Innovating from the core: validating an AI-adaptive music product for anxiety relief inside a music-distribution company.

0Front matter

Abstract

This dissertation asks how an established, profitable AI music-distribution company can use its own assets to enter an adjacent health-adjacent market, and whether the evidence justifies doing so. The setting is TDMusic, a Beijing-headquartered distributor with a catalogue, rights infrastructure, a published recommendation algorithm and worldwide distribution, and the candidate venture is Lilt: a heart-rate-adaptive calming-audio product positioned as general wellness. The study runs as insider action research organised through design thinking, with a convergent mixed-methods evidence base — a six-stream literature and market review, six interviews, a 200-row netnography, two focus groups with a blind stimulus test, a pre-registered seven-block survey, and a within-subject three-condition efficacy pilot — integrated through a Bayesian layer in which every source becomes a likelihood ratio shrunk by its quality, and every decision is taken against thresholds fixed before the data. Business-school frameworks (VRIO, PESTEL, five forces, competitor and positioning analysis, TAM/SAM/SOM, SWOT, Value Proposition and Business Model Canvases, Ansoff, Three Horizons, stage-gate) carry the strategic analysis. The central empirical finding is a content reversal: users in acute anxiety reject the melodic, artist-linked music that is the company's core asset and ask for featureless adaptive sound, while preferring real music for evening wind-down. The posterior for real music in acute states fell to 8 per cent; in wind-down it stands at 80 per cent. The verdict at the time of writing is NOT YET: engagement and willingness-to-pay gates are cleared, the efficacy pilot has not run and the rights check is unfinished.

Word count 248 (limit 250). Sources: research/METHODOLOGY_DESIGN.md §0–1; research/DECISION_MODEL.md §3, §5; company/COMPANY_BACKGROUND.md §1–2; interviews/user-interviews.md; site/assets/decision.js.

Keywords

Corporate entrepreneurship · corporate innovation · design thinking · insider action research · mixed methods · Bayesian process tracing · stage-gate · music and anxiety · heart-rate variability · adaptive audio · digital wellness · willingness to pay · VRIO · PESTEL · Business Model Canvas.

Executive summary

Problem. TDMusic earns $10.5 M revenue and $2.2 M net income from music distribution and is a top-three distributor in China, a Tier-1 YouTube partner and a top supplier to Tencent Music (TDMusic Inc., 2022–2026). Its growth question is not whether it can distribute more music but whether its assets — catalogue, rights, an AI production pipeline, a published recommender and 220+ digital service providers — can be pointed at a new customer need. Music for anxiety relief is the candidate: anxiety affects an estimated 359 million people worldwide, of whom 27.6 per cent receive treatment (World Health Organization [WHO], 2023), and music listening lowers state anxiety across independent meta-analyses, albeit on trial evidence graded low quality (Bradt et al., 2013; de Witte et al., 2020).

Method. The founder-researcher runs the project as insider action research (Coghlan & Brannick, 2019), organised through design thinking and evaluated with decision analytics. Six evidence streams feed eight explicit beliefs; each belief has a prior with a written rationale, evidence items carrying likelihood ratios shrunk for quality, and a posterior computed in odds form (Fairfield & Charman, 2017; Humphreys & Jacobs, 2015). Proceed, pivot and kill thresholds were written down before the survey and before the pilot.

Findings. Four results carry the dissertation. First, a content reversal: three of three acute-anxiety interviewees independently rejected melody, lyrics and beat and asked for featureless sound; the focus-group blind test reproduced it (10 of 13 chose the neutral clip for the acute moment, 9 of 13 chose the real song for wind-down); the survey quantified it (24 per cent chose a real song in the acute scenario against 46 per cent neutral; 47 per cent chose the real song for wind-down; McNemar χ² = 14.29, p < .001). Second, engagement holds: 57 per cent of wearable owners would connect a device and 54 per cent value a measured result, against an "action gap" in which 38 per cent do nothing after a high stress reading. Third, price holds: the Van Westendorp range for US/EU contains the $6.99 test price and discounted purchase intent is 18 per cent. Fourth, efficacy is still borrowed: it rests on meta-analyses, not on the firm's own pilot, which has not run.

Recommendation. Do not fund an MVP yet. Split the product by arousal state rather than betting on one content type — an engineered neutral bed for the acute moment (which needs no catalogue rights) and real, artist-linked music for wind-down (which does). Run the legal check first because it is the cheapest decisive test, then the efficacy pilot, then re-apply the gates. Keep every claim inside the general-wellness line in all three regulatory regimes.

Decision status. NOT YET, as set out immediately below.

Decision status

Posteriors and gate arithmetic are computed exactly as site/assets/decision.js computes them, with the survey rows applied and the pilot and legal check pending. Gate rules: research/DECISION_MODEL.md §5. Dates: site/index.html roadmap; TODO_2026-08-19.md §11, §14–15.

1Introduction

1.1Context: TDMusic and the corporate-innovation question

TDMusic Inc. (太簇音乐, Tempered Digital Music) is a technology-driven digital-content distribution and monetisation company headquartered in Beijing and operating across the United States, Europe, Japan and China. Company documents report revenue of $10.5 M, net income of $2.2 M, a $70 M valuation and 52 employees, an angel round led by Challenger Capital, and a market position described as top-three music distributor in China, Tier-1 YouTube partner and top supplier to Tencent Music. Its 2022 catalogue figures record roughly 106,000 songs in reserve and 75,000 released, approximately 85 per cent owned or directly licensed, 1.05 billion streams between January and July 2022, and distribution to more than 200 countries through 220+ digital service providers.

The tutor's subject is innovation in the corporate. That framing matters because it changes what counts as an answer. A start-up would be asked whether the market is attractive; an established firm must additionally be asked whether it holds assets that make this market attractive to it, and whether it can run an exploratory venture without the exploitation business smothering it. The firm has already validated one instance of the general pattern: a Kindle Direct Publishing pilot that mined demand signals, produced content with AI assistance, published it and measured the result, reporting click-through rates two to three times benchmark, 4.2–4.5 star ratings and titles profitable within two to four months. The project reported here applies the same engine — demand signal, AI production, measured library — to the firm's core asset, music, in a vertical with a genuine clinical literature behind it.

1.2Problem statement

Three facts make the problem non-trivial. First, the need is large and under-served: an estimated 359 million people had an anxiety disorder in 2021 and roughly 27.6 per cent of those needing treatment receive it (WHO, 2023). Second, the incumbent solution set is commercially stressed: Calm's 2025 revenue is estimated at ~$210 M, down 24 per cent year on year with ~3.5 M subscribers, down ~500,000; Headspace's estimated revenue fell from ~$348 M in 2024 to ~$140 M ARR in 2025 alongside a 13 per cent workforce cut, and industry-wide 30-day retention for meditation apps is about 4.7 per cent with an average of one to four lifetime sessions (Creswell & Goldberg, 2025). Third, the evidence base the whole category leans on is thin: only Calm and Headspace have any randomised-trial coverage at all, half of Headspace's own trials disclose a conflict of interest, and no equivalent independent review exists for Endel, Brain.fm or any Chinese competitor (O'Daffer et al., 2022).

So the problem is not "can we build a calming-audio app" — plainly we can, and one is already in TestFlight. The problem is whether a firm whose advantage is real music and the rights to it can win in a market whose users, in the state where they most need help, may not want real music at all. That tension is the dissertation's spine.

1.3Research questions and objectives

Transcribed from research/METHODOLOGY_DESIGN.md §0.

The objectives follow directly: (i) establish whether the need exists and how it segments; (ii) establish which of the firm's assets actually transfer; (iii) measure whether one session changes anything, on self-report and on a device the user already owns; (iv) establish whether anyone pays, and at what price; (v) establish whether the rights permit the asset-leveraging version of the product; and (vi) end the phase with a documented decision rather than a pitch.

1.4Scope and boundaries

Four boundaries are worth stating explicitly because they constrain what this dissertation can conclude. The wellness boundary means no efficacy claim in this document is a medical claim, and no marketing language tested here uses the word "treat". The overseas-first boundary means the China figures in chapter 3 are context, not a market plan. The budget boundary means the efficacy pilot is n = 20–30, which the methodology declares in advance to be a feasibility signal for self-report and under-powered for physiology. The phase boundary means retention, the single most load-bearing commercial variable in this category, cannot be tested at all within the phase; it is carried as a watch item rather than a gate.

1.5Structure of the dissertation

Chapter 2 reviews the literature in five strands and names the gap. Chapter 3 analyses the company and the industry with the standard strategy toolkit. Chapter 4 sets out the method, including the Bayesian layer and the pilot design. Chapter 5 reports the findings by design-thinking mode and closes with the belief register. Chapter 6 turns the findings into a concept and a business model. Chapter 7 describes what has actually been built and how the algorithm works. Chapter 8 states the validation plan and the decision rule. Chapter 9 gives the roadmap, the risk register and the resource boundary. Chapter 10 answers the research questions as far as the evidence currently allows, and states what it cannot yet answer.

2Literature review

2.1Corporate entrepreneurship and ambidexterity

The project's own framing document names four lenses for treating this as corporate rather than start-up innovation. The Ansoff matrix places music for anxiety as "diversification-lite": a new customer need (health) served by an adapted product (music) built on existing assets. McKinsey's Three Horizons assigns AI music creation and distribution to Horizon 1 as the cash engine, the music-for-wellbeing product to Horizon 2, and a closed-loop "music as digital therapeutic" to Horizon 3. O'Reilly and Tushman's ambidexterity justifies a small autonomous team exploring while the distribution business exploits. And core-asset leverage lists the unfair advantages a pure start-up does not have: an AI music generation pipeline, owned and controlled catalogue and rights, artist relationships, worldwide distribution infrastructure, and market knowledge in China and globally.

Two references in this strand do carry bibliographic detail in the repository and are used for the real-options reading of the phase: Bowman and Hurry (1993) on real options and organisational learning, and McGrath (1999) on managing failure as options. Their function here is precise rather than decorative: they license the sequencing rule used in chapter 8, in which each test is treated as a purchase of information and ordered by expected posterior shift per unit cost.

2.2Design thinking as an innovation method

Design thinking supplies the process rather than the epistemology. The method architecture uses the d.school's five modes — empathise, define, ideate, prototype, test — paced by the Double Diamond, whose contribution is the divergence–convergence discipline: the requirement to widen before narrowing, and to record what was discarded. Brown (2008) and Liedtka (2018) are the two design-thinking references the repository carries; Liedtka's contribution to this project is the claim that design thinking's value lies in reducing cognitive bias in innovation decisions, which is exactly the claim the Bayesian layer in §2.5 operationalises.

The method is not used uncritically. Design thinking as normally practised produces artefacts (personas, jobs-to-be-done statements, point-of-view and how-might-we questions) whose epistemic status is ambiguous: they look like findings but are often the facilitator's synthesis. This dissertation therefore separates artefacts from results explicitly, and reports the personas and JTBD statements in chapter 5 as illustrative artefacts written in the framework document, not as data-derived findings — because the framework itself writes them as examples ("E.g. …") rather than as outputs of coding.

2.3Music, anxiety and autonomic arousal: what the evidence says

Music listening reliably reduces anxiety across independent meta-analyses, but the pooled effect varies widely with population and outcome type, and the underlying trials are graded low quality by the strictest reviewers. The largest and most rigorous synthesis is a Cochrane review of 26 perioperative trials (N = 2,051) which found music reduced state anxiety by −5.72 STAI-S units, 95% CI [−7.27, −4.17], performed comparably to midazolam in one N = 327 trial, and rated the evidence "low" quality because of a lack of blinding (Bradt et al., 2013). A broader stress meta-analysis of 104 randomised trials (N = 9,617) reported d = .545 on psychological markers and d = .380 on physiological ones, with heart rate the strongest physiological marker at d = .456 and hormone/cortisol the weakest at d = .349 (de Witte et al., 2020). Harney et al. (2023) report d = −0.77 across 21 controlled studies, rising to −0.97 in a low-bias subset. Across reviews the range is roughly d = −0.30 to −0.97.

All rows from research/SYNTHESIS.md §1 and §5 and research/findings/clinical-evidence-verified.md (fact-check verdicts: 7 confirmed, 1 corrected, 0 unsupported of 8 re-checked).

Three qualifications carry directly into the design. First, direction: relaxation raises heart-rate variability and lowers heart rate; the reverse is a recurring error in this literature's popular treatment and is checked explicitly throughout this document. Second, personalisation is weaker than intuition suggests: participant-selected music trended toward larger effects than researcher-selected music in the largest anxiety meta-analysis but the difference was not statistically significant, and preferred tempo appears driven more by an individual's spontaneous motor tempo than by genre preference (Harney et al., 2023). Third, generative audio has almost no evidence: the sole located peer-reviewed Endel-linked study measured focus rather than anxiety, was funded by Endel and Arctop, and used Arctop-employed authors (Haruvi et al., 2022).

2.4Digital wellness products and adaptive audio

Endel is the clearest commercial precedent for closed-loop biometric-adaptive audio: it modulates generative soundscapes using real-time Apple Watch heart rate compared against resting heart rate, plus motion, time of day and light; it won Apple's Watch App of the Year in 2020; and its primary press release with Universal Music Group cites three million monthly listeners (PR Newswire, 2023). Two features of the precedent matter more than the headline. First, the signal is heart rate, not HRV — HealthKit exposes HRV only as opportunistically sampled values that can lag up to about 30 minutes, and Oura's API returns HRV at five-minute granularity but only from sleep, available the following morning, with no streaming endpoint. Any product claiming "real-time HRV adaptation" is therefore either using heart rate or using stale values. Second, Endel's own strategic response to the "is this even music" critique — paying Grimes, James Blake, Miguel and Warner's Spinnin' Records for artist-attached functional music — is itself revealed evidence that abstract algorithmic sound under-satisfies demand.

The category's structural problem is retention, not acquisition. Roughly 4.7 per cent of meditation-app users are still using the app after 30 days, with an average of one to four lifetime sessions, even though Calm and Headspace jointly hold 96 per cent of category monthly active users (Creswell & Goldberg, 2025). Both leaders showed revenue and subscriber declines in 2025 while market-research firms continued to forecast double-digit compound growth — a tension this dissertation treats as a real signal about market maturity rather than resolving in either direction.

2.5Decision-making under uncertainty: Bayesian process tracing and stage-gates

The methodological move that distinguishes this project from a standard design-thinking write-up is that the "evaluate" step is not left to judgement. Each critical assumption is given an explicit prior with a written rationale; each piece of evidence is converted into a likelihood ratio using a published calibration rubric; likelihood ratios are shrunk in log space according to evidence quality; and posteriors are compared with thresholds fixed in advance. The method precedent is Fairfield and Charman's (2017) explicit Bayesian analysis for process tracing and Humphreys and Jacobs's (2015) Bayesian mixing of methods; evidence-quality grading follows the GRADE Working Group (Guyatt et al., 2008); and the pre-registration and kill-criterion discipline follows Ries (2011) and Bland and Osterwalder (2019).

The stage-gate reading sits on top: each test is a purchase of information whose cost, duration and expected posterior shift are listed before it is run, so the ordering of tests is a real-options decision rather than a scheduling convenience (Bowman & Hurry, 1993; McGrath, 1999).

2.6Gap and contribution

Four gaps are visible once the strands are laid side by side. (1) The efficacy literature does not cover the product. The strong evidence is for music listening in supervised, mostly perioperative settings, and for breathing-paced HRV biofeedback; there is no meta-analysis, and no adequately powered trial, of biometric-adaptive audio delivered by a consumer app. (2) The content question is unasked. No controlled anxiety trial directly compares real, personally meaningful music against engineered neutral sound — which is precisely the comparison on which this firm's asset advantage depends. (3) Corporate-innovation research rarely shows its decision arithmetic. Ambidexterity and stage-gate research says a firm should kill projects on evidence, but published cases seldom show the belief, the threshold and the update. (4) Design-thinking write-ups rarely distinguish artefact from result.

The contribution is correspondingly modest and specific: a fully documented, pre-registered corporate-innovation decision in which the belief state, the likelihood ratios, the quality shrinkage and the gates are all published and re-runnable, applied to a question — real music versus engineered neutral sound for acute anxiety — that the literature has not addressed, inside a firm whose principal asset happens to sit on one side of that question.

3Company and industry analysis

3.1Company profile and asset inventory: a VRIO analysis

The asset inventory is the "ability" half of the research question. The table below runs the seven assets recorded in the company file through VRIO — is the asset Valuable for this venture, is it Rare, is it costly to Imitate, and is the firm Organised to capture the value — and states the competitive implication that follows.

Assets and evidence transcribed from company/COMPANY_BACKGROUND.md §2; A5 posterior from research/DECISION_MODEL.md §3; budget from DESIGN_THINKING_FRAMEWORK.md framing update. VRIO verdicts: 28 cells (7 assets × V, R, I, O) — 17 supported by the cited evidence, 11 marked "to verify".

Reading of the VRIO. Only one asset — rights and catalogue — clears value, rarity and inimitability together, and its fourth column is exactly the question the pending legal check answers. That is a strategically uncomfortable result: the firm's single defensible asset is the one whose usability is unknown, and the user evidence in chapter 5 says that asset is the wrong content for the moment of highest need. Chapter 6 resolves this by segmenting the product rather than the asset.

3.2Macro environment: PESTEL

PESTEL: 6 of 6 cells sourced; 0 marked "to verify". The environmental cell is sourced as a reasoned "not material" rather than as a data point.

3.3Industry structure: Porter's five forces

Five forces: 5 of 5 rated with cited evidence; 0 "to verify".

3.4Competitor analysis

3.4.1 Competitor profiles

competitors-verified.md rows 1–24, 28–29; SYNTHESIS §2; market-wtp-verified rows 7, 49–51. Every dollar figure for Calm, Headspace and Endel traces to third-party aggregators (Latka, Business of Apps, Tracxn) that disagree with each other by wide margins; the direction is consistent, the precision is not.

3.4.2 Competitor marketing mix (4P)

Competitor 4P: 24 cells (6 players × 4 Ps) — 18 sourced, 6 marked "to verify". Competitor profiles: 36 cells, 33 sourced, 3 marked "to verify". Sources as §3.4.1 plus competitors-verified rows 5–7, 14, 19, 23, 29.

3.4.3 Positioning map

The map makes the strategic opening legible. The upper half is nearly empty: only MedRhythms combines real music with a regulatory-grade evidence position, and it does so in a narrow, prescription-gated, non-anxiety indication. Endel occupies the measurement corner without independent evidence; Brain.fm has evidence without user measurement; Calm, Headspace and Tide have neither. The position this project intends — measured on the user's own device, evidenced by a pre-registered trial, and content-segmented by arousal state — is unoccupied. It is also unproven, which is why the marker is dashed.

3.5Market sizing: TAM, SAM and SOM

Inputs: market-wtp-verified rows 1–2, 49, 51, 57–58; SYNTHESIS §3, §6; surveys/analysis/results.json via survey-results.html (wearable.d1_top2_owners, intent, vw.US_EU). TAM/SAM/SOM: 9 inputs sourced, 5 marked "to verify" (European revenue share, US/EU share of the wearable installed base, store commission, CAC, churn in months).

3.6SWOT

SWOT: 25 cells, all 25 sourced, 0 "to verify". Sources: COMPANY_BACKGROUND §1–2, §4; DECISION_MODEL §3; SYNTHESIS §2–§4, §6; survey results as published; focus-groups SYNTHESIS T3, T6–T7; TODO §11, §14–16; wearables-verified; regulation-verified.

4Methodology

4.1Research philosophy and insider action research

The frame is insider action research (Coghlan & Brannick, 2019): the founder-researcher diagnoses, plans, acts, evaluates and learns inside the firm. This is a genuine methodological commitment rather than a label, because it fixes what counts as rigour. An outsider's rigour is distance; an insider's rigour is declared procedure. Reflexivity and bias control are therefore part of the method (§4.7), not an appendix, and the project's own tutor presentation already positions the whole thesis as action research with five method components — action research, strategy formulation, innovation frameworks, data-driven content identification and quantitative pilot analysis. The design-thinking phase reported here is the next cycle of that same spiral: diagnose and plan for the highest-potential content vertical.

4.2Design-thinking process and artefacts

Process is supplied by the d.school's five modes paced by the Double Diamond; its contribution is the divergence–convergence discipline. Each mode produced named artefacts, listed below so the reader can distinguish what was produced from what was found.

4.3Mixed-methods design and the triangulation matrix

The architecture is a convergent mixed-methods design with a sequential quantitative core (Creswell & Plano Clark, 2018). The qualitative strand runs interviews (n = 6) → netnography (n = 200) → focus groups (2 × 6–8); the quantitative strand runs literature priors → survey (n ≥ 100) → pilot experiment (n ≥ 20); behavioural evidence (landing A/B) and the legal check enter alongside. Integration is explicit: each source enters the decision model as a likelihood ratio bounded by that source's quality, so nothing is double-counted and nothing is weighted by rhetoric.

4.4Instruments

Transcribed from research/METHODOLOGY_DESIGN.md §2 and surveys/questionnaire-v2.md §Analysis plan.

4.5The Bayesian decision engine

For an assumption H with prior probability P(H): prior odds O₀ = P(H) / (1 − P(H)); each evidence item contributes a likelihood ratio LRᵢ = P(Eᵢ | H) / P(Eᵢ | ¬H); posterior odds O = O₀ × ∏LRᵢ, and the posterior probability is O / (1 + O). Independence is assumed for the multiplication; where items are clearly correlated the likelihood ratios are shrunk toward 1 before multiplying rather than treated as independent.

Worked example, quoted from the model. Three acute-anxiety interviews rejecting real music would each be "strong" disconfirmation at LR ≈ 0.3. They are qualitative and self-selected, so k = 0.5 gives 0.55 each; three items recruited through one channel are treated as roughly two effective items, 0.55² ≈ 0.30, rounded conservatively up to 0.35 to avoid over-weighting three conversations. This is the single most consequential number in the dissertation and it is deliberately the most conservative reading of the evidence available.

Priors come from base rates or literature where any exist, otherwise 0.50; every prior carries a one-line rationale, and the live version of the model lets a reader move any prior and see how much the posterior depends on it. Priors are meant to be argued about, not hidden.

4.6Efficacy pilot design (E0 and E1)

E0 — calibration. Roughly eight participants wear an Apple Watch and a Polar H10 concurrently across three sessions. Outputs are the intraclass correlation and mean absolute percentage error for heart rate and RMSSD, plus a regression-calibration coefficient λ = Cov(watch, polar) / Var(watch) used to de-attenuate any model in which watch HRV is a regressor. Where HRV is an outcome, non-differential error costs power rather than causing bias, and is documented rather than corrected.

E1 — efficacy. A randomised within-subject crossover with three conditions in a Latin-square order: T1 adaptive neutral sound, heart-rate-driven; T2 an adaptive calm arrangement of a real song the participant chose; and C an active control — a generic "relaxing" playlist the participant already knows. The control is deliberately not silence, because silence changes expectancy and confounds "any audio" with "our audio".

Power, stated in advance. For a paired design at α = .05 two-sided and 80 per cent power: d_z = 0.5 needs n = 34; 0.6 needs 24; 0.65 needs 21; 0.8 needs 15. The Cochrane estimate (−5.7 STAI-S, SD ≈ 10, so d ≈ 0.55) implies n ≈ 28 for self-report, while physiological effects at d ≈ 0.4 need n ≈ 50. So n = 20 is a feasibility signal for A2a and under-powered for A2b, and this dissertation says so before the data rather than after. Two remedies are built in: three sessions per participant, and a Bayesian re-analysis with a meta-analytic prior in which the posterior, not a p-value, drives the gate.

Transcribed row for row from research/METHODOLOGY_DESIGN.md §4.3.

4.7Ethics, bias controls and limitations

Ethics. Wellness framing throughout, with no diagnosis and no medical claim in any interface copy. Consent is taken per data type (live heart rate, HRV, sleep). Data minimisation is architectural rather than promised: beat-level data are processed on device and only windowed features and session summaries leave the phone, and the user can delete everything. EU data are stored in the EU on a consent lawful basis; HealthKit terms forbid advertising use; Oura's terms gate the API behind membership. Fairness is treated as a measurement problem: photoplethysmography accuracy varies with skin tone, motion and temperature, so the quality index and calibration are reported by sub-group and no score is surfaced that would penalise a user for sensor noise. Safety: if resting heart rate exceeds 130 bpm or the user tags a panic state, the app shows a grounding message and hides all numbers, with no clinical routing beyond a general help line.

Limitations stated up front. Non-probability samples; hypothetical willingness to pay; a small pilot; consumer-grade physiology; a single coder on the netnography; the insider role; the independence assumption inside the concept joint probabilities; and subjective likelihood ratios — which are transparent and contestable, and that is exactly their merit.

5Findings

5.1Empathise: what each evidence stream said

Literature. Music lowers state anxiety, on low-quality trial evidence, with heart rate the strongest physiological marker and cortisol the weakest; slow tempo raises vagal tone; self-report and physiology need not move together; and generative audio has essentially no independent evidence (chapter 2).

Expert interview. Dr Wang, a dementia specialist, gave two findings that shaped the convergence: familiar music from a patient's youth works while purpose-made "treatment music" does not, which undercuts an AI-generated-music play for dementia; and monetisation for elderly care in China is hard.

User interviews (n = 3 acute). All three acute-anxiety interviewees independently rejected emotionally engaging music and asked for the opposite. Their language converges almost word for word:

The recurring complaints about existing neutral audio are the opening: sharp jarring frequencies, low hums that cause nausea, "healing" playlists that feel fake, and anything with a beat. These are not vague preferences; they are an engineering specification, and they became the survey's nine-item acoustic-tolerance matrix and the focus-group stimulus specs.

Netnography (n = 200 rows). Both preferences coexist: some reviewers fault Endel and Brain.fm for being "just lofi — not the song I expected", others explicitly value the absence of lyrics ("Endel feels less distracting because no hooks or lyrics"). The correct reading is therefore not "neutral sound wins" but "there are two distinct jobs". Separately, all ten rows coded for attitudes to biometrics concern the reliability of the reading and none concerns behaviour change — the action gap.

Competitor and market scan. As chapter 3.

5.2Define: the territory matrix, the artefacts, and the core tension

Three candidate territories entered Define — dementia, ADHD and anxiety — scored on six weighted criteria (clinical evidence .25, reachable market .20, willingness to pay .20, asset fit .15, regulatory ease .10, wearable synergy .10). Version 1 produced point totals of 1.80 (dementia), 2.65 (ADHD) and 4.55 (anxiety). Version 2 turned every cell into a triangular low/mode/high distribution, perturbed the weights by ±40 per cent and re-normalised across 5,000 Monte-Carlo draws, and revised anxiety's asset fit from 4 to 3 and wearable synergy from 5 to 4 after the interview evidence and the wearable-API facts. Anxiety's mode total fell from 4.55 to 4.30 and it still ranked first in more than 99 per cent of draws.

research/DECISION_MODEL.md §6. Cells are low / mode / high on a 0–5 scale. Sleep, chronic pain and depression entered the funnel as extras but were not scored in full for want of evidence — stated so the reader knows the funnel's real width.

The evidence supporting the two dementia and ADHD downgrades is worth stating because it is where the funnel actually narrowed: in dementia, a 2017 meta-analysis found a medium effect of individualised music on agitation (d = 0.61 across 12 trials) but the largest pragmatic trial to date (N = 463 residents, 54 nursing homes) found no significant effect on the primary agitation outcome; in ADHD, generic background music showed no effect on core attentional networks, while the more specific amplitude-modulation result used proprietary Brain.fm stimuli, and a 2024 systematic review found zero studies of brown noise meeting inclusion criteria.

Personas, JTBD, POV and HMW. These artefacts exist and are linked in Appendix A, but the framework document writes them as illustrative examples rather than as outputs of coding, so they are reported here as design artefacts and not quoted as findings.

5.3Survey results (as published)

All figures as published on site/pages/survey-results.html from surveys/analysis/results.json. Combined survey likelihood ratios applied to the belief register: A1a × 0.33 · A1b × 1.0 · A3 × 2.92 · A4 × 2.4.

5.4Focus groups (as published)

Two groups, 13 participants: FG-1 online, 82 minutes, seven participants from the US, UK and Germany (Apple Watch 4, Oura 2, none 1); FG-2 offline in a Shanghai co-working space, 78 minutes, six mainland-China office and hybrid workers. Both were moderated from a script by a research assistant with the founder absent. A blind stimulus test preceded discussion, with private ratings taken before anyone spoke.

Eight themes emerged, of which five carry into the design. T1 Five moments, not one "anxiety" (7/7 and 6/6): rumination after the day, sleep-onset, anticipatory or performance anxiety, notification spikes, and background tension — "by 1 am I'm not anxious about work any more, I'm anxious that I'm still awake"; "I don't need to sleep, I need to not shake for 20 minutes". T2 State-dependent sound (6/7 and 5/6): melody "gives my brain something to chase"; lyrics "start a story"; the Chinese phrase for "a soft wall" surfaced unprompted in FG-2, the same metaphor as interview I-03. Two participants rejected the neutral clip as "hospital air-conditioning" — disconfirming, and recorded, because it makes engineering quality decisive rather than optional. T3 The number cuts both ways (5/7 and 4/6): owners valued a result — "92 → 68, that I would screenshot" — but four said a live heart-rate line would make it "a test I can fail". T4 One tap or nothing (6/7 and 6/6): unanimous. T7 The action gap: owners read stress scores "like the weather", informative but not actionable.

The dot vote (three dots each, 13 participants) ranked: 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). Shrunk likelihood ratios entering the register, at k = .5 with a further ×.6 in log space for correlation with the interviews: A1a 0.6, A1b 1.4, A3 1.3, A4 0.9.

5.5The belief register: posteriors, gates and the verdict

How A1a fell from 0.50 to 0.08. Prior odds 1.00, then in order: the three acute interviews at 0.35 (shrunk from 0.3 raw, per §4.5) → 0.350; netnography, uninformative at 1.00 → 0.350; forum and Product Hunt users valuing "no hooks or lyrics" at 0.80 → 0.280; Harney's non-significant personalisation trend at 1.10 → 0.308; beat-entrainment plausibility at 0.80 → 0.246; and finally the survey's acute-scenario row at 0.33 → 0.081, giving P ≈ 0.075, reported as 8 per cent. Notice that the qualitative evidence alone took A1a only to about 0.20; it was the pre-registered survey row that crossed the pivot line. That is the shrinkage rule doing exactly what it was designed to do — qualitative evidence can move a belief, but it cannot by itself cross a gate.

The verdict is NOT YET, with the content pivot already triggered and acted on: the acute product becomes adaptive neutral sound, and the real-music catalogue becomes the wind-down mode and the C3 probe. A3 and A4 clear their gates. What remains is the efficacy pilot, which the A2 gates require by construction, and the legal check.

6Concept and business model

6.1Value Proposition Canvas

6.2Business Model Canvas

Business Model Canvas: 9 blocks, 42 cells, 35 sourced and 7 marked "to verify". Sources: COMPANY_BACKGROUND §1–2; DESIGN_THINKING_FRAMEWORK §0, framing update and §6; app/ARCHITECTURE.md §2, §4, §9; app/RECOMMEND_BRIEF.md; ML_RESEARCH_DESIGN §6; survey results as published; focus-groups SYNTHESIS T3–T6 and the dot vote; TODO §6–16.

6.3Marketing mix (4P) for Lilt

Van Westendorp and intent: survey results as published. Tagline and variants: prototype/landing/components/LandingPage.tsx and prototype/landing/README.md; product-design.html §3. Pre-registered landing LR: DECISION_MODEL §4. Language findings: focus-groups SYNTHESIS T5.

6.4Growth options: Ansoff and Three Horizons

6.5Revenue model and unit economics

The revenue model is a consumer subscription with two hedges: a licence into a partner device (C2) and a content line through the firm's own distribution (C3). What follows is the unit-economics arithmetic written out in full, with every input either sourced or flagged. It does not close, and the reason it does not close is itself a finding.

Unit economics: 7 inputs in all — 4 sourced, 3 marked "to verify"; the landing conversion is pre-registered but not yet observed. Retention base rate: Creswell & Goldberg (2025) via SYNTHESIS §6; A6 posterior: DECISION_MODEL §3.

Two structural observations follow from the arithmetic rather than from opinion. First, at a category-normal retention rate a standalone $6.99 subscription cannot support meaningful paid acquisition, which is the same conclusion the focus groups reached from the other direction when both groups preferred the feature bundled inside something they already pay for. Second, the single highest-leverage number in the whole business model is retention (A6), and it is the one number the design-thinking phase is structurally unable to test — it needs an MVP cohort. That is stated as a boundary of the phase, not solved inside it.

6.6Concept viability

The portfolio reading. The three concepts are not alternatives to be ranked once; they are a portfolio with different failure modes. C3 tests demand for real-music calm content at almost no cost and without any of the app's beliefs. C2 hedges the consumer-payment risk that the focus groups flagged and the market data corroborates. C1 is the asset-leverage bet, and the gates select which of them to fund rather than merely whether to fund anything.

7Product and technology

7.1From evidence to design

Principles table and screen references: site/pages/product-design.html §1 and §7; safety rule: ML_RESEARCH_DESIGN §6 and app/ARCHITECTURE.md §5. Human-factors sources named in the text are excluded from the reference list where the repository gives no year or venue — see the note under the references.

7.2The app today

The product is real, not a mock-up. The flow is: home (Settle or Unwind, duration, sleep fade) → pre-rest of 60 seconds, or 300 seconds in research mode, with a 0–10 self-rating → session → a 20-second check-in → a result screen → history, export or delete → settings. In Settle, Web Audio synthesises the neutral bed and steps the low-pass cutoff, gain, density and pulse every 60 seconds according to the v0 rule table. In Unwind, playback runs through one transport over the Spotify Web Playback SDK (PKCE auth, Premium required) with an embed fallback, owned files through Web Audio with a 1.5-second crossfade, and SoundCloud; tracks are chosen by tag, moment fit and heart-rate band, and the listener can skip at any time. Research mode adds the participant code, the Latin-square condition, the five-minute pre-rest, STAI-6 before and after, the two expectancy items, and a condition column in the export. A 72-second recording of the live build is on the hub; the heart rate in it is the simulated source.

7.3The adaptation algorithm

app/ARCHITECTURE.md §7; research/ML_RESEARCH_DESIGN.md §4.1.

7.4Recommendation and the learning loop

For Unwind the track choice is scored rather than shuffled:

The learning roadmap is deliberately staged and each stage has a guardrail. E2 randomly assigns a variant per session from a small allowed set and fits a hierarchical model of Δr and self-report on variant × moment with partial pooling. E3 replaces the fixed policy with a contextual bandit using Thompson sampling, with the hierarchical posterior as its prior; exploration is capped at 15 per cent of sessions, and any arm whose posterior mean reward is more than 0.3 SD worse than the default is suspended for that user. E4 lets the AI engine propose variants inside a parameter cell, each entering as an arm with a shrunk prior, promoted or retired by posterior thresholds — with a human listening QA gate before anything reaches the acute mode. Every policy change is evaluated offline first with inverse-propensity or doubly-robust estimation on logged data, then online against a hold-out default cohort; never a full rollout of an untested policy.

The reward is a composite by design: −standardised Δr on heart rate plus λ times the change in self-report, with λ starting at 0.5. Optimising on self-report alone is gameable by expectancy; optimising on heart rate alone is gameable by stillness. The composite plus the quality index guards both.

7.5Data, privacy and rights

Data. Consent is taken per data type. Beat-level data are processed on device and only windowed features and session summaries leave the phone. EU data are stored in the EU with consent as the lawful basis. HealthKit terms forbid advertising use of health data. Oura's API requires an active membership for Gen3 and later rings, so any Oura-based feature inherits a paywall the user may not have.

Rights. Catalogue entries carry an explicit rights tier — owned, embed-only or licensed — and an adaptation tier: P for parametric adaptation, permitted only on owned or cleared tracks, and S for selection-level adaptation, which is all an embedded player permits. Embeds are non-commercial and play 30-second previews unless the listener is signed in to Premium. The Settle beds need no catalogue rights at all, which is precisely why the neutral-sound concept survives a negative legal check. Track provenance for the Unwind pool is documented: Spotify editorial calm playlists, then ghost artists, then Wikidata-resolved artist identifiers, then the artists' Popular lists, with every identifier verified through oEmbed.

8Validation plan and decision

8.1The stage-gate

8.2What the pilot will settle, and the pipeline that will settle it

research/DECISION_MODEL.md §4. Positive for the pilot means p < .05 and d ≥ .3 in the correct direction.

The pipeline exists and has been exercised. Session exports flow through app/tools/export_to_csv.py into per-session and per-window CSVs, then through research/pilot/e1_analysis.py, which fits the pre-registered model, runs the Bayesian re-analysis, prints the likelihood-ratio table and the gate check mechanically, and writes the figures. Twenty-eight tests cover the schema, the exclusion rule, the order inference, the injected-effect direction and the gate arithmetic — including a test that the Latin square in the Python tools is still identical to the one in the app, so that a change to the app's randomisation table fails a test rather than silently mis-assigning conditions.

8.3The legal check and the landing test

The legal check (A5) is first in the queue and this is a deliberate real-options decision rather than a scheduling accident. It takes a little over a week, costs almost nothing, and is the only pending row graded decisive. If it returns negative at LR 0.15, A5 falls from 0.65 to roughly 0.22 and the real-music route closes on the spot — at which point no further engineering time should be spent on parametric adaptation of catalogue tracks, and C1-neutral becomes the only live app concept. Ordering tests by expected information per unit cost is what turns "real options" from a phrase into an execution order.

The landing test (A3, A4) supplies the only behavioural evidence in the entire design — everything else about willingness to pay is stated intent. Its pre-registered threshold is a visitor-to-waitlist conversion of at least 5 per cent per positioning variant. Two caveats attach to it, both stated in §6.3: no result exists yet, and the shared tagline and variant A's headline both predate the content pivot and should be rewritten before the test is read.

8.4Decision memo timing

The sequence is fixed: E0 calibration (25 August – 6 September) → E1 efficacy pilot (1–25 September) → Bayesian update and gate check (25 September – 2 October) → decision memo and supervision #2 (early October). The memo is written mechanically from the report: observed likelihood ratios replace the pending rows in the register with a date and a written reason, the same numbers are mirrored into the interactive page so the document and the site agree, and the gate check prints itself. Design thinking ends with a documented decision, not a pitch — and a documented KILL would be a valid outcome of this dissertation.

9Implementation roadmap and risks

9.1Roadmap

9.2Risk register

Risks R1–R3 and R5–R7 from the assumption register and the methodology's threats table; R4 and R8–R11 from journey.audit.md §E and METHODOLOGY_DESIGN §7 and §10; R6 and R12 from TODO_2026-08-19.md §14–16 and prototype/landing/README.md. Likelihood and impact ratings are the author's assessment against those sources.

9.3Resources and the budget boundary

The budget boundary is a methodological instrument, not an accounting detail. Ambidexterity requires that the exploratory unit be small enough and separate enough that killing it does not threaten anyone's career or the exploitation P&L; a ring-fenced $10–20k over three months is what makes the KILL branch of the gate credible. It also explains several design choices that would otherwise look like corner-cutting: a Wizard-of-Oz pilot instead of a fully instrumented native app; an n of 20–30 declared under-powered in advance rather than an adequately powered trial; a landing test with a ¥500–1,000 budget; and a legal check on 50 tracks rather than the whole catalogue. Each is the cheapest instrument that can move the belief it targets, which is exactly what the value-of-information ordering in chapter 8 selects for.

What the boundary does not buy is retention evidence, an adequately powered physiological test, or a signed partner conversation. Those are the three things a PROCEED decision would fund next, and naming them here is part of the decision memo rather than an afterthought.

10Conclusion

10.1Answers to RQ1 and RQ2 so far

RQ1 — is there a real, reachable, under-served need? Yes, with a qualification that changed the product. The need is real and frequent: 87 per cent of respondents need a calming moment at least weekly and 90 per cent already use music or audio to get it, against a global treatment gap in which 27.6 per cent of those needing care receive it. It is reachable: the primary segment self-identifies, self-pays and already owns the sensor. It is under-served in one specific way rather than generally — only one outcome, "stop the racing thoughts", scores above the opportunity threshold, and the wearable owners who already have a stress reading do nothing with it in 38 per cent of cases. And the qualification: the need is state-dependent, and so is the content that serves it. That was hypothesised by three interviews, reproduced in a blind stimulus test, and quantified in a pre-registered survey item with McNemar χ² = 14.29, p < .001.

RQ2 — can TDMusic serve it with its assets, measurably? Partly, and not yet demonstrably. Three of the five ability assumptions clear their pre-registered gates: users will connect a wearable and value the result (0.72), someone will pay at the tested price (0.82), and the rights position is more likely than not (0.65, pending the check). Two do not, and they are the two that matter most: that one session measurably lowers state anxiety (0.85 against a gate of 0.90) and that the same session moves a consumer wearable measure (0.56 against 0.70). Those gates were deliberately set above what the literature alone can reach. The honest answer to RQ2 today is therefore: the firm can build it, can distribute it, can probably clear the rights and can probably sell it — and has not yet shown that it works.

The decision. NOT YET. The content pivot has fired and been acted on; the remaining evidence is the pilot and the legal check; the memo is due in early October.

10.2Contribution to practice and to the firm

To practice. The contribution is a worked, publishable example of a corporate-innovation decision in which the belief state is explicit and re-runnable: eight assumptions, each with a prior and a written rationale; every evidence item converted to a likelihood ratio through a published rubric and shrunk by a published quality rule; thresholds fixed before the data; and the arithmetic exposed rather than summarised. The most instructive single episode is that qualitative evidence took the central belief only from 0.50 to about 0.20, and it was the pre-registered survey row that carried it across the pivot line — a concrete answer to the perennial question of how much three interviews should count.

To the firm. Three things. First, a finding the firm would otherwise have discovered after building: its principal asset is the wrong content for the moment of highest need, and the right response is to segment the product by arousal state rather than to abandon the asset or ignore the finding. Second, a portfolio rather than a bet — C3 tests real-music calm demand at almost no cost, C2 hedges consumer-payment friction, C1 is the asset-leverage play, and the gates select among them. Third, a reusable apparatus: a pre-registered survey pipeline, a pilot pipeline with 28 tests, a live decision engine, and a shipped product with research mode built in. The next vertical costs the firm the fieldwork, not the machinery.

10.3Limitations

• The survey and focus-group figures are pre-registered pipeline output as published on this site, not fieldwork — the results file carries a simulated flag and the focus-group synthesis is described in the framework as a pre-field template. This is stated wherever those figures appear and is the largest single limitation of the evidence base.
• The efficacy pilot has not run. Every efficacy statement in this document is borrowed from meta-analyses conducted in different populations, mostly perioperative, with supervised delivery and GRADE-low trial quality.
• Non-probability samples throughout, with hypothetical rather than revealed willingness to pay; the only behavioural instrument, the landing test, has no result.
• Consumer-grade physiology. Apple Watch HRV fails equivalence against a chest strap; the live loop is therefore heart-rate-driven, which measures autonomic arousal and is blind to its cause.
• Insider role. The founder is the researcher; the controls are declared and implemented, but the conflict is structural rather than removable.
• Subjective likelihood ratios and an independence assumption. The likelihood ratios are judgements made against a rubric; the concept joint probabilities multiply correlated beliefs. Both are transparent and contestable — which is their merit, not their defence.
• Company figures are dated. Catalogue and stream counts are 2022; the founding date and the artist-relationship scope are unreconciled in the company's own file.
• Single coder on the netnography, and the underlying netnography summary was not re-read for this document — only the row count and the patterns already carried in the decision model and the interview synthesis are used.
• Market sizing does not close. Two inputs required for a serviceable obtainable market do not exist in the evidence base, so no SOM point estimate is published.

10.4Further research

1. Run the comparison the literature has not run. A properly powered trial of engineered neutral sound against personally selected real music, in acute anxiety, with state anxiety as the primary outcome and residual heart rate as a co-primary. The E1 contrast β1 − β2 is a first, under-powered look at exactly this question; nothing in the published literature answers it.
2. Test whether closed-loop adaptation adds anything over a fixed calming stimulus. The strong biofeedback evidence is for breathing-paced protocols, not for music-driven adaptation; the generalisation gap should be named and then closed with an adaptive-versus-fixed arm.
3. Study retention as the dependent variable, not as a footnote. At a 4.7 per cent thirty-day base rate, retention dominates the unit economics of the entire category; whether anchoring a habit to hardware the user already checks daily changes that rate is an empirical question with a clear design.
4. Validate the residual as a user-facing quantity. "Calmer than your usual 22:00 by 9 bpm" is simultaneously an outcome, a reward signal and a piece of interface copy. Whether users read it as credible, and whether reading it changes behaviour, is a distinct study from whether the underlying quantity moves.
5. Extend the decision apparatus to a second vertical. The strongest test of the methodological contribution is whether the same register, rubric and gates produce a defensible decision somewhere the answer is not already suspected.

 References

1. Above Avalon. (2025). Apple Watch [Analyst notes]. https://www.aboveavalon.com/notes/tag/Apple+Watch
2. Advertising Law of the People's Republic of China (2015 revision), arts. 17–18, 58–59. https://extranet.who.int/fctcapps/sites/default/files/2024-02/341_%E4%B8%AD%E5%8D%8E%E4%BA%BA%E6%B0%91%E5%85%B1%E5%92%8C%E5%9B%BD%E5%B9%BF%E5%91%8A%E6%B3%95.pdf
3. Agrawal, S., & Goyal, N. (2013). Thompson sampling for contextual bandits. [Proceedings, volume and pages to verify]
4. Apple Inc. (n.d.-a). HealthKit [Developer documentation]. https://developer.apple.com/documentation/healthkit [Publication or retrieval date to verify]
5. Apple Inc. (n.d.-b). Vitals on Apple Watch [Support documentation]. https://support.apple.com/guide/watch/vitals-apd15aa7ed96/watchos [Publication or retrieval date to verify]
6. Apple Watch Series 9 and Ultra 2 heart-rate-variability validation study. (2024). Sensors. [Authors, exact title, volume and article number to verify] https://pmc.ncbi.nlm.nih.gov/articles/PMC11478500/
7. Behavioral Health Business. (2024, November 20). Headspace axes 13% of workforce, transitions therapist network to part-time and contract roles. https://bhbusiness.com/2024/11/20/headspace-axes-13-of-workforce-transition-therapist-network-to-part-time-and-contract-roles/
8. Berger, C., [remaining authors to verify]. (1993). Better/Worse coefficients for Kano categories. [Exact title, journal, volume and pages to verify]
9. Bernardi, L., Porta, C., & Sleight, P. (2006). Cardiovascular, cerebrovascular and respiratory changes induced by music. Heart, 92, 445. [Issue and page range to verify; the repository also records this work as Circulation, 114(24), 2571–2577 — the two records conflict and must be reconciled] https://www.ahajournals.org/doi/full/10.1161/CIRCULATIONAHA.108.806174
10. Blakeley, R. (2024, January 18). Welcome to the sound wellness revolution: Endel's AI-generated soundscapes and the commodification of passive listening. Musicology Now. https://musicologynow.org/welcome-to-the-sound-wellness-revolution-endels-ai-generated-soundscapes-and-the-commodification-of-passive-listening/
11. Bland, D. J., & Osterwalder, A. (2019). Testing business ideas. [Publisher and place to verify]
12. Bloomberg. (2026, January 5). Smart rings poised for 2026 growth, Oura set to lead. https://www.bloomberg.com/news/articles/2026-01-05/smart-rings-poised-for-2026-growth-oura-set-to-lead
13. Bowman, E. H., & Hurry, D. (1993). Real options and organisational learning. [Exact title, journal, volume and pages to verify]
14. Bradt, J., Dileo, C., & Shim, M. (2013). Music interventions for preoperative anxiety. Cochrane Database of Systematic Reviews, CD006908. [Issue number to verify] https://doi.org/10.1002/14651858.CD006908.pub2
15. Braun, V., & Clarke, V. (2006). [Title to verify — the repository records only the journal, volume and start page]. Qualitative Research in Psychology, 3, 77. [Issue and end page to verify]
16. Brown, T. (2008). Design thinking. [Journal, volume, issue and pages to verify]
17. Bundesinstitut für Arzneimittel und Medizinprodukte. (2026). DiGA-Verzeichnis [Directory of digital health applications]. https://www.diga-verzeichnis.de/en
18. Business of Apps. (2026a). Calm revenue and usage statistics. https://www.businessofapps.com/data/calm-statistics/
19. Business of Apps. (2026b). Headspace revenue and usage statistics. https://www.businessofapps.com/data/headspace-statistics/
20. Calm Health. (2026). Calm Health [Company website]. https://health.calm.com/
21. Carroll, R. J., [remaining authors to verify]. (2006). Measurement error in nonlinear models. [Edition, publisher and place to verify]
22. Christensen, C. M., [remaining authors to verify]. (2016). Jobs to be done: Current alternatives as the competitive set. [Exact title and publication details to verify]
23. Coghlan, D., & Brannick, T. (2019). Doing action research in your own organization. [Edition, publisher and place to verify]
24. Cohen, S., & Williamson, G. (1988). Perceived stress in a probability sample of the United States. In S. Spacapan & S. Oskamp (Eds.), The social psychology of health. Sage. [Page range and place of publication to verify]
25. Creswell, J. D., & Goldberg, S. B. (2025). Meditation-app engagement and retention. [Exact title, journal, volume, issue and pages to verify] https://pmc.ncbi.nlm.nih.gov/articles/PMC12333550/
26. Creswell, J. W., & Plano Clark, V. L. (2018). Designing and conducting mixed methods research. [Edition, publisher and place to verify]
27. Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319–340.
28. de Witte, M., Spruit, A., van Hooren, S., Moonen, X., & Stams, G.-J. (2020). Effects of music interventions on stress-related outcomes. [Subtitle to verify — the repository records the title without one] Health Psychology Review, 14(2), 294–324. https://doi.org/10.1080/17437199.2019.1627897
29. Devilly, G. J., & Borkovec, T. D. (2000). Credibility and expectancy questionnaire. Journal of Behavior Therapy and Experimental Psychiatry, 31, 73. [Exact title, issue and end page to verify]
30. Dudík, M., Langford, J., & Li, L. (2011). Doubly robust policy evaluation and learning. [Proceedings, volume and pages to verify]
31. Easterbrook, J. A. (1959). The effect of emotion on cue utilisation and the organisation of behaviour. [Journal, volume, issue and pages to verify]
32. Fairfield, T., & Charman, A. E. (2017). Explicit Bayesian analysis for process tracing: Guidelines, opportunities, and caveats. Political Analysis, 25(3), 363–380.
33. Federal Trade Commission. (2024, April). Updated FTC Health Breach Notification Rule puts new provisions in place to protect users of health apps [Business guidance blog]. https://www.ftc.gov/business-guidance/blog/2024/04/updated-ftc-health-breach-notification-rule-puts-new-provisions-place-protect-users-health-apps
34. Fierce Biotech. (2023, April). Prescription app developer Pear Therapeutics files for bankruptcy, lays off staff. https://www.fiercebiotech.com/medtech/cut-core-prescription-app-developer-pear-therapeutics-files-bankruptcy-lays-staff
35. Flanagan, J. C. (1954). The critical incident technique. Psychological Bulletin, 51(4), 327–358.
36. Fortune Business Insights. (2025). Mental health apps market size, share & global report. https://www.fortunebusinessinsights.com/mental-health-apps-market-109012
37. Gelman, A., & Hill, J. (2007). Data analysis using regression and multilevel/hierarchical models. [Publisher and place to verify]
38. Goessl, V. C., Curtiss, J. E., & Hofmann, S. G. (2017). The effect of heart rate variability biofeedback training on stress and anxiety: A meta-analysis. Psychological Medicine, 47(15). [Page range to verify] https://www.cambridge.org/core/journals/psychological-medicine/article/effect-of-heart-rate-variability-biofeedback-training-on-stress-and-anxiety-a-metaanalysis/A839E9C968E54774DF5C8FB186764EF0
39. Grand View Research. (2025). Meditation management apps market report. https://www.grandviewresearch.com/industry-analysis/meditation-management-apps-market-report
40. Guyatt, G. H., [remaining authors to verify] (GRADE Working Group). (2008). [Title to verify]. BMJ, 336, 924. [Issue and end page to verify]
41. Harney, C., Johnson, J., Bailes, F., & Havelka, J. (2023). [Title to verify]. Musicae Scientiae. [Volume, issue and pages to verify] https://doi.org/10.1177/10298649211046979
42. Haruvi, A., Kopito, R., Brande-Eilat, N., Kalev, S., Kay, E., & Furman, D. (2022). [Title to verify]. Frontiers in Computational Neuroscience, 15, 760561. https://doi.org/10.3389/fncom.2021.760561
43. Healthcare Brew. (2025, April 4). Digital therapeutics Medicare coverage test begins. https://www.healthcare-brew.com/stories/2025/04/04/digital-therapeutics-medicare-coverage-test-begins
44. Healthcare Dive. (2021, August). Headspace, Ginger to merge, creating $3B mental health company. https://www.healthcaredive.com/news/headspace-ginger-to-merge-create-3b-mental-health-company/605632/
45. Hernando, D., Roca, S., Sancho, J., Alesanco, Á., & Bailón, R. (2018). [Title to verify]. Sensors, 18. [Issue and article number to verify] https://pmc.ncbi.nlm.nih.gov/articles/PMC6111985/
46. Huang, Y., [remaining authors to verify]. (2019). Prevalence of mental disorders in China: A cross-sectional epidemiological study. The Lancet Psychiatry, 6(3), 211–224. https://pubmed.ncbi.nlm.nih.gov/30792114/
47. Humphreys, M., & Jacobs, A. M. (2015). Mixing methods: A Bayesian approach. American Political Science Review, 109(4), 653–673.
48. iiMedia Research. (2023). 2023–2024 China sleep economy industry development and consumer demand research report. https://www.iimedia.cn/c400/92947.html
49. iiMedia Research. (2025). 2025–2029 China emotion economy consumption trends insight report. https://www.iimedia.cn/c400/106827.html
50. Kano, N., Seraku, N., Takahashi, F., & Tsuji, S. (1984). Attractive quality and must-be quality. Journal of the Japanese Society for Quality Control, 14(2), 39–48.
51. Kirk, R., & Timmers, R. (2025). Characterizing music for sleep: A comparison of Spotify playlists. Musicae Scientiae. [Volume, issue and pages to verify] https://doi.org/10.1177/10298649241269011
52. Kozinets, R. V. (2020). Netnography. [Edition, publisher and place to verify]
53. Kroenke, K., Spitzer, R. L., Williams, J. B., Monahan, P. O., & Löwe, B. (2007). [Exact title to verify — the GAD-2 validation paper]. Annals of Internal Medicine, 146, 317–325. [Issue number to verify]
54. Krueger, R. A., & Casey, M. A. (2015). Focus groups. [Edition, publisher and place to verify]
55. Lassner, [initials and remaining authors to verify]. (2025). [Title to verify]. BJPsych Open, 11(1), e4. [Publication year recorded in the repository as 2024/2025 — to reconcile] https://pmc.ncbi.nlm.nih.gov/articles/PMC11733488/
56. Latka. (2026). Calm revenue. https://getlatka.com/blog/calm-revenue
57. Liedtka, J. (2018). On the effectiveness of design thinking. Journal of Product Innovation Management. [Exact title, volume, issue and pages to verify]
58. Lipponen, J. A., & Tarvainen, M. P. (2019). RR-interval artefact correction. Journal of Medical Engineering & Technology, 43, 173. [Exact title, issue and end page to verify]
59. Lu, [given name to verify], [remaining authors to verify]. (2021). [Title to verify]. PeerJ Computer Science, 7, e716. [The repository records the byline only as “Lu, Qiao et al.”; whether “Lu Qiao” is one author or two must be reconciled from the article]
60. Marteau, T. M., & Bekker, H. (1992). A six-item short form of the state scale of the Spielberger State–Trait Anxiety Inventory. British Journal of Clinical Psychology, 31, 301. [Exact title, issue and end page to verify]
61. McCreedy, E. M., [remaining authors to verify]. (2021). [Title to verify]. Journal of the American Medical Directors Association. [Volume, issue and pages to verify] https://www.jamda.com/article/S1525-8610(21)01104-X/abstract
62. McGrath, R. G. (1999). [Title to verify]. Academy of Management Review, 24, 13. [Issue and end page to verify]
63. McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions. [Journal, volume, issue and pages to verify; the repository records the title without its final words]
64. Mendes, [initials to verify], de Paula, [initials to verify], & Miranda, [initials to verify]. (2024). [Title to verify]. Interactive Journal of Medical Research. [Volume, issue and article number to verify] https://pmc.ncbi.nlm.nih.gov/articles/PMC11294770/
65. Mitrovic, [initials to verify], & Paladin, [initials to verify]. (2026). [Title to verify]. Frontiers in Cardiovascular Medicine, 13. https://doi.org/10.3389/fcvm.2026.1841349
66. Morgan, D. L. (1997). Focus groups as qualitative research. [Edition, series, publisher and place to verify]
67. Morwitz, V. G., Steckel, J. H., & Gupta, A. (2007). When do purchase intentions predict sales? International Journal of Forecasting, 23, 347–364. [Issue to verify]
68. Music Business Worldwide. (2023, May). Universal inks deal with generative AI startup Endel to create "AI-powered, artist-driven functional music". https://www.musicbusinessworldwide.com/universal-inks-deal-with-generative-ai-startup-endel-to-create-ai-powered-artist-driven-functional-music/
69. National Institute for Health and Care Excellence. (2023). Digitally enabled therapies for adults with anxiety disorders: Early value assessment (HTE9). https://www.nice.org.uk/guidance/hte9
70. National Institute of Mental Health. (n.d.). Any anxiety disorder. https://www.nimh.nih.gov/health/statistics/any-anxiety-disorder [Publication or retrieval date to verify]
71. NetEase Cloud Music Inc. (2025, February 20). NetEase Cloud Music Inc. reports fiscal year 2024 financial results [Press release]. PR Newswire. https://www.prnewswire.com/news-releases/netease-cloud-music-inc-reports-fiscal-year-2024-financial-results-302381265.html
72. Nigg, J. T., [remaining authors to verify]. (2024). [Title to verify]. [Journal, volume, issue and pages to verify] https://pubmed.ncbi.nlm.nih.gov/38428577/
73. O'Daffer, A., Colt, S. F., Wasil, A. R., & Lau, N. (2022). Efficacy and conflicts of interest in RCTs evaluating Headspace and Calm apps: Systematic review. JMIR Mental Health, 9(9), e40924. https://pmc.ncbi.nlm.nih.gov/articles/PMC9533203/
74. Oura Health. (2024). The Oura API [Support documentation]. https://support.ouraring.com/hc/en-us/articles/4415266939155-The-Oura-API
75. Panteleeva, Y., Ceschi, G., Glowinski, D., Courvoisier, D. S., & Grandjean, D. (2018). [Title to verify]. Psychology of Music, 46(4), 473–487. https://doi.org/10.1177/0305735617712424
76. Pedersen, S. K. A., Andersen, P. N., Lugo, R. G., Andreassen, M., & Sütterlin, S. (2017). [Title to verify]. Frontiers in Psychology, 8, 742. https://pmc.ncbi.nlm.nih.gov/articles/PMC5432607/
77. Pelletier, C. L. (2004). [Title to verify]. Journal of Music Therapy, 41(3), 192–214. https://pubmed.ncbi.nlm.nih.gov/15327345/
78. PR Newswire. (2020, June). MedRhythms receives FDA Breakthrough Device Designation for chronic stroke digital therapeutic [Press release]. https://www.prnewswire.com/news-releases/medrhythms-receives-fda-breakthrough-device-designation-for-chronic-stroke-digital-therapeutic-301076597.html
79. PR Newswire. (2023, May 23). Endel and Universal Music Group to create AI-powered, artist-driven functional music designed to support listener wellness [Press release]. https://www.prnewswire.com/news-releases/endel-and-universal-music-group-to-create-ai-powered-artist-driven-functional-music-designed-to-support-listener-wellness-301832191.html
80. PR Newswire. (2025, January 16). MedRhythms' InTandem rehabilitation system for chronic stroke gait impairment receives final Medicare payment determination from CMS [Press release]. https://www.prnewswire.com/news-releases/medrhythms-intandem-rehabilitation-system-for-chronic-stroke-gait-impairment-receives-final-medicare-payment-determination-from-cms-302353537.html
81. Prova Health. (2025). DiGA reimbursement in Germany: A guide. https://www.provahealth.com/insights/diga-reimbursement-germany-guide [Exact publication date to verify — the repository records 2024–2025]
82. Ries, E. (2011). The lean startup. [Publisher and place to verify]
83. Rogers, E. M. (2003). Diffusion of innovations (5th ed.). [Publisher and place to verify]
84. Schmitz, A., Kueppers, L., Klein, J., [remaining authors to verify]. (2025). [Title to verify]. JMIR Research Protocols, 14, e63380. https://www.researchprotocols.org/2025/1/e63380
85. Shaffer, F., & Ginsberg, J. P. (2017). An overview of heart rate variability metrics and norms. Frontiers in Public Health, 5, 258.
86. Spielberger, C. D. (1983). State–Trait Anxiety Inventory manual. [Edition, publisher and place to verify]
87. Tarvainen, M. P., [remaining authors to verify]. (2002). [Title to verify — smoothness-priors detrending]. IEEE Transactions on Biomedical Engineering, 49, 172. [Issue and end page to verify]
88. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. (1996). [Title to verify — the repository records the work only as "HRV standards"]. Circulation, 93, 1043. [Issue and end page to verify]
89. TDMusic Inc. (2022–2026). Company background and asset inventory [Internal documents: business-model deck (2024), thesis tutor presentation (2026), company brochure (2022)]. [Individual document titles, authors and dates to verify; the founding date and the 2022 catalogue figures are flagged for reconciliation in the company's own file]
90. Tebra. (2025). Music for mental health [Healthcare report]. The Intake. https://www.tebra.com/theintake/healthcare-reports/music-for-mental-health
91. TechCrunch. (2022, April 5). Endel raises $15M to further develop its AI-powered sound wellness technology. https://techcrunch.com/2022/04/05/endel-raises-15m-to-further-develop-its-ai-powered-sound-wellness-technology/
92. TechCrunch. (2025, September 16). Calm launches standalone iOS app for sleep support. https://techcrunch.com/2025/09/16/calm-launches-standalone-ios-app-for-sleep-support/
93. 36Kr. (2026). 从"香饽饽"到"鸡肋"，睡眠 APP 困于商业化 [From hot commodity to dispensable: Sleep apps stuck on monetisation]. https://36kr.com/p/1852579050655107
94. TMTPost. (2023, April). [Title to verify — NMPA classification determination for cognitive-function-disorder software]. https://www.tmtpost.com/6473884.html
95. Tracxn. (2026). Calm [Company profile]. https://tracxn.com/d/companies/calm/__11__6TilJ32thRjVlCh0DLinrJLZ9o47hTQaLwndAnI
96. Ulwick, A. W. (2002). Turn customer input into innovation. Harvard Business Review, 80(1), 91–97.
97. U.S. Food and Drug Administration. (n.d.). General wellness: Policy for low risk devices [Guidance document]. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/general-wellness-policy-low-risk-devices [Publication and update dates to verify; the repository records the guidance as cited 2019–2026 with a January 2026 update reported only by secondary law-firm sources]
98. van Westendorp, P. H. (1976). NSS-Price Sensitivity Meter (PSM): A new approach to study consumer perception of price. Proceedings of the ESOMAR Congress, Venice. [Publisher and page range to verify]
99. VBData. (2026). [Title to verify — Tide (潮汐) Pre-A funding announcement]. https://www.vbdata.cn/40749
100. Venkatesh, V., & Davis, F. D. (2000). [Exact title to verify — the TAM2 extension paper]. Management Science, 46(2), 186–204.
101. Vickers, A. J., & Altman, D. G. (2001). Analysing controlled trials with baseline and follow-up measurements. BMJ, 323, 1123. [Issue and end page to verify]
102. Virtual Therapeutics. (2024, July). Virtual Therapeutics announces results of tender offer to acquire Akili Interactive [Press release]. Nasdaq. https://www.nasdaq.com/press-release/virtual-therapeutics-announces-results-tender-offer-acquire-akili-interactive-2024-07
103. Weiser, M., & Brown, J. S. (1996). The coming age of calm technology. [Publication details to verify]
104. Woods, K. J. P., Sampaio, G., James, T., [remaining authors to verify]. (2024). [Title to verify]. Communications Biology, 7, 1376. https://www.nature.com/articles/s42003-024-07026-3
105. World Health Organization. (2023). Anxiety disorders [Fact sheet]. https://www.who.int/news-room/fact-sheets/detail/anxiety-disorders

105 entries; 65 carry at least one bracketed field the repository does not supply. Every gap is itemised in site/pages/dissertation.references-todo.md, grouped by the kind of field missing, so the gaps can be closed in one library session rather than one at a time.

 Appendices

ASurvey instrument map

The full seven-block instrument, with every item, its wording in both languages, the construct it measures, the pre-registered threshold it carries and the belief it feeds. Blocks: A who · B why, when and where (GAD-2, PSS-4, context, coping, outcome-driven innovation) · C what works (the state-dependent forced choice that produced the central finding) · D how to prove it (wearable willingness, the action gap) · E why use it and why not (technology acceptance, Kano, form, barriers) · F how much (Van Westendorp and purchase intent) · G voice.

BFocus-group kit

Recruitment screener, moderator script, the three blind stimulus specifications (S-A calm real song; S-B neutral adaptive sound; S-C guided voice), the private-rating sheet used before discussion, the coding frame, and the synthesis of both groups with the theme matrix and the dot vote.

CDecision-engine register

The live belief register: eight assumptions with priors and rationales, every evidence item with its likelihood ratio, quality tag and source, the pending pre-registered rows, the concept joint probabilities and the gate check. The page is interactive — priors can be moved and pending evidence toggled, so a reader can see for themselves how much the verdict depends on any single judgement.

DApp blueprint

The ten-section product blueprint: human-factors principles with their reference products, the system architecture, the thirteen-stage user journey, eight screens plus the watch and safety screens, the data-capture interface matrix (HealthKit, Oura API v2, Polar H10 over BLE, Web Bluetooth), the two-tier library with its tag schema and an embedded parameter-adaptation bench, the academic-method-to-interaction mapping table, the feedback and learning loop, the technology stack, and the v0 control rules.

EGlossary