「轮次间隔」而非「响应时间」，以免它被引用为时延基准。
What It Means
含义
- "Multimodal" is currently an overstatement. The honest description of the engine today is text-first with an optional voice-prosody layer on roughly half of sessions. Face signal exists in the schema but not in the data.
- 「多模态」目前是夸大。 对引擎当下状态的诚实描述是：以文本为主，约半数会话可选叠加语音韵律层。画面信号在数据结构里存在，在数据里不存在。
- No face or expression claim is made on the site. Because coverage is 4.8%, pages describe face signal as a capability that is defined but not deployed. Present-tense marketing language for it would be false.
- 站点上不作任何面部或表情判断的宣称。 因为覆盖率仅 4.8%，页面把画面信号描述为「已定义但尚未部署」的能力。对其使用现在时的营销措辞将是不实陈述。
- The voice layer is real but partial. 12 sessions carry genuine prosody frames (up to 121 of them), so voice features can be built on real data — but any claim about them must be scoped to voice-mode sessions, not to all users.
- 语音层是真实存在的，但是部分的。 12 场会话带有真实的韵律帧（最多 121 帧），因此语音特征可以建立在真实数据上——但相关结论必须限定在语音模式会话范围内，不能推广到全部用户。
- Harvest coverage is the cheaper win. Raising sample coverage from 28.4% toward 100% costs engineering, not new data collection, and would roughly triple every text-derived figure in Reports 01 and 02.
- 提高收割覆盖率是成本最低的改进。 把样本覆盖率从 28.4% 提升到接近 100% 花的是工程量，而不是新的数据采集，且会让报告 01、02 中所有由文本推导的数字大约翻三倍。
The honest framing: one channel is complete, one is half-complete, one is unused. Stating that precisely is worth more than a plausible-sounding multimodal claim, because it tells a prospective user, a partner and our own roadmap exactly where the next unit of work pays off.
诚实的表述： 一路通道完整，一路半完整，一路未启用。精确说明这一点，比一句听起来合理的多模态宣称更有价值——因为它同时告诉潜在用户、合作方和我们自己的路线图：下一份投入应该花在哪里。
Reproduce it
复现方式
Pure aggregation over two stores — no probe required. For each record in
data/mirror_training_samples.jsonl, test whether voice_prosody, video_signal_seq and turn_timing are present and non-empty. Permission counts come from data/mirror_analytics.json, filtered on event_type equal to mic_permission and camera_permission.
对两个存储的纯聚合，无需探针。对 data/mirror_training_samples.jsonl 中的每条记录，检验 voice_prosody、video_signal_seq、turn_timing 是否存在且非空。授权计数来自 data/mirror_analytics.json，按 event_type 等于 mic_permission 与 camera_permission 过滤。
Limitations
研究局限
- 21 samples is a very small base. A single percentage point on this page is a fraction of one sample. The qualitative conclusion (face signal is absent) is robust; the exact arrival rates are not yet stable.
- 21 份样本是极小的基数。 本页每 1 个百分点都不到一份样本。定性结论（画面信号缺失）是稳健的；精确到位率尚不稳定。
- Consent is opt-in, so absence can be legitimate. Low camera coverage reflects users declining, which is the intended behaviour. This report measures supply, not whether users were asked fairly.
- 授权是选择加入的，因此缺失可能是正当的。 摄像头覆盖率低反映的是用户拒绝授权，这正是预期行为。本报告衡量的是供给侧，而非用户是否被公平地询问过。
- The training store is not the only data path. Signals may exist in session records without a corresponding training sample. The 21/74 harvest rate therefore understates total captured signal, and a reconciliation between the two stores is still outstanding.
- 训练库不是唯一的数据路径。 信号可能存在于会话记录中而没有对应的训练样本。因此 21/74 的收割率低估了总采集信号量，两个存储之间的对账仍未完成。
- Frame counts are not quality. We report that prosody frames exist and how many; we do not assess whether the frames are clean, aligned, or sufficient for any particular acoustic feature.
- 帧数不等于质量。 我们报告韵律帧是否存在及其数量；并不评估这些帧是否干净、对齐，或对任何特定声学特征是否充分。
What would prove this report wrong Falsifiable
什么能推翻本报告 可证伪
- If face data is found in session records that never produced a training sample, then face coverage is materially higher than 4.8% and this report's headline must be corrected.
- 若在从未产出训练样本的会话记录中发现了画面数据，则画面覆盖率显著高于 4.8%，本报告的核心结论必须更正。
- If the camera-permission counter is shown to be under-instrumented (permissions granted but never logged), then the "demand-side absence" explanation is wrong and a silent capture failure becomes the leading hypothesis.
- 若能证明摄像头授权计数埋点缺失（已授权但未记录），则「需求侧空缺」的解释不成立，「静默采集失败」将成为首要假设。
- If a deliberate test run with camera enabled for ten sessions produces ten samples with face data, the pipeline is proven healthy and the low coverage is purely a traffic property.
- 若刻意在开启摄像头的条件下跑十场会话，产出十份含画面数据的样本，则管线被证明健康，低覆盖率纯属流量属性。
Related: Report 01 — Repeatability · Report 02 — Honest confidence
相关： 报告 01 — 重复性 · 报告 02 — 置信诚实性