Lilt 产品设计蓝图 v1Lilt product blueprint — functions, data capture, library, journey, and how the research method becomes interaction

一个 App、两种模式：Settle（安定）用无旋律的自适应中性声音接住急性焦虑；Unwind（舒缓）用真歌的舒缓版陪你从一天里退出来。手表/戒指给出心率，App 静默地调整声音，结束后才把"比你平时更平静了多少"告诉你。每一次聆听同时是一次被测量的实验。

One app, two modes: Settle catches acute anxiety with melody-free adaptive neutral sound; Unwind eases you out of the day with calm arrangements of real songs. The watch or ring supplies heart rate, the app adapts the sound silently, and only afterwards tells you how much calmer you were than your usual. Every listen is also a measured experiment.

依据 Sources: research/ML_RESEARCH_DESIGN.md（信号、E0–E4、架构）· research/METHODOLOGY_DESIGN.md §4（试点识别）· research/DECISION_MODEL.md（A1a…A6 · 门槛 · 内容转向）· surveys/questionnaire-v2.md（B3 / C3 / D / E2 Kano）· research/FOCUS_GROUP_KIT.md（刺激规格 · "实时数字 vs 事后结果"）· interviews/user-interviews.md（"软墙"）· 现有 5 屏点击原型 site/demo/。
交付 Delivery: 网页 http://38.180.150.12/pages/product-design.html · 应用 App https://38.180.150.12.sslip.io/app/ · Artifact 链接见 TODO。

0. 一页总览The whole product on one page

1. 设计原则：人因工程与参考产品Design principles — human factors, and what we borrow from good apps

焦虑发作时注意力会收窄（Easterbrook 1959 的"线索利用"），工作记忆被占满（Sweller 认知负荷）。所以这个 App 的第一原则不是"功能多"，而是在用户最不能思考的时刻，把决定降到零。下面每条原则都写明它约束了哪个界面。

Under acute anxiety attention narrows (Easterbrook's cue-utilisation) and working memory is full (Sweller's cognitive load). So the first principle is not feature count but zero decisions at the moment the user can least think. Each principle below names the screen it constrains.

参考产品：借什么、避什么Reference apps — what to borrow, what to avoid

2. 系统架构图System architecture — sensors to sound to evidence

这张图把 ML_RESEARCH_DESIGN.md §5 的学习架构展开成产品部件。左边是数据从哪来，中间是手机上做什么，右边是声音怎么出、证据往哪去。粗线是每次会话必经的路径，虚线是离线 / 次日路径。

This expands the learning architecture in ML_RESEARCH_DESIGN.md §5 into product components: where data comes from (left), what the phone computes (middle), how sound is produced and where evidence goes (right). Solid lines run every session; dashed lines are offline or next-day.

3. 用户路径：从安装到听歌到反馈The user journey — from install through listening to feedback

横向滚动看 13 个阶段。每个阶段写明：用户做什么、采集什么数据、喂给哪个学术构念 / 信念 / 实验。高亮的是核心循环（7–11），用户每天走的就是这几步。

Scroll sideways through 13 stages. Each names what the user does, what data is captured, and which construct, belief or experiment it feeds. Highlighted stages (7–11) are the daily core loop.

← 横向滚动 · scroll sideways → · * 仅研究参与者 / pilot weeks only

每一步的成本预算The effort budget per step

4. 界面设计：八屏 + 手表 + 安全屏Screens — eight phone screens, the watch companion and the safety screen

视觉方向：一个温暖的深夜世界。炭褐底、亚麻白字、琥珀色只给"可以按的东西"，青色只在会话后表示"更平静"。没有插画堆叠，没有紫色渐变，没有表情符号；字体用系统字（SF Pro / 苹方）配等宽数字。每一屏下面写明"为什么这样"和"采集了什么"。

Visual direction: one warm night world. Charcoal-brown ground, linen text, amber only for what can be tapped, teal only after the session to mean "calmer". No illustration piles, no purple gradients, no emoji; system type (SF Pro / PingFang) with monospaced numerals. Each screen states why it is built this way and what it captures.

次级界面（不在核心循环里）Secondary screens outside the core loop

• 洞察 Insights：每周一张卡，"你的 Settle 在 22–23 点最有效 · 床 B 比床 A 更适合你 · 昨晚睡眠 HRV +6 ms"。柱状图七根，青色表示比基线更平静。One card per week: best hour, best bed, last night's sleep HRV; seven bars, teal = calmer than baseline.
• 曲库 Library：Unwind 按艺人 / 情绪 / 慢度浏览，标签见 §6；Settle 只有 4–6 张"床"，用质地命名（Fog · Linen · Low tide · Warm static），不用情绪词，避免暗示。Unwind browsable by artist / mood / slowness; Settle offers 4–6 beds named by texture, not emotion, to avoid suggestion.
• 数据与隐私 Data：同意开关、"哪些数据离开手机"、导出 JSON、删除全部。Consent toggles, what leaves the phone, export JSON, delete all.
• 研究模式 Research mode（隐藏，RA 用密码开启）：拉丁方条件、5 分钟预休息、STAI-6、期待值、Polar H10 配对、会话导出。Hidden, RA-unlocked: Latin-square condition, 5-min pre-rest, STAI-6, expectancy, Polar pairing, session export.

5. 数据采集接口Data-capture interfaces — function × signal × source × API

每个功能需要什么信号、从哪个接口来、多久一次、要什么权限、没有时怎么办。iOS 是首发平台（因为只有 Apple Watch 能给实时心率流）；Oura 只能给第二天；Polar H10 既是校准金标准，也是网页演示里唯一能实时读心率的途径（Web Bluetooth）。

For each function: the signal, its interface, rate, permission and fallback. iOS ships first because only Apple Watch gives a live HR stream; Oura is next-day only; Polar H10 is both the calibration reference and the one way a web demo can read live HR (Web Bluetooth).

6. 曲库与标签Library and tags — two catalogues, one schema, two playback tiers

6.1 两个曲库Two catalogues

6.2 标签规范（每首曲目 / 每张床一条）Tag schema — one record per track or bed

标签同时服务三件事：规则表选曲（slot）、bandit 的"参数格"（acoustic）、E4 的晋升 / 退役（posterior）。Spotify 的 audio-features 接口 2024 年 11 月起对新应用关闭，所以声学标签由我们自己标：自有音源用 librosa / Essentia 算，嵌入曲目人工标 + 试听核对。

Tags serve three jobs at once: rule-table selection (slot), the bandit's parameter cell (acoustic), and E4 promotion / retirement (posterior). Spotify's audio-features endpoint closed to new apps in November 2024, so we tag ourselves: librosa / Essentia on owned audio, manual tagging plus listening checks for embeds.

{ "id": "unwind-0007", "mode": "unwind",
  "title": "Weightless", "artist": "Marconi Union",
  "source": { "spotify": "6kkwzB6hXLIONkEk9JciA6", "soundcloud": "665599673" },
  "rights": "embed-only",            // owned | embed-only | licensed
  "adapt_tier": "S",                 // P parametric | S selection
  "fit": ["wind-down", "sleep-onset"],
  "acoustic": { "bpm": 60, "pulse_eq": 60, "brightness": "dark", "density": "sparse",
                "transients": "none", "lyrics": "none", "melody_salience": 1,
                "sub60hz": false, "lufs": -18, "key": "Dm" },
  "slot": { "start_hr": "<75", "role": "lead-down" },   // ≥90 | 75–90 | <75
  "qa": { "human": "2026-09-01", "notes": "long fades, no sharp onsets" },
  "posterior": { "n": 0, "mean": null, "sd": null, "prior": "pop:unwind:dark-sparse" } }

{ "id": "settle-B2", "mode": "settle", "name": "Linen",
  "source": { "owned": "/audio/settle/linen/", "stems": ["bed", "air", "warmth"] },
  "rights": "owned", "adapt_tier": "P",
  "fit": ["acute", "sleep-onset", "pre-performance"],
  "acoustic": { "pitch_centre": "none", "rhythm": "none", "centroid_hz": [1200, 600],
                "sub60hz": false, "transient_max_db": 6, "lufs": -16 },
  "params": { "cutoff_hz": [600, 2400], "density": [0, 1], "pulse_eq": [50, 66], "gain_db": [-6, 0] },
  "controller": "E1-rule-table-v0",
  "qa": { "human": "pending", "gate": "acute mode requires human QA" } }

20 首 Unwind 候选曲目见下表：从 12 个 Spotify 官方舒缓歌单出发，官方歌单里多为匿名"幽灵艺人"，因此又经 Wikidata 解析 17 位知名艺人的 Spotify ID，取各自"热门"榜（Spotify 按播放量排序）前 4 名随机抽一首、每位艺人一首。所有 ID 已用 Spotify oEmbed 校验；尚未人工试听，声学标签（bpm、亮度、密度）待试听后填写，急性模式不使用这些曲目。

The 20 Unwind candidates below start from 12 Spotify editorial calm playlists; because those are dominated by pseudonymous "ghost" artists, 17 well-known artists were resolved to Spotify IDs via Wikidata and one track per artist was drawn at random from the top four of each artist's Popular list (Spotify's own play-count order). Every ID is verified with Spotify's oEmbed endpoint; none has been listened to yet — acoustic tags (bpm, brightness, density) are filled after listening, and none of these enters the acute mode.

6.3 嵌入播放 + 参数适配：可行性实验台Embedded playback + parametric adaptation — a feasibility bench

下面这个小实验台证明两件事在浏览器里就能做：① 官方嵌入播放器可以被页面控制（Spotify iFrame API / SoundCloud Widget API：播放、暂停、跳转、事件）；② 自有音源走 Web Audio 时，低通截止与增益可以跟着心率每秒变化。心率在这里是模拟的（从 92 沉到 68），网页版可换成 Polar H10 的 Web Bluetooth 实时读数。这个实验台也部署在生产应用的"实验"页面里，可以用实时心率测试。

This bench shows two things work in a browser: ① official embeds can be driven by the page (Spotify iFrame API / SoundCloud Widget API: play, pause, seek, events); ② owned audio through Web Audio lets low-pass cutoff and gain follow heart rate second by second. HR here is simulated (92 settling to 68); the web version can swap in a Polar H10 over Web Bluetooth. The bench is also live in the deployed app's Lab screen, where you can test it with live heart rate.

6.4 权利与商用边界Rights and the non-commercial line

• 演示阶段不商用、不推广：真歌只通过 Spotify / SoundCloud 官方嵌入播放，不下载、不改音频、不去水印，符合两家的嵌入条款；Spotify 未登录只播 30 秒预览，登录 Premium 播全曲。Demo is non-commercial and unpromoted: real songs play only through official embeds — no download, no audio alteration — within both platforms' embed terms; Spotify plays 30-s previews unless the viewer is logged in to Premium.
• 参数级适配只对 TDMusic 自有或已确认改编权的曲目做（A5 法务核查 ≥ 50 首仍在进行）；急性模式的中性床全部自有，不需要曲库权利。Parametric adaptation only on TDMusic-owned or rights-cleared tracks (the A5 legal check on ≥ 50 tracks is still running); Settle beds are all owned and need no catalogue rights.
• 试点里 T2 的"舒缓版"应优先用自有曲目，避免把法务风险带进效果检验。In the pilot, T2 arrangements should prefer owned tracks so the efficacy test carries no rights risk.

7. 学术方法 → 产品交互：逐项映射Method → interaction — the mapping, item by item

这是老师最想看到的一张表：论文里的每个方法元素，在产品里长成了哪个交互，产出什么数据，推动哪个信念。反过来读也成立：产品里没有一个交互是没有研究理由的。

The table the supervisor will want: each method element in the dissertation, the interaction it becomes, the data it yields and the belief it moves. It reads backwards too: no interaction in the product lacks a research reason.

8. 反馈与学习闭环Feedback and the learning loop

9. 技术栈与"周一上线"计划Stack and the plan to release on Monday

需要你拍板的项目（第一周）Blocking items you must provide (week one)

1. Apple Developer Program（P4 需要；需要 24–48 小时注册）— 已有吗？Already have it?
2. Xcode 安装在 Mac 上（≈ 1–2 小时下载）— P4 需要。Needed for P4.
3. Spotify Premium 登录（后续使用；嵌入已用 30 秒预览工作）。Later; embeds already work with 30-s previews.
4. 自有音源：≥ 3 首慢 TDMusic 曲目文件；如可能 4 张 Settle 床按 §6.1 规格（合成噪声 + 缓慢 pad 作为临时方案）。≥ 3 slow TDMusic tracks as files; 4 Settle beds to §6.1 spec if possible (synth beds as fallback).
5. 20 曲候选表核查：保留 / 删除 / 增加。Confirm the 20-track candidate list: keep / drop / add.

10. 附录：v0 控制规则与安全规则Appendix — the v0 controller and safety rules

START   pulse_eq = HR ≥ 90 → 66 · 75–90 → 60 · < 75 → 56   (entrainment: start near the listener, then lead down)
        cutoff   = 600 + clamp((HR − 60) / 40, 0, 1) × 1800 Hz
        density  = HR ≥ 90 → dense → sparse over 3 min · else sparse
EVERY 60 s
        Δ = residual HR (this window) − residual HR (previous window)
        Δ ≤ −2 bpm            → pulse_eq −2 · cutoff −1 notch (−250 Hz) · keep layers
        Δ ≥ +3 bpm, no motion → hold · remove any transient layer
        else                  → hold
GUARDS  pulse_eq ≥ 50 · no onsets > bed +6 dB · loudness −16 LUFS fixed · nothing < 60 Hz
SAFETY  resting HR > 130 or panic tag → grounding screen, numbers hidden, sound continues
QUALITY window flagged if accel SD > 0.05 g or HR jump > 12 bpm; session < 70 % clean → no outcome shown
UNWIND  (tier S) next track = same rules applied to the catalogue's pulse_eq / brightness tags; 8-s crossfade
        (tier P, owned) identical to Settle, applied to the calm arrangement's stems

The same table drives the Wizard-of-Oz pilot (a human applies it), the v0 controller in the app, and the default arm of the later bandit — one rule set, three implementations, so the pilot result transfers to the product.