为什么是这个产品，证据凭什么这么说。

0. 一句话与全图The chain in one picture

一句话：把每一个关键判断变成一个概率，让每条证据按"强度 × 质量"去推动这个概率，最后用事先写死的门槛做决定。设计思维负责流程（共情 → 定义 → 构思 → 原型 → 测试），贝叶斯层负责"评估"这一步不靠感觉（Fairfield & Charman 2017 [11]；Humphreys & Jacobs 2015 [12]）。

In one sentence: turn every key judgement into a probability, let each piece of evidence push that probability in proportion to its strength × quality, and decide against thresholds written down in advance. Design thinking owns the process (empathize → define → ideate → prototype → test); the Bayesian layer makes the "evaluate" step run on something better than gut feel (Fairfield & Charman 2017 [11]; Humphreys & Jacobs 2015 [12]).

1. 为什么是企业创新，而不是又一个创业项目Why this is corporate innovation, not a new venture

整篇论文的框架是内部人行动研究（insider action research，Coghlan & Brannick 2019 [3]）：创始人—研究者在自己的公司里"诊断 → 计划 → 行动 → 评估 → 学习"。这不是为了方便，而是决定了方法论必须自带偏差控制（§10 与 METHODOLOGY §7）：反身性、预注册、由研究助理执行现场、分析时对条件盲化。

The frame for the whole dissertation is insider action research (Coghlan & Brannick 2019 [3]): the founder-researcher diagnoses, plans, acts, evaluates and learns inside their own firm. That is not a convenience — it is what forces bias control into the method itself (§10 and METHODOLOGY §7): reflexivity, pre-registration, a research assistant running the sessions, and the analyst blinded to condition labels.

公司已经有一台被验证过的引擎：需求信号 → AI 生产 → 被测量的曲库（KDP 图书试点跑通过）。这一轮行动研究是把同一台引擎接到一个新的需求上——音乐用于身心健康。它是相邻扩张，不是全新赌局：新的客户需求（健康）× 被改造过的产品（音乐），建立在既有资产上。

The firm already runs a proven engine: demand signal → AI production → measured library (validated in a books pilot). This action-research cycle applies the same engine to a new need — music for wellbeing. It is an adjacent move, not a fresh bet: a new customer need (health) × an adapted product (music), standing on assets the firm already owns.

资产盘点与它推动的信念 · Asset inventory and the belief each asset moves

四条锁定的决定（框架文件 2026-07-09 更新）：海外优先（美/欧付费意愿更高、话术更干净），中国只作低成本探针；试点硬件 = Oura（创始人自有）+ Apple Watch；本阶段预算 $10–20k / ≤ 3 个月；监管路线 = 现在只做"一般健康（wellness）"定位，数字疗法路线推迟到融资之后。

Four locked decisions (framework document, updated 9 Jul 2026): overseas-first go-to-market (better willingness to pay in the US/EU, cleaner claims), with China as a low-cost probe only; pilot hardware = Oura (founder-owned) + Apple Watch; a ring-fenced budget of $10–20k over ≤ 3 months; and a regulatory strategy of general-wellness positioning now, with the digital-therapeutics route deferred until after funding.

2. 设计思维如何组织证据Design thinking as the process, not as decoration

设计思维在这里只做一件事：规定发散与收敛的顺序（d.school 五模式 / 双钻，Brown 2008 [1]；Liedtka 2018 [2]）。真正让结论站得住的是每个模式的产出物——每一格都必须交出一个可被检查的文件，而不是一次工作坊的感受。方法架构本身是收敛式混合方法、量化为核心的序贯设计（Creswell & Plano Clark 2018 [4]）。

Design thinking does exactly one job here: it imposes the order of divergence and convergence (d.school five modes / Double Diamond; Brown 2008 [1]; Liedtka 2018 [2]). What makes the conclusions defensible is the artefact each mode must hand over — an inspectable file, not the feeling in a workshop. The method architecture underneath is a convergent mixed-methods design with a sequential quantitative core (Creswell & Plano Clark 2018 [4]).

3. 共情：六条证据流各说什么Empathize — six evidence streams and what each is allowed to claim

混合方法研究有两种经典翻车方式：一条证据被计两次分；一条弱证据说了它没资格说的话。所以在收集之前先分配角色（METHODOLOGY §1.1）：文献只给起点信念（先验）；访谈、网络民族志、焦点小组只负责提出假设和解释原因（质性，权重要打折）；问卷负责量化比例与价格；试点负责因果检验。

Mixed-methods research has two classic failure modes: the same evidence counted twice, and weak evidence making a claim it has no standing to make. So roles are assigned before collection (METHODOLOGY §1.1): literature sets the starting belief (the prior); interviews, netnography and focus groups generate hypotheses and interpret (qualitative, heavily discounted); the survey quantifies shares and prices; the pilot tests causality.

六条流分别说了什么What each stream actually said

① 文献（先验）。音乐聆听降低焦虑在多个独立元分析里方向一致，但效应量跨度很大（d ≈ −0.30 到 −0.97），且最严格的评审给底层试验打了"低质量"。最大且最严谨的一份是 26 项围手术期 RCT 的 Cochrane 综述（N = 2,051），状态焦虑降低 −5.72 STAI-S 单位（Bradt, Dileo & Shim 2013 [33]）；更宽口径的压力元分析（104 项 RCT，N = 9,617）给出心理指标 d = .545、生理指标 d = .380，其中心率是最强的生理标记（d = .456）（de Witte et al. 2020 [34]）。HRV 方向上证据一致：慢速音乐提升迷走张力（Bernardi, Porta & Sleight 2006 [38]；Mitrovic & Paladin 2026 [39]）。但自评与生理并不总是同步：Panteleeva et al. 2018 [37] 发现自评显著（d = −.30）而心理生理不显著。另有两条同样计入先验：Harney et al. 2023 [35] 在 21 项对照研究上给出 d = −0.77；Lassner et al. 2024/2025 [36] 的头对头综合发现被动聆听与有治疗师的音乐治疗效应相当（SMD = 0.47，k = 23，n = 1,084），但 GRADE 极低，且效应在随访时不维持。最接近闭环产品的间接证据是 HRV 生物反馈的元分析（Goessl, Curtiss & Hofmann 2017 [40]，g = 0.81），但它是呼吸导引而非音乐驱动——这个泛化落差必须明说，不能假装它支持我们。

① Literature (the priors). Music listening reduces anxiety consistently in direction across independent meta-analyses, but the pooled effect ranges widely (d ≈ −0.30 to −0.97) and the strictest reviewers grade the underlying trials low quality. The largest and most rigorous is a Cochrane review of 26 perioperative RCTs (N = 2,051): −5.72 STAI-S units (Bradt, Dileo & Shim 2013 [33]). A broader stress meta-analysis (104 RCTs, N = 9,617) gives d = .545 psychological and d = .380 physiological, with heart rate the strongest physiological marker (d = .456) (de Witte et al. 2020 [34]). On HRV the direction is consistent — slow music raises vagal tone (Bernardi, Porta & Sleight 2006 [38]; Mitrovic & Paladin 2026 [39]) — but self-report and physiology do not always move together (Panteleeva et al. 2018 [37]: the self-report effect is significant at d = −.30 while the psychophysiological one is not). Two further rows enter the prior: Harney et al. (2023 [35]) report d = −0.77 across 21 controlled studies, and Lassner et al. (2024/2025 [36]) find passive listening roughly equivalent to therapist-led music therapy (SMD = 0.47, k = 23, n = 1,084) at GRADE very low, with the effect not maintained at follow-up. The closest indirect evidence for a closed loop is a meta-analysis of HRV biofeedback (Goessl, Curtiss & Hofmann 2017 [40], g = 0.81) — but that is breathing-paced rather than music-driven, and the generalisation gap has to be stated rather than glossed over.

② 专家访谈。王医生（老年痴呆专科）给了两个对收敛至关重要的发现：熟悉的老歌有效，专门制作的"治疗音乐"无效——这直接削弱了"AI 生成音乐做痴呆"的路；以及中国养老场景的变现极难。

② Expert interviews. Dr. Wang (dementia specialist) produced two findings that were decisive for convergence: familiar music from a patient's youth works, purpose-made "treatment music" does not — which undercuts an AI-generated-music play for dementia — and monetising elderly care in China is very hard.

③ 用户访谈（n = 3 急性焦虑者）——本项目最重要的一条反证。三位受访者各自独立地拒绝有情绪起伏的音乐，要求的是相反的东西：无特征、无旋律、无歌词、情绪中性、低信息量的声音。他们用了几乎相同的语言：I-03 说要"一堵柔软的墙"（"一种完全'空洞'、没有情绪起伏、像一堵柔软的墙一样的声音"），I-04 要"让脑电波平缓下来的声音，不要有旋律"，I-05 要"既能填补空间空白，又能带来绝对平静的声音"（他目前的替代方案是把空气净化器开到最大）。对现有中性音频的抱怨非常具体：白噪音/自然音有刺耳频率（鸟叫、高音钢琴）；脑波音轨的低频嗡鸣引起恶心；"治愈系"歌单让人觉得虚伪并触发逆反；任何有节拍的东西会让心率跟着节奏走——方向反了。

③ User interviews (n = 3 acute sufferers) — the single most important piece of disconfirming evidence in the project. All three independently rejected emotionally engaging music and asked for the opposite: featureless, non-melodic, lyric-free, emotionally neutral, low-information sound. They converged on almost identical language — a "soft wall" (I-03), "sound that flattens my brainwaves, no melody" (I-04), "something that fills the emptiness and brings absolute calm" (I-05, whose current substitute is an air purifier on maximum). Their complaints about existing neutral audio are specific: white noise and nature tracks contain sharp jarring frequencies; brainwave tracks have a low hum that causes nausea; "healing" playlists feel fake and trigger reactance; anything with a beat makes the heart entrain to the rhythm — the wrong direction.

④ 网络民族志（n = 200 行）。它回火了访谈：在 Calm / Endel / Brain.fm 的应用商店与 Product Hunt 评论里，两种偏好同时存在——有人抱怨 Endel/Brain.fm"不是我期待的那首歌 / 只是 lofi"，也有人明确欣赏它"没有 hook 和歌词"。所以正确的读法不是"中性声音赢了"，而是"存在两个不同的活儿/两个分群"。这条流还贡献了 A3 的关键负面信号：10 行关于生物指标态度的记录全都在讨论读数可不可信，没有一行在讨论因此改变了什么行为——这就是"行动缺口"。

④ Netnography (200 coded rows). It tempered the interviews: across App Store and Product Hunt reviews of Calm, Endel and Brain.fm, both preferences coexist — some users fault Endel/Brain.fm for being "not the song I expected / just lofi", others explicitly value the absence of hooks and lyrics. So the correct reading is not "neutral sound wins" but "there are two distinct jobs". This stream also produced the key negative signal for A3: all ten rows coded for attitudes to biometrics are about whether the reading is trustworthy, and none is about a behaviour it changed — the "action gap".

⑤ 竞品扫描。Endel 是最接近的商业先例：用 Apple Watch 的实时心率（不是 HRV）加动作、时间、光线做闭环，2020 年拿下 Apple 年度手表 App，主新闻稿称 300 万月听众；它与 UMG / WMG 的艺人功能音乐合作，本身就是"纯生成音频满足不了需求"的揭示性证据。同时它唯一一份同行评议研究由 Endel 与 Arctop 资助、作者受雇于 Arctop，且测的是"专注"而不是焦虑（Haruvi et al. 2022 [43]）。Calm 与 Headspace 是收缩中的高消耗订阅方（2025 年收入与订阅数双降）；MedRhythms 拿到 FDA 突破性设备认定与 Medicare 报销，但适应症很窄（卒中/帕金森步态）；潮汐（¥218/年）代表中国结构性低付费。

⑤ Competitor scan. Endel is the closest commercial precedent: a closed loop on live Apple Watch heart rate (not HRV) plus motion, time of day and light; Apple's Watch App of the Year 2020; 3 million monthly listeners per its primary press release. Its UMG/WMG artist deals are themselves revealed evidence that purely generative audio under-satisfies. Its only peer-reviewed study is industry-funded, authored partly by Arctop employees, and measures "focus" rather than anxiety (Haruvi et al. 2022 [43]). Calm and Headspace are shrinking high-burn subscription incumbents (2025 revenue and subscriber declines); MedRhythms holds FDA Breakthrough Device designation and Medicare reimbursement but on a narrow indication; Tide (¥218/yr) represents structurally low Chinese willingness to pay.

⑥ 类别层面最有分量的一条量化发现来自用户情绪流：冥想类 App 全行业 30 天留存约 4.7%，平均终身使用 1–4 次（Creswell & Goldberg 2025 [41]）。这条直接压低了 A6 的先验，也解释了为什么产品设计必须围绕"零决策 + 可测量的结果"，而不是内容目录。

⑥ The heaviest single quantitative finding at category level comes from the user-sentiment stream: industry-wide 30-day retention for meditation apps is about 4.7%, with an average of 1–4 lifetime sessions (Creswell & Goldberg 2025 [41]). It is what holds the A6 prior down, and it is why the product is designed around zero decisions and a measured result rather than a content catalogue.

4. 定义：矩阵 v1 → v2，以及本项目的核心张力Define — the territory matrix, and the tension the thesis has to confront

v1 是一张六维加权评分表，把三个领域打成 4.55（焦虑）/ 2.65（ADHD）/ 1.80（老年痴呆）。对一次 Define 工作坊够用，但对方法论导向的论文太弱：它掩盖不确定性（4.55 是拍脑袋的中值），无法显示每条证据推动了多少，也不让读者做敏感性分析。v2 保留同样的六个准则和权重（.25 / .20 / .20 / .15 / .10 / .10），但把每一格换成三角分布（低 / 众数 / 高），把权重扰动 ±40% 并重新归一，跑 5,000 次蒙特卡洛，报告"排第一的概率"。

Version 1 was a six-criterion weighted score that ranked the three territories 4.55 (anxiety) / 2.65 (ADHD) / 1.80 (dementia). Adequate for a Define workshop, weak for a methodology-led dissertation: it hides uncertainty, cannot show how much each piece of evidence moved, and gives the reader no sensitivity analysis. Version 2 keeps the same six criteria and weights (.25 / .20 / .20 / .15 / .10 / .10) but replaces every cell with a triangular distribution (low / mode / high), perturbs the weights by ±40% with renormalisation, runs 5,000 Monte-Carlo draws and reports the probability of ranking first.

六个准则里权重最高的是临床证据强度（.25），它的打分有具体依据。老年痴呆方向：个体化音乐对激越的元分析给出中等效应（Pedersen et al. 2017 [46]，d = 0.61，12 项 RCT），但迄今最大的实用性 RCT（N = 463，54 家养老院）在主要激越结局上未见显著效应（McCreedy et al. 2021 [47]）——一个真实的矛盾；加上王医生的临床观察（熟悉的老歌有效，专门制作的治疗音乐无效），AI 生成音乐在这条路上没有位置。ADHD 方向：普通背景音乐对核心注意网络无效（Mendes et al. 2024 [48]）；经特殊调幅处理的音乐显示出更具体的获益，但用的是厂商自有刺激（Woods et al. 2024 [45]）；网络声量极大的棕噪音，在 2024 年的系统综述里零篇研究符合纳入标准（Nigg et al. 2024 [44]）。可达市场那一项同样有据：WHO 估计 2021 年全球有 3.59 亿人患焦虑障碍，只有 27.6% 得到治疗（WHO 2023 [49]）；中国全国流调的 12 个月患病率 5.0%、终身 7.6%（Huang et al. 2019 [50]），与美国口径相差约四倍——更可能是诊断口径与污名的差异，而不是真实差异。最后，整个商业品类的独立疗效证据都薄：只有 Calm 与 Headspace 有 RCT 覆盖，且 Headspace 自家试验中有一半带已披露的利益冲突（O'Daffer et al. 2022 [42]）。

The heaviest criterion is clinical-evidence strength (.25), and its scores are sourced. Dementia: a meta-analysis of individualised music finds a medium effect on agitation (Pedersen et al. 2017 [46], d = 0.61, 12 RCTs), but the largest pragmatic RCT to date (N = 463 across 54 nursing homes) finds no significant effect on the primary agitation outcome (McCreedy et al. 2021 [47]) — a genuine contradiction; together with Dr. Wang's clinical observation that familiar old songs work while purpose-made treatment music does not, it leaves AI-generated music with no role there. ADHD: ordinary background music does not engage the core attentional networks (Mendes et al. 2024 [48]); specially amplitude-modulated music shows a more specific benefit but on the vendor's own proprietary stimuli (Woods et al. 2024 [45]); and brown noise, despite enormous online enthusiasm, has zero studies meeting inclusion criteria in a 2024 systematic review (Nigg et al. 2024 [44]). Reachable market is sourced too: the WHO estimates 359 million people had an anxiety disorder in 2021 with only 27.6% receiving treatment (WHO 2023 [49]), while China's national survey reports 5.0% 12-month and 7.6% lifetime prevalence (Huang et al. 2019 [50]) — roughly a fourfold gap against US figures, most plausibly diagnostic and stigma-related rather than real. Finally, independent efficacy evidence across the whole commercial category is thin: only Calm and Headspace have any RCT coverage, and half of Headspace's own trials carry a disclosed conflict of interest (O'Daffer et al. 2022 [42]).

这里有一个方法论上的要点：v2 的重评分发生在共情证据之后，而且降低了自家方案的分数。资产契合从 4 降到 3，可穿戴协同从 5 降到 4——两处都是被自己的证据打下来的。这正是"Define 被 Empathize 更新过"的可见证据，而不是事先定好结论再补材料。

A methodological point sits here: the v2 re-score happened after the empathy evidence and it lowered our own option's score. Asset fit went 4 → 3 and wearable synergy 5 → 4 — both knocked down by our own findings. That is the visible evidence that Define was updated by Empathize, rather than a conclusion fixed in advance and backfilled.

核心张力：用户需要的，和公司资产擅长的，不是同一件事The core tension: what the user needs is not what the asset is good at

公司的不公平优势是已授权曲库 + 艺人关系 + AI 音乐生产——它指向"你熟悉的真歌的舒缓版"。急性焦虑用户的需求指向无特征的生成式声音——那是 Endel 的地盘，用不上曲库。这是整个项目最锋利的战略问题，论文的讨论章必须正面回答它。

The firm's unfair advantage is a licensed catalogue + artists + AI music production, which argues for "calm versions of songs you already love". The acute-anxiety need argues for featureless generative sound — Endel's turf, which does not lever the catalogue. This is the sharpest strategic question in the project and the discussion chapter has to meet it head-on.

本项目给出的解法是按唤醒状态分群，而不是二选一：急性 / 睡前 / 上台前用中性自适应声音（不需要曲库权利）；下班后的放松与情绪修复用真歌舒缓版（用得上曲库和艺人）。这个解法不是拍脑袋选的——它是被 §5 的问卷情景题与 §6 的盲听测试量化过的（急性场景选真歌 24%，放松场景 47%）。

The resolution taken here is segmentation by arousal state rather than a single content bet: neutral adaptive sound for acute, sleep-onset and pre-performance (which needs no catalogue rights); real-song calm arrangements for wind-down and mood repair (which does use the catalogue and the artists). This was not chosen by intuition — it was quantified by the survey scenarios in §5 and the blind listening test in §6 (real song chosen by 24% in the acute scenario, 47% in wind-down).

5. 问卷怎么设计、怎么打分Survey v2 — how it is built and how it scores

问卷 v2 用 5W1H 的框架分成七块，每一块对应一个构念、一套已发表的量表、一个统计量，以及一个在看到数据之前就写死的阈值：A 谁（分群）· B 为什么/何时/何地（GAD-2、PSS-4、触发、身心表现、当前替代方案、Ulwick 机会分）· C 什么内容有效（分状态情景选择、声学特征反应矩阵）· D 怎么证明（连接意愿、信任、行动缺口、结果价值、隐私）· E 为什么用/为什么不用（TAM、Kano、形态选择、相对优势、障碍）· F 多少钱（Van Westendorp、购买意向、付费模式）· G 质量与开放题（注意力检测、关键事件法）。

Survey v2 is built on a 5W1H frame in seven blocks. Each block maps to a construct, a published instrument, a statistic and a threshold fixed before the data was seen: A who (segmentation) · B why / when / where (GAD-2, PSS-4, triggers, somatic pattern, current alternatives, Ulwick opportunity scores) · C what works (state-dependent scenario choice, acoustic-feature reaction matrix) · D how to prove it (willingness to connect, trust, the action gap, value of a result, privacy) · E why use / why not (TAM, Kano, concept form, relative advantage, barriers) · F how much (Van Westendorp, purchase intent, payment model) · G quality and voice (attention check, critical-incident technique).

结果（按站点公布的口径）Results, as the site reports them

样本：投放 150，剔除 16 份注意力检测失败、0 份速答（中位作答 7.4 分钟），保留 n = 134；可穿戴拥有者 63（47%），美/欧 80、中国 41、其他 13。需求侧：GAD-2 阳性（≥ 3）38%（n = 51），PSS-4 均值 6.54/16，r(GAD-2, PSS-4) = 0.69；87% 每周至少一次"需要平静下来"的时刻，90% 每周至少一次用音乐/音频应对。发生时点集中在床上 57%、工作时 52%、傍晚 33%。当前"雇佣"的替代方案里音乐 61% 居首，刷手机 57% 紧随其后——真正的竞品是刷手机。ODI 只有一个结果越过 6 分门槛："停止思绪飞转" 6.4；"几分钟内平静下来" 5.8、"不要让情况更糟" 5.6、"知道它确实起效了" 5.5 紧贴门槛。

Sample: 150 fielded, 16 attention-check failures and 0 speeders removed (median completion 7.4 min), n = 134 kept; 63 wearable owners (47%), 80 US/EU, 41 China, 13 other. Need side: GAD-2 positive (≥ 3) 38% (n = 51); PSS-4 mean 6.54/16; r(GAD-2, PSS-4) = 0.69; 87% want a calming moment at least weekly and 90% already use music or audio at least weekly. The moments cluster in bed (57%) and at work (52%), then evening (33%). Among the alternatives currently "hired", music leads at 61% with scrolling right behind at 57% — the real competitor is the phone. Only one desired outcome crosses the ODI threshold of 6: "stop racing thoughts" at 6.4; "calm down within minutes" (5.8), "not make things worse" (5.6) and "know it is actually working" (5.5) sit just under it.

内容侧（A1 的直接检验）：在急性情景里选"我爱的歌的舒缓版"的只有 24%［17–31%］，选"无特征的中性声音"的 46%；在放松情景里两者翻转为 47%［39–56%］与 19%。同一批人在两个情景下的选择差异显著（McNemar χ² = 14.29，p < 0.001，b = 16、c = 47）。GAD-2 阳性子群在急性情景下更极端：真歌 16%、中性 49%。C4 的声学反应给出了中性床的规格：让情况变糟的是突兀/尖锐声 −1.5、强节拍 −1.1、歌词 −0.8、低频嗡鸣 −0.7；有帮助的是非常慢的速度 +0.9 和会随人变化的声音 +0.8。

Content side (the direct test of A1): in the acute scenario only 24% [17–31%] choose "a calm version of a song I love" against 46% for featureless neutral sound; in the wind-down scenario that flips to 47% [39–56%] and 19%. The within-person difference is significant (McNemar χ² = 14.29, p < 0.001, b = 16, c = 47). The GAD-2-positive segment is sharper in the acute scenario: real song 16%, neutral 49%. C4 supplies the neutral-bed spec: what makes it worse is sudden or sharp sounds (−1.5), a strong beat (−1.1), lyrics (−0.8) and a low continuous hum (−0.7); what helps is a very slow tempo (+0.9) and sound that changes as you calm (+0.8).

测量与产品侧：拥有者中 57%［45–69%］愿意连接设备；54% 看重一个被测量的结果（"你的心率从 92 降到 68"）；对现有压力/HRV 读数的信任只有 3.06/5；隐私顾虑者 18%。行动缺口很直接：看到"高压力"读数后 38% 什么也不做，30% 做呼吸，18% 说反而更焦虑。TAM：PU 4.9/7、PEOU 5.2/7、BI 4.9/7，BI top-2 仅 25%（拥有者 32%），BI ~ PU β 0.52 + PEOU β 0.34，R² 0.64。Kano：K1 实时适配 = 魅力属性 A，K2 前后结果 = 一维属性 O（GAD-2 阳性子群里升为必备 M）；而 K3 真歌与 K4 中性声音都被判为无差异 I——功能本身不打动人，闭环才打动人。形态偏好：内置在穿戴 App 里 45% > 独立 App 32% > 艺术家舒缓专辑 17%。价格：美/欧可接受区间 [$5.19, $8.46]，OPP $6.28、IPP $7.37；中国 [¥14.16, ¥28.6]，OPP ¥15.47。$6.99 的购买意向 top-2 35%（折半后 18%），中国 ¥30 只有 15%。最大的障碍是"又一个订阅" 49%。

Measurement and product side: 57% [45–69%] of owners would connect a device; 54% value a measured result ("your heart rate settled 92 → 68"); trust in current stress/HRV scores is only 3.06/5; 18% are privacy-concerned. The action gap is blunt: after a "high stress" reading 38% do nothing, 30% breathe, and 18% say it makes them more anxious. TAM: PU 4.9/7, PEOU 5.2/7, BI 4.9/7, but BI top-2 only 25% (owners 32%); BI ~ PU β 0.52 + PEOU β 0.34, R² 0.64. Kano: K1 live adaptation is attractive (A) and K2 the before/after result is one-dimensional (O) — and must-be (M) inside the GAD-2-positive segment — while both K3 (real songs) and K4 (neutral sound) come back indifferent (I): the content type is not what moves people, the loop is. Form preference: inside the wearable's own app 45% > standalone app 32% > artist calm album 17%. Price: US/EU acceptable range [$5.19, $8.46] with OPP $6.28 and IPP $7.37; China [¥14.16, ¥28.6], OPP ¥15.47. Purchase intent at $6.99 is 35% top-2 (18% after the ×0.5 discount); at ¥30 in China only 15%. The largest barrier is "another subscription" at 49%.

6. 焦点小组与盲听Focus groups and the blind stimulus test

焦点小组坐在一对一访谈（提出了唤醒状态假设）和问卷/试点（量化并检验它）之间，做三件访谈做不到的事：让人们互相接话地把需求时刻聚成类；在不知道听的是什么的情况下盲评三段音频；对概念、测量方式与价格当场反应（Krueger & Casey 2015 [7]；Morgan 1997 [8]）。两组：FG-1 美/欧可穿戴拥有者线上 7 人 82 分钟；FG-2 中国办公/混合办公上海线下 6 人 78 分钟；由研究助理按脚本主持，创始人不在场。

Focus groups sit between the one-to-one interviews (which generated the arousal-state hypothesis) and the survey and pilot (which quantify and test it), doing three things interviews cannot: letting people build on each other while clustering the moments of need; rating three audio clips blind; and reacting to the concept, the measurement idea and the price in the room (Krueger & Casey 2015 [7]; Morgan 1997 [8]). Two groups: FG-1, seven US/EU wearable owners online, 82 min; FG-2, six China office / hybrid workers offline in Shanghai, 78 min. Both moderated from a script by a research assistant, with the founder absent.

盲听用三段等响度、顺序轮换的 60 秒片段：S-A 一首广为人知的歌的舒缓器乐编配（60–66 bpm，无打击乐，−14 LUFS）；S-B 中性自适应床（无旋律、无歌词、无节拍、无尖锐瞬态、无低频嗡鸣，频谱重心 60 秒内从约 1.2 kHz 漂到约 600 Hz，60 Hz 以下无内容，瞬态不超过床 +6 dB）；S-C 柔和铺底上的引导语音（对照 / 现有产品风格）。参与者先私下打分再讨论。结果：S-B 在"急性/夜里"这个时刻上以 10/13 胜出，S-A 在"白天放松"上以 9/13 胜出；S-C 被 8/13 判为"哪个时刻都不合适"。

The blind test uses three loudness-matched 60-second clips in rotated order: S-A, a calm instrumental arrangement of a widely known song (60–66 bpm, no percussion, −14 LUFS); S-B, a neutral adaptive bed (no melody, lyrics, beat, sharp transients or low drone; spectral centroid drifting ~1.2 kHz → ~600 Hz over the minute; nothing below 60 Hz; transients no more than +6 dB above the bed); S-C, a guided voice over a soft pad as the incumbent-style control. Participants rate privately before discussing. S-B wins the acute/at-night moment 10/13; S-A wins daytime wind-down 9/13; S-C is assigned to "neither moment" by 8 of 13.

八个主题里，三个直接改了产品：T2 分状态的声音（FG-1 6/7、FG-2 5/6）——拒绝 S-A 的理由与访谈一模一样，"旋律给了我的大脑一个可以追的东西"、"歌词会开始讲一个故事"；"一堵软墙"这个比喻在 FG-2 里未经提示地再次出现。T3 数字是把双刃剑（5/7、4/6）——拥有者珍视结果（"92 → 68，那个我会截图"），但四位说盯着实时心率线会"把它变成一场我会失败的考试"。T4 一次点击，或者干脆不用（6/7、6/6）——凌晨一点没有人会选心情、时长或曲风。还有两个必须记录的反对声音：两位参与者把 S-B 形容为"医院的空调"——中性声音的工程质量是决定性的，这条风险后来变成 E4 的人工试听 QA 门。

Three of the eight themes changed the product directly. T2, state-dependent sound (FG-1 6/7, FG-2 5/6): the reasons for rejecting S-A at night echo the interviews — melody "gives my brain something to chase", lyrics "start a story" — and the "soft wall" metaphor surfaced unprompted in FG-2. T3, the number cuts both ways (5/7, 4/6): owners value the result ("92 → 68 — that I would screenshot") but four said watching a live heart-rate line would "make it a test I can fail". T4, one tap or nothing (6/7, 6/6): at 1 a.m. nobody picks a mood, a duration or a genre. Two disconfirming voices are kept on the record: two participants called S-B "hospital air-conditioning" — the engineering quality of neutral sound is decisive, a risk that later became E4's human listening QA gate.

点票（每人 3 票，13 人）：不问我就适应我 14 票 · 给我看事后结果 11 票 · 手表/戒指一键启动 9 票 · 中性声音做得"干净" 3 票 · 我喜欢的艺人的真歌 2 票。这是产品优先级的直接来源。

Dot vote (3 dots each, 13 participants): adapts to me without asking 14 · shows me the after-result 11 · works from the watch or ring with one tap 9 · neutral sound engineered "clean" 3 · real songs by artists I like 2. That ranking is where the product's priorities come from.

进入模型时，这些发现被收缩：质性 k = 0.5，且与访谈相关（同一假设、部分同源）再乘 0.6（对数空间）——A1a ↓ LR 0.6、A1b ↑ 1.4、A3 ↑ 1.3（附带设计条件）、A4 独立订阅 ↓ 0.9 而捆绑形态权重上调。

Entering the model, these findings are shrunk: qualitative k = 0.5, then × 0.6 in log space for correlation with the interviews — A1a ↓ LR 0.6, A1b ↑ 1.4, A3 ↑ 1.3 (with a design condition attached), A4 ↓ 0.9 for standalone B2C while the weight on the bundled C2 form goes up.

7. 贝叶斯决策引擎The Bayesian layer — priors, likelihood ratios, posteriors, gates

规则只有三步（DECISION_MODEL §1）：先验概率 P 换成先验赔率 O₀ = P/(1−P)；每条证据给一个似然比 LR = P(看到这条证据 | 假设为真) ÷ P(看到这条证据 | 假设为假)；后验赔率 O = O₀ × ∏LR，再换回概率 O/(1+O)。LR > 1 支持，< 1 反对，= 1 无信息。用主观似然比整合定性与定量证据在社会科学里有明确先例：Fairfield & Charman 2017 [11] 的过程追踪显式贝叶斯分析，以及 Humphreys & Jacobs 2015 [12] 的混合方法贝叶斯框架；证据分级仿 GRADE（Guyatt et al. 2008 [13]）；把预注册与书面的停止判据用在创新决策上的先例是 Ries 2011 [14] 与 Bland & Osterwalder 2019 [15]。

The rule has three steps (DECISION_MODEL §1): turn the prior probability P into prior odds O₀ = P/(1−P); give each piece of evidence a likelihood ratio LR = P(evidence | hypothesis true) ÷ P(evidence | hypothesis false); posterior odds O = O₀ × ∏LR, converted back with O/(1+O). LR > 1 supports, < 1 disconfirms, = 1 is uninformative. Using subjective likelihood ratios to integrate qualitative and quantitative evidence has an explicit social-science precedent: Fairfield & Charman's (2017 [11]) Bayesian process tracing and Humphreys & Jacobs's (2015 [12]) mixed-methods Bayesian framework; the quality grading follows GRADE (Guyatt et al. 2008 [13]); the precedent for pre-registration and written kill criteria in innovation decisions is Ries (2011 [14]) and Bland & Osterwalder (2019 [15]).

校准：一条证据怎么变成一个数字Calibration — how evidence becomes a number

算例（DECISION_MODEL §2.2 原文）：三个急性焦虑访谈拒绝真歌。原始"强反对"每人约 LR 0.3；定性 + 自选样本 → k = 0.5 → 每人 0.55；三条来自同一渠道，按约 2 条有效计 → 0.55² ≈ 0.30；保守取整到 0.35，以免让三段对话权重过大。

Worked example (DECISION_MODEL §2.2 verbatim): three acute-anxiety interviewees rejecting real music. A raw "strong" disconfirmation would be LR ≈ 0.3 per interview; qualitative and self-selected → k = 0.5 → 0.55 each; three items from one recruitment channel are treated as roughly two effective ones → 0.55² ≈ 0.30; rounded conservatively to 0.35 so that three conversations are not over-weighted.

八个信念现在站在哪Where the eight beliefs stand

门槛与判定Gates and the verdict

概念可行性用关键信念的联合概率表示（假设独立，并同时点名最弱一环）：C3 艺术家舒缓内容线 65%（A1b 80% × A4 82%，最弱 A1b）> C1 中性声音 App 46%（¬A1a 92% × A2a 85% × A3 72% × A4 82%，最弱 A3）≈ C2 自适应音频 SDK 45%（A2a × A4 × A5，最弱 A5 65%）> C1 真歌 App 26%（A1b × A2a × A3 × A4 × A5，最弱 A5）。

Concept viability is the independence-assumed joint probability of the critical beliefs, with the weakest link named: C3, an artist calm-content line, 65% (A1b 80% × A4 82%, weakest A1b) > C1-neutral, 46% (¬A1a 92% × A2a 85% × A3 72% × A4 82%, weakest A3) ≈ C2, an adaptive-audio SDK, 45% (weakest A5 at 65%) > C1-real, 26% (weakest A5).

把这个排序和问卷 E3 的形态偏好（穿戴内置 45% > 独立 App 32%）以及三分之一偏好捆绑/雇主付费放在一起，产品方向就出来了：一个双模式自适应音频引擎（急性 = 中性 / 放松 = 真歌），以独立 App（C1）与穿戴品牌 SDK（C2）两种形态交付，C3 作为最便宜的真歌需求探针先行。

Read that ranking together with the survey's form preference (45% inside the wearable's own app vs 32% standalone) and the third of respondents who prefer bundled or employer payment, and the direction follows: one dual-mode adaptive-audio engine (acute = neutral, wind-down = real music), delivered both as a standalone app (C1) and as a wearable-brand SDK (C2), with C3 running first as the cheapest probe of real-music demand.

8. 从证据到产品：为什么这样设计Evidence → product — why the app looks like this

产品的每一个交互都有一条研究理由，反过来读也成立：产品里没有一个交互是没有研究理由的（这张映射表的完整版在 蓝图 §7）。人因工程的部分同样有据：急性焦虑时注意力收窄（Easterbrook 1959 [59]），工作记忆被占满（Sweller 的认知负荷理论），所以第一原则不是"功能多"，而是在用户最不能思考的时刻，把决定降到零；配合 Hick 定律（选项越多决策越慢）、Fitts 定律与 Apple HIG 的 ≥ 44 pt 触控目标、Weiser & Brown 1996 [60] 的"安静的技术"，以及自我决定理论（自主 / 胜任，不设内疚）。

Every interaction in the product has a research reason, and the mapping reads backwards too: no interaction in the product lacks one (the full table is in blueprint §7). The human-factors half is equally sourced: under acute anxiety attention narrows (Easterbrook 1959 [59]) and working memory is full (Sweller's cognitive-load account), so the first principle is not feature count but zero decisions at the moment the user can least think — supported by Hick's law, Fitts's law and Apple's ≥ 44 pt targets, Weiser & Brown's (1996 [60]) calm technology, and self-determination theory (autonomy and competence, no guilt).

看它跑起来See it running

9. 算法怎么工作How the algorithm actually works

9.1 信号链：心率开车，HRV 记账The signal chain — heart rate steers, HRV is measured

自评是焦虑这个构念的测量；心率类指标测的是自主神经唤醒——对身体诚实，对原因盲目（运动、咖啡因、兴奋和恐惧都会抬高心率、压低 HRV）。所以分工是硬性的：HR（bpm）腕部 PPG 误差约 6%、1–5 秒可得 → 用作控制信号与共同主要生理结局；HRV（RMSSD / SDNN，ms）是文献里与放松相关的迷走张力标记（越高越平静），但腕表噪声大（Apple Watch SDNN MAPE ≈ 29%）、稀疏（HealthKit 机会性采样、滞后可达 30 分钟；Oura 只有夜间、次日早晨才给）→ 只做前后对比，本阶段绝不用于实时控制。

Self-report measures the construct of anxiety; heart-rate measures index autonomic arousal — honest about the body, blind to the cause (exertion, caffeine, excitement and fear all raise HR and lower HRV). So the division of labour is hard: HR (bpm) is accurate on wrist PPG (≈ 6% error) and available every 1–5 s → the control signal and co-primary physiological outcome; HRV (RMSSD / SDNN, ms) is the vagal-tone marker the literature links to relaxation (higher = calmer) but is noisy on the wrist (Apple Watch SDNN MAPE ≈ 29%) and sparse (HealthKit samples opportunistically with lag up to 30 min; Oura is sleep-only, next morning) → a before/after outcome, never the live control signal in this phase.

这张表之所以要"简单到一个人能手动执行"，是因为 E1 试点是 Wizard-of-Oz：由人来当算法。这样效能检验测的是"自适应中性声音有没有用"，而不是"我们的代码有没有 bug"。同一张表随后被写进 core/rules.js，成为产品的默认臂。

The table is deliberately simple enough for a person to execute by hand, because E1 is a Wizard-of-Oz pilot: a human plays the algorithm. That way the efficacy test measures whether adaptive neutral sound works, not whether our code has bugs. The same table is then written into core/rules.js as the product's default arm.

9.2 推荐：真歌那一路怎么选曲Recommendation — how the real-song layer chooses

嵌入播放器（Spotify / SoundCloud）不能改音频，所以 Unwind 的"自适应"只能体现在选曲与衔接（tier S）；自有曲目走 Web Audio 才能做参数级适配（tier P）。评分是五个乘数的积，权重来自 app/RECOMMEND_BRIEF.md 与 app/js/core/recommend.js。

Embedded players cannot alter audio, so adaptation on the Unwind layer is selection and sequencing (tier S); only owned files routed through Web Audio allow parametric adaptation (tier P). The score is a product of five multipliers, specified in app/RECOMMEND_BRIEF.md and implemented in app/js/core/recommend.js.

9.3 学习路线：从规则表到自我改进的曲库The learning roadmap — from a rule table to a self-improving library

奖励设计与护栏（ML §5）：每次会话的奖励 = −标准化 Δr 心率 + λ × Δ自评（λ 是待定的权重，设计文件没有预设值）。绝不只用自评（会被期待效应骗），也绝不只用心率（会被"一动不动"骗）；质量指数 < 0.70 的会话不进奖励。任何策略变更先离线评估再灰度上线，从不 100% 直接铺开；某个臂比默认差超过 0.3 个标准差就对该用户暂停。

Reward design and guardrails (ML §5): the per-session reward is −standardised Δr in heart rate + λ × Δ self-report (λ is a weight still to be set; the design document fixes no value for it). Never optimise on self-report alone (gameable by expectancy) and never on heart rate alone (gameable by sitting still); sessions below the 0.70 quality index do not enter the reward. Every policy change is evaluated offline first, then online against a held-out default cohort — never a 100% rollout — and any arm more than 0.3 SD worse than default is suspended for that user.

10. 怎么证明有效How efficacy will be proved — E0, E1 and what they feed

10.1 识别策略Identification

E0 校准子研究先跑：n ≈ 8，同时佩戴 Apple Watch 与 Polar H10，各 3 节，估计 ICC、MAPE 与回归校准系数 λ = Cov(watch, polar)/Var(watch)。理由是测量误差的两种后果不同：作为结局时，非差异性误差只损失功效、不产生偏倚；作为回归自变量时（例如把基线 HRV 当调节变量）会造成衰减偏倚，必须去衰减（Carroll et al. 2006 [58]）。

E0, the calibration sub-study, runs first: n ≈ 8 wearing an Apple Watch and a Polar H10 concurrently for 3 sessions each, estimating ICC, MAPE and the regression-calibration coefficient λ = Cov(watch, polar)/Var(watch). The reason is that measurement error has two different consequences: as an outcome, non-differential error costs power but does not bias; as a regressor (say, baseline HRV as a moderator) it causes attenuation bias and has to be de-attenuated (Carroll et al. 2006 [58]).

E1 是随机化的被试内三条件交叉设计，顺序按拉丁方平衡：T1 自适应中性声音（心率驱动）· T2 自适应真歌舒缓版（参与者自选曲目）· C 主动对照 = 参与者自己常听的一份"放松"歌单，原样播放。对照不是静音——静音会改变期待并把"任何音频有效"和"我们的音频有效"混在一起。每人 3 节 × 15 分钟，不同日、同一时段、间隔 ≥ 24 小时（洗脱），坐姿、只用手机。

E1 is a randomised within-subject three-condition crossover, order-balanced by Latin square: T1 adaptive neutral sound (HR-driven) · T2 an adaptive calm arrangement of a real song the participant chose · C, an active control — a generic "relaxing" playlist they already know, played untouched. The control is not silence: silence changes expectancy and confounds "any audio" with "our audio". Three 15-minute sessions per participant on separate days, same time of day, ≥ 24 h apart (washout), seated, phone only.

10.2 威胁与控制Threats and controls

功效的诚实说法。配对 t 检验（α = .05 双侧，功效 = .80）：d_z = 0.5 → n = 34；0.6 → 24；0.65 → 21；0.8 → 15。Cochrane 的围手术期估计（−5.7 STAI-S，SD ≈ 10 → d ≈ 0.55）对应 n ≈ 28；生理效应（d ≈ 0.4）需要 n ≈ 50。所以：n = 20–30 对 A2a 是可行性信号，对 A2b 功效不足——这句话必须写在论文里（METHODOLOGY §3）。两个补救：每人 3 节（增加个体内信息）；用元分析先验做贝叶斯分析，让后验而不是 p 值去撞门槛。

Power, stated honestly. Paired-t power (α = .05 two-sided, power = .80): d_z = 0.5 → n = 34; 0.6 → 24; 0.65 → 21; 0.8 → 15. The Cochrane perioperative estimate (−5.7 STAI-S, SD ≈ 10 → d ≈ 0.55) implies n ≈ 28; physiological effects (d ≈ 0.4) need n ≈ 50. So n = 20–30 is a feasibility signal for A2a and under-powered for A2b — and the dissertation says so (METHODOLOGY §3). Two remedies: three sessions per person for more within-person information, and a Bayesian analysis with the meta-analytic prior so that the posterior, not a p-value, meets the gate.

10.3 管线已建好，并已在合成数据上跑过The pipeline is built and has been exercised — on synthetic data

从 App 导出到决策模型的整条路径已经存在：app/tools/export_to_csv.py（导出 JSON → sessions.csv + windows.csv）→ research/pilot/e1_analysis.py（预注册模型、贝叶斯再分析、似然比、门槛、报告与图）→ 把观测到的 LR 替换回 DECISION_MODEL.md 与 decision.js。28 个测试覆盖了 schema、拉丁方一致性、0.70 质量剔除、顺序推断、注入效应的方向以及 LR / 门槛算术。

The whole path from app export to decision model already exists: app/tools/export_to_csv.py (export JSON → sessions.csv + windows.csv) → research/pilot/e1_analysis.py (the pre-registered model, the Bayesian re-analysis, the LRs, the gates, a report and figures) → the observed LRs replace the pending rows in DECISION_MODEL.md and decision.js. Twenty-eight tests cover the schema, Latin-square agreement, the 0.70 quality exclusion, order inference, the direction of injected effects, and the LR / gate arithmetic.

结果怎么回到决策：report.md §4 打印三条预注册的行——A2a（STAI-S 的 β1）LR 4.0 / 0.30；A2b（残差心率 β1，Holm 后）LR 4.0 / 0.30；A1a（β1 − β2 的对比）LR 3.0 / 1.0 / 0.33。把这三个数替换回信念登记册，§5 的门槛检查就会机械地打印出判定——决策备忘录不需要再作一次判断。

How the result returns to the decision: report.md §4 prints the three pre-registered rows — A2a (β1 on STAI-S) LR 4.0 / 0.30; A2b (β1 on residual HR, after Holm) LR 4.0 / 0.30; A1a (the β1 − β2 contrast) LR 3.0 / 1.0 / 0.33. Substituting those three numbers into the belief register makes the §5 gate check print the verdict mechanically — the decision memo requires no further judgement.

11. 给谁用、放什么音乐、为什么Who it is for, what plays, and why

不是"焦虑"这一个人群，而是一组可分开的时刻。焦点小组的 T1 主题里，参与者用自己的便签聚出五个簇（下班后的复盘 · 入睡 · 上台前 · 通知惊跳 · 低度背景紧张），问卷 B3 的时间分布把它们量化（床上 57%、工作 52%、傍晚 33%、深夜 23%）；产品把其中四个做成了显式的 moment 标签（app/ARCHITECTURE.md 的 type Moment：急性 / 放松 / 入睡 / 上台前）。

Not one "anxiety" audience but a set of separable moments. In focus-group theme T1 participants clustered their own sticky notes into five groups (rumination after the day, sleep-onset, anticipatory/performance, notification spikes, low-grade background tension); the survey's timing item quantifies them (in bed 57%, at work 52%, evening 33%, night 23%); and the product turns four of them into explicit moment tags (type Moment in app/ARCHITECTURE.md: acute / wind-down / sleep-onset / pre-performance).

选模式这个动作本身就是时刻标签：一次点击同时完成了分群与上下文采集，这既是产品的简化，也是研究的数据（moment tag 进入 bandit 的上下文，也进入 A1a / A1b 的行为证据）。

Choosing a mode is the moment tag: one tap does segmentation and context capture at once — a simplification for the user and a datum for the research (the moment tag becomes bandit context and behavioural evidence on A1a / A1b).

放什么声音，规格从哪来What plays, and where the spec comes from

Settle 的中性床不是"氛围音乐"，是一份从用户抱怨里逆推出来的工程规格：无旋律、无节拍、无歌词、无可辨音高中心；频谱重心在一分钟内从约 1.2 kHz 漂到约 600 Hz（越平静越暗）；60 Hz 以下无内容（直接回应 I-05 的"低频嗡鸣引起恶心"）；瞬态不超过床 +6 dB（回应"尖锐刺耳声"，问卷里最负的一项 −1.5）；响度固定 −16 LUFS；参数每 60 秒按 v0 规则表变化一次。四张自有的床已经在产品里：Arctic Reverberation · Zen Furnace · Fluorescent Hum · Silent Monolith。

The Settle neutral bed is not ambient music but an engineering spec reverse-engineered from complaints: no melody, no beat, no lyrics, no discernible pitch centre; a spectral centroid drifting from ~1.2 kHz to ~600 Hz across a minute (darker as you settle); nothing below 60 Hz (answering I-05's nauseating low hum); transients no more than +6 dB above the bed (answering "sudden / sharp sounds", the most negative item in the survey at −1.5); loudness fixed at −16 LUFS; parameters stepped every 60 s by the v0 rule table. Four owned beds ship today: Arctic Reverberation, Zen Furnace, Fluorescent Hum and Silent Monolith.

Unwind 的真歌规格是：器乐钢琴 / 吉他改编、60–66 bpm、无打击乐、柔和低通、无突变动态。产品负责人的补充要求很具体——曲池需要更多慢速、混响厚重的钢琴。现在的曲库是 69 首 Unwind 曲目，其中 53 首是慢钢琴；9 首 TDMusic 自有 WAV 已按 −16 LUFS 转码入库，其中 5 首进 Unwind（Suburban Nights · CRT Light · Rewind the Years · Sunset Rewind · 钢琴弹唱），另外 4 首就是上面那四张 Settle 中性床。曲目来源可追溯：从 Spotify 官方舒缓歌单出发，因为这些歌单里多是匿名"幽灵艺人"，又通过 Wikidata 解析知名艺人的 Spotify ID 取其"热门"榜；每个 ID 都用 Spotify 的 oEmbed 端点校验过。

The Unwind real songs are specified as instrumental piano or guitar arrangements, 60–66 bpm, no percussion, gentle low-pass, no sudden dynamics — with an explicit product request for more slow, reverb-heavy piano. The catalogue today holds 69 Unwind tracks, 53 of them slow piano; nine TDMusic-owned WAVs have been ingested at −16 LUFS, five of them into Unwind (Suburban Nights, CRT Light, Rewind the Years, Sunset Rewind, 钢琴弹唱) and the other four being the Settle beds named above. Provenance is traceable: starting from Spotify's editorial calm playlists, which are dominated by pseudonymous "ghost" artists, well-known artists were resolved to Spotify IDs via Wikidata and drawn from their Popular lists; every id was verified against Spotify's oEmbed endpoint.

12. 我们现在在哪Where the app stands today — 6 Sep 2026

已经做出来的东西What exists

• 产品蓝图 v1（3 Sep）：中英对照，十节——人因原则、系统架构、13 阶段用户路径、八屏 + 手表 + 安全屏、数据采集接口矩阵（HealthKit / Oura API v2 / Polar H10 BLE / Web Bluetooth）、曲库两档与标签规范、学术方法 → 产品交互逐项映射、反馈闭环、技术栈与四周计划、v0 控制规则。打开蓝图 →
• 可安装的 App（PWA）：https://38.180.150.12.sslip.io/app/（HTTPS，iPhone Safari"添加到主屏幕"即可）。首页（Settle / Unwind、时长、睡眠淡出）→ 预休息 60 秒（研究模式 300 秒）+ 自评 0–10 → 会话 → 20 秒回访 → 结果（相对预休息 / 相对"你的平时"、沉降曲线、信号质量、有 RR 时给 RMSSD）→ 历史 / 导出 / 删除 → 设置；安全屏（HR > 130 持续 10 秒或用户点"我需要接地"）；Service Worker 离线壳。
• 研究模式（v0.2.0 起）：参与者编号 → 拉丁方条件 T1 / T2 / C、5 分钟预休息、STAI-6 前后测、期待值两题、导出含条件字段。这不是演示功能——它就是 E1 的现场工具。
• 推荐与播放（v0.5.x）：按标签 × 时刻 × 心率带 × 喜好 × 新鲜度 × 质量评分，加权洗牌，过渡约束，"为什么是这首"；Spotify SDK（PKCE）/ 嵌入回退 / 自有交叉淡入 / SoundCloud；Like · Slower · Next · Shuffle 随时可用。
• 测试（TODO §12 在 v0.4.0 时点的计数，此后仍在增加）：单元 70 · 端到端 11 · 音频离线渲染检查 6 · 场景引擎 14；另有 WebKit（iPhone 15 模拟）6/6 场景通过（§16）。核心逻辑是纯函数，测试即规格。
• 自有音乐入库：9 首自有 WAV 经 ingest_music.py 转成 −16 LUFS AAC 并做声学分析（32 MB）；Settle 床 Arctic Reverberation / Zen Furnace / Fluorescent Hum / Silent Monolith，Unwind 自有曲目 Suburban Nights / CRT Light / Rewind the Years / Sunset Rewind 等。
• 试点数据管线：research/pilot/ 完成，28 个测试；导出 → CSV → 预注册模型 → 报告 / 图 / 门槛输入。

App Store 与 TestFlight 的真实状态The real store status

13. 下一步，以及每一步是为了什么What comes next, and which belief each step is meant to move

路线图不按"功能"排，按它买到多少信息排。每个测试都是一次信息采购：成本、时长与预期的后验移动量都写在 DECISION_MODEL.md §4，排序遵循"每单位成本的期望信息量"——法务核查最便宜也最决定性（LR 8 / 0.15），所以先跑；然后是问卷；最后是最贵的试点。这就是设计思维阶段的实物期权视角（Bowman & Hurry 1993 [17]；McGrath 1999 [16]）。

The roadmap is ordered not by feature but by how much information each step buys. Every test is a purchase of information whose cost, duration and expected posterior shift are listed in DECISION_MODEL.md §4, and the ordering follows expected information per unit cost — the legal check is the cheapest and most decisive (LR 8 / 0.15) so it runs first, then the survey, then the expensive pilot. That is the real-options view of the design-thinking phase (Bowman & Hurry 1993 [17]; McGrath 1999 [16]).

14. 评估方法一览Evaluation methods at a glance — construct, instrument, statistic, threshold, decision node

15. 参考文献References

仅列出本项目文档中已经引用的著作；本页不引入任何新文献。条目按主题分组编号。Only works already cited in this project's own documents appear below; this page introduces no new literature. Entries are numbered within thematic groups.

方法与框架Method and frame

1. Brown, T. (2008). Design thinking.
2. Liedtka, J. (2018). On the effectiveness of design thinking. Journal of Product Innovation Management.
3. Coghlan, D., & Brannick, T. (2019). Doing Action Research in Your Own Organization.
4. Creswell, J. W., & Plano Clark, V. L. (2018). Designing and Conducting Mixed Methods Research.
5. Braun, V., & Clarke, V. (2006). Qualitative Research in Psychology, 3, 77. (Also 2019, reflexive thematic analysis.)
6. Kozinets, R. V. (2020). Netnography.
7. Krueger, R. A., & Casey, M. A. (2015). Focus Groups.
8. Morgan, D. L. (1997). Focus groups as qualitative research.
9. Flanagan, J. C. (1954). The critical incident technique. Psychological Bulletin, 51(4), 327–358.
10. Christensen, C. M., et al. (2016). Jobs to be done — current alternatives as the competitive set.

决策分析与创新管理Decision analytics and innovation management

11. Fairfield, T., & Charman, A. E. (2017). Explicit Bayesian analysis for process tracing: guidelines, opportunities, and caveats. Political Analysis, 25(3), 363–380.
12. Humphreys, M., & Jacobs, A. M. (2015). Mixing methods: a Bayesian approach. American Political Science Review, 109(4), 653–673.
13. GRADE Working Group (Guyatt, G. H., et al., 2008). BMJ, 336, 924.
14. Ries, E. (2011). The Lean Startup.
15. Bland, D. J., & Osterwalder, A. (2019). Testing Business Ideas.
16. McGrath, R. G. (1999). Academy of Management Review, 24, 13.
17. Bowman, E. H., & Hurry, D. (1993). Real options and organisational learning.

量表与调查方法Instruments and survey methods

18. Kroenke, K., et al. (2007). GAD-2 anxiety screener.
19. Cohen, S., & Williamson, G. (1988). PSS-4 perceived stress scale.
20. Spielberger, C. D. (1983). STAI manual.
21. Marteau, T. M., & Bekker, H. (1992). STAI six-item short form. British Journal of Clinical Psychology, 31, 301.
22. Devilly, G. J., & Borkovec, T. D. (2000). Credibility / expectancy. Journal of Behavior Therapy and Experimental Psychiatry, 31, 73.
23. Ulwick, A. W. (2002). Outcome-driven innovation. Harvard Business Review.
24. Davis, F. D. (1989). Technology acceptance model.
25. Venkatesh, V., & Davis, F. D. (2000). Extension of the technology acceptance model.
26. Kano, N., et al. (1984). Attractive quality and must-be quality.
27. Berger, C., et al. (1993). Better / Worse coefficients for Kano categories.
28. Rogers, E. M. (2003). Diffusion of Innovations, 5th ed.
29. van Westendorp, P. (1976). Price sensitivity meter.
30. Morwitz, V. G., Steckel, J. H., & Gupta, A. (2007). When do purchase intentions predict sales? International Journal of Forecasting, 23, 347–364.
31. McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions.
32. Vickers, A. J., & Altman, D. G. (2001). Analysing controlled trials with baseline and follow-up measurements. BMJ, 323, 1123.

临床与生理证据Clinical and physiological evidence

33. Bradt, J., Dileo, C., & Shim, M. (2013). Music interventions for preoperative anxiety. Cochrane Database of Systematic Reviews (26 RCTs, N = 2,051; −5.72 STAI-S; GRADE low).
34. de Witte, M., et al. (2020). Effects of music interventions on stress-related outcomes (104 RCTs, N = 9,617; psychological d = .545, physiological d = .380, heart rate d = .456).
35. Harney, C., et al. (2023). Musicae Scientiae — music listening and anxiety (21 controlled studies, d = −0.77).
36. Lassner, et al. (2024/2025). Delivery modes compared (SMD = 0.47, k = 23, n = 1,084; GRADE very low; effect not maintained at follow-up).
37. Panteleeva, Y., et al. (2018). Music-induced anxiety reduction: self-report significant (d = −.30), psychophysiological not significant.
38. Bernardi, L., Porta, C., & Sleight, P. (2006). Cardiovascular, cerebrovascular and respiratory changes induced by music. Heart, 92, 445.
39. Mitrovic & Paladin (2026). Tempo–HRV relationship, review. Frontiers in Cardiovascular Medicine.
40. Goessl, V. C., Curtiss, J., & Hofmann, S. G. (2017). HRV-biofeedback and stress / anxiety, meta-analysis (g = 0.81). Psychological Medicine.
41. Creswell, J. D., & Goldberg, S. B. (2025). Meditation-app engagement: ~4.7% 30-day retention; 1–4 lifetime sessions.
42. O'Daffer, A., et al. (2022). Efficacy and conflicts of interest in commercial mindfulness-app trials.
43. Haruvi, A., et al. (2022). Effects of personalised soundscapes on focus. Frontiers in Computational Neuroscience (industry-funded; measures focus, not anxiety).
44. Nigg, J. T., et al. (2024). Systematic review finding no controlled studies of brown noise and ADHD.
45. Woods, K. J. P., et al. (2024). Amplitude-modulated music and attention. Communications Biology.
46. Pedersen, S. K. A., et al. (2017). Individualised music and agitation in dementia, meta-analysis (d = 0.61, 12 RCTs).
47. McCreedy, E. M., et al. (2021). Pragmatic RCT of personalised music in nursing homes (N = 463, 54 homes): no significant effect on the primary agitation outcome.
48. Mendes, et al. (2024). Background music and attentional networks in ADHD, case-control.
49. World Health Organization (2023). Anxiety disorders fact sheet (359 million people in 2021; 27.6% receiving treatment).
50. Huang, Y., et al. (2019). Prevalence of mental disorders in China (anxiety 5.0% 12-month, 7.6% lifetime).

信号处理与学习Signal processing and learning

51. Task Force of the ESC and NASPE (1996). Heart rate variability standards. Circulation, 93, 1043.
52. Tarvainen, M. P., et al. (2002). Smoothness-priors detrending. IEEE Transactions on Biomedical Engineering, 49, 172.
53. Lipponen, J. A., & Tarvainen, M. P. (2019). RR-interval artefact correction. Journal of Medical Engineering & Technology, 43, 173.
54. Shaffer, F., & Ginsberg, J. P. (2017). An overview of heart rate variability metrics and norms. Frontiers in Public Health, 5, 258.
55. Agrawal, S., & Goyal, N. (2013). Thompson sampling for contextual bandits.
56. Dudík, M., Langford, J., & Li, L. (2011). Doubly robust policy evaluation and learning.
57. Gelman, A., & Hill, J. (2007). Data Analysis Using Regression and Multilevel / Hierarchical Models.
58. Carroll, R. J., et al. (2006). Measurement Error in Nonlinear Models.

人因工程Human factors

59. Easterbrook, J. A. (1959). The effect of emotion on cue utilisation and the organisation of behaviour.
60. Weiser, M., & Brown, J. S. (1996). The coming age of calm technology.

在正文中按名字提到、但项目文档未给出年份或出处的框架与定律，不进入上面的编号列表：Ansoff 矩阵 · McKinsey 三层面 · O'Reilly & Tushman 的双元性 · Sweller 的认知负荷 · Hick 定律 · Fitts 定律 · Nielsen 启发式 · Deci & Ryan 的自我决定理论 · Efraimidis–Spirakis 加权抽样。Frameworks and laws named in the text but for which the project's own documents give no year or venue are deliberately left out of the numbered list: the Ansoff matrix, McKinsey's Three Horizons, O'Reilly & Tushman's ambidexterity, Sweller's cognitive load, Hick's law, Fitts's law, Nielsen's heuristics, Deci & Ryan's self-determination theory, and Efraimidis–Spirakis weighted sampling.