KK Effect Reports

REPORT 01 Repeatability 重复性

Does the same conversation produce the same model twice?

同一段对话，重复建模会得到同一个模型吗？

One fixed 20-turn conversation, modelled three times independently. Stability score 0.7333, mean per-dimension σ 13.33, worst-case range 55 points.

同一段 20 轮对话，独立建模 3 次。稳定性分 0.7333，逐维平均 σ 13.33，最坏极差 55 分。

2026-09-12 · N=3 runs2026-09-12 · N=3 次

REPORT 02 Honest Confidence 置信诚实性

74 real conversations → 35 models: what 100% coverage really means

74 场真实对话 → 35 份模型：100% 覆盖率真正意味着什么

All 385 trait slots were filled — but 40% of models returned nothing except an explicit "insufficient conversation data" flag. Coverage is not quality.

385 个特质槽位全部填充——但有 40% 的模型除了显式的“对话数据不足”标记外什么都没给出。覆盖率不等于质量。

2026-09-12 · 74 sessions / 35 models2026-09-12 · 74 会话 / 35 模型

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？

Text is present in 21/21 samples. Voice prosody in 12/21. Face signal in 1/21. We publish the gap instead of hiding it.

文本通道 21/21 样本到位。语音韵律 12/21。画面信号 1/21。我们公布这个缺口，而不是把它藏起来。

2026-09-12 · 21 training samples2026-09-12 · 21 份训练样本

Why these three reports

为什么是这三篇

A personality score is a measurement, and a measurement is only as good as the instrument that produced it. Before asking anyone to trust a Human Model, we owe three answers about the instrument itself: does it repeat (Report 01), does it know when it does not know (Report 02), and does it actually receive the signals it advertises (Report 03). Together these are the instrument's calibration sheet — its repeatability, its honesty, and its channel integrity.

人格分数是一种测量，而测量的可信度取决于产生它的仪器。在要求任何人信任一份人类模型之前，我们欠仪器本身三个答案：它是否可重复（报告 01）、它是否知道自己不知道（报告 02）、它是否真的收到了它所宣称的信号（报告 03）。合起来，这就是这台仪器的标定表——重复性、诚实性、通道完整性。

Data provenance

数据来源

All three reports read the same first-party stores, snapshotted on 2026-09-12. Nothing is synthetic, and no record was excluded to improve a figure.

三篇报告读取同一批第一方数据，快照时间 2026-09-12。数据均非合成，且未为改善任何指标而剔除任何记录。

Store数据文件	What it holds内容	Rows记录数
`mirror_models.json`	One model per completed session, 11 traits each每次完成的会话产出一份模型，每份 11 维	35
`mirror_sessions.json`	Raw session transcripts (text / voice / video)原始会话记录（文本 / 语音 / 视频）	74
`mirror_training_samples.jsonl`	Per-sample multimodal signal frames逐样本的多模态信号帧	21
`stability_baseline.json`	Repeat-run probe output (Report 01)重复运行探针的输出（报告 01）	N=3
`mirror_analytics.json`	Instrumented client/server events埋点的客户端 / 服务端事件	551

How to reproduce

如何复现

Report 01 is produced by a probe that ships with the service. It replays one fixed transcript through the full modelling endpoint N times, then writes a baseline file that later model changes are compared against.

报告 01 由随服务一起部署的探针脚本产出。它把一段固定文稿重复送入完整建模端点 N 次，然后写出一个基线文件，供后续模型改动做退化对比。

On the application server:在应用服务器上：
python _stability_probe.py --n 3 --tag my-baseline
Reads data/mirror_training_samples.jsonl, writes data/stability_baseline.json. Use --cleanup to delete the probe models it creates. Report 02 and 03 are pure aggregations over the JSON stores above — no probe needed. 读取 data/mirror_training_samples.jsonl，写出 data/stability_baseline.json。加 --cleanup 可删除探针创建的模型。报告 02 与 03 是对上述 JSON 存储的纯聚合，无需探针。

Desensitization

脱敏口径

Only aggregate statistics are published. No transcript text, no user identifier, no session identifier appears in any report.
只发布聚合统计量。任何报告中都不出现对话原文、用户标识或会话标识。
Sample references are shown as a short hash prefix (e.g. 050bdc34) so a figure can be re-derived without exposing the conversation.
样本引用以短哈希前缀表示（如 050bdc34），以便在不暴露对话的前提下重新推导数字。
Consent is a precondition, not an afterthought: each session records camera / audio / analysis / storage consent before any signal is stored.
授权是前置条件而非事后补办：每场会话在任何信号落库之前先记录摄像头 / 音频 / 分析 / 存储授权。

What these reports do not claim

这些报告不宣称什么

No accuracy claim. None of the three reports measures whether a trait score is correct — that requires an external criterion we do not yet have. They measure whether the instrument is stable, honest, and properly fed. 不宣称准确率。 三篇报告都不衡量特质分数是否正确——那需要一个我们目前尚不具备的外部效标。它们衡量的是这台仪器是否稳定、是否诚实、是否被正确喂入信号。

Sample sizes are small (35 models, 21 samples, N=3 reruns). These are calibration readings, not population claims — every report states this in its Limitations section.
样本量很小（35 份模型、21 份样本、3 次重复）。这些是标定读数，不是总体结论——每篇报告都在“研究局限”一节写明这一点。
Where a number looks bad (40% fallback rate, 1/21 face coverage), it is published as-is. A report that only contains good news is marketing, not measurement.
凡数字不好看之处（40% 降级率、1/21 画面覆盖率），一律原样公布。只含好消息的报告是营销，不是测量。

Back to Research返回研究

KK 效果实测报告