> For the complete documentation index, see [llms.txt](https://tonify-1.gitbook.io/tonify-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://tonify-1.gitbook.io/tonify-docs/reference/sources.md).

# Sources

Every input with its status — measured, derived, assumed, or a model parameter.

Every quantity in this study carries one of four statuses, and the status is part of the claim:

| Status                  | Meaning                                                                        |
| ----------------------- | ------------------------------------------------------------------------------ |
| **M — measured**        | comes from data or a published measurement, with the source named              |
| **D — derived**         | arithmetic over measured quantities                                            |
| **A — assumed**         | a modelling assumption; the study says so and shows what breaks if it is wrong |
| **P — model parameter** | a dial of the simulation, not a claim about the world                          |

The rule that makes this useful: an assumed number may never be reported as if it were measured. Where a source is weaker than it looks — an industry figure without a methodology, a panel statistic normalised over a lifetime rather than a year — the caveat travels with the number everywhere it is used.

***

*🇬🇧 English |* [*🇷🇺 Русский*](https://github.com/ProximaCA/tonify-sims/blob/main/SOURCES.ru.md)

## SOURCES — every parameter, its value, its provenance

One row per model parameter: the value the simulations actually use, the source as verified in the research vault (tonify-research, 1,152 notes), the vault note id, and where the parameter enters the code/figures. Rows where the vault does **not** confirm the repo's attribution say so explicitly — see *Discrepancies* below the table. Nothing in this file is cited from memory; every link was read from a vault note. Where the vault has no source, the row says `assumption` or `not in vault` instead of inventing one.

| #  | Parameter                                   | Value used                                                                                | Source (as verified in vault)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           | Used in                                                             |
| -- | ------------------------------------------- | ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
| 1  | Share below 1,000 streams/yr (target T1)    | 87%                                                                                       | Closest verified: Luminate year-end 2023 — 86.2% of **tracks** (not artists) ≤1,000 plays, via MBW 14 Mar 2024 ([link](https://www.musicbusinessworldwide.com/deezer-has-deleted-26m-useless-tracks-since-it-launched-artist-centric-model-with-universal-music-group/)); artist-side analogue: Chartmetric 2025 — 86% of Spotify artists <10 monthly listeners, via Kullick & Petry 2025 ([PDF](https://www.cambridge.org/core/services/aop-cambridge-core/content/view/61644B64A790CB6B4B15A36C3D6DF83C/S2059599925100150a.pdf)). See discrepancy D1. | sim1 world construction, validation T1; fig2                        |
| 2  | Rightsholders above $1,000/yr               | 2.6% (259,700 of 10M+ uploaders, 2023)                                                    | Spotify Loud & Clear (2023 data), via Music Ally, 14 Jan 2025 ([link](https://musically.com/2025/01/14/chartmetric-tracks-11m-spotify-artists-fewer-than-1-6m-have-over-10-listeners/))                                                                                                                                                                                                                                                                                                                                                                 | sim1 validation T2; CRITIC §5 bound (rightsholders, not artists)    |
| 3  | Top share of streams (target T3)            | top-0.28% ≈ 50%; Gini 0.72                                                                | Òscar Celma, PhD thesis *Music Recommendation and Discovery*, UPF Barcelona (Last.fm crawl, July 2007: top-737 of 260,525 artists = 50% of playcounts) ([PDF](http://www.mtg.upf.edu/static/media/PhD_ocelma.pdf)). **Not a CMA number** — CMA's own: top-0.4% → 63–65% of streams 2014–2020 ([final report](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1120610/Music_and_streaming_final_report.pdf)). See D2.                                                                                    | sim1 validation T3 (world Gini validates at 0.97 by streams)        |
| 4  | Plays per listener per artist               | median 5.16, mean 21.21                                                                   | Schedl & Hauger, *Int. J. Multimedia Information Retrieval* 6:71–84, 2017, Table 6 (LFM-1b: 120,322 users, 1.09B events) ([PDF](https://link.springer.com/content/pdf/10.1007%2Fs13735-017-0118-y.pdf)); dataset intro: Schedl, ICMR 2016 ([PDF](https://www.cp.jku.at/people/schedl/Research/Publications/pdf/schedl_icmr_2016.pdf)). Measures panel lifetime, not a year → red team widened to 8–21 (CRITIC §2).                                                                                                                                      | every MVA number (PL=21.21 in all three sim1 scripts); fig1/5/7     |
| 5  | Pro-rata independent payout                 | $4.43 per 1,000 streams (US, Jan 2026)                                                    | Cited in PAPER §3 as Duetti; **not in vault** — no Duetti note exists, no public link verified here. See D3.                                                                                                                                                                                                                                                                                                                                                                                                                                            | anchor of the independent pool; fig1/5/7, breakeven                 |
| 6  | Signed artist per-stream take               | $0.0003/stream                                                                            | **Not in vault** as a direct figure. Nearest independent points in vault: E\&Y/SNEP 2015 waterfall via Techdirt ([link](https://www.techdirt.com/2015/02/05/yes-major-record-labels-are-keeping-nearly-all-money-they-get-spotify-rather-than-giving-it-to-artists/)); AEPO-ARTIS ≈£0.00065/stream (Sony data in CMA); CMA ≈£0.001/stream. See D3.                                                                                                                                                                                                      | anchor of the signed pocket; 188,590; fig1/5/7                      |
| 7  | Label pass-through                          | 6.772% — **derived** as 0.0003/0.00443, not an input (PAPER CHANGELOG v0.5.1)             | Vault-independent corroboration of the order: 8–20% pass-through (Rose, *Streaming in the Dark*, Berkeley J. Ent. & Sports Law 13:1, May 2024, [PDF](https://publicknowledge.org/wp-content/uploads/2024/05/Streaming-in-the-Dark_Competitive-Dysfunction-Within-the-Music-Streaming-Ecosystem_Berkeley-Journal-of-Entertainment-Sports-Law_May-2024.pdf)); \~10.6% reaches recording artists (CMU via AEPO-ARTIS). The "\~6.8% (CNM)" corroboration named in code comments is **not in vault**. See D4.                                                | v05 matrix signed column; ×14.8 contract multiplier; fig7           |
| 8  | Superfan share of audience                  | 0.6–1.7% (sensitivity range; 1.7% base)                                                   | SoundCloud × MIDiA, *Building a fan economy with Fan-Powered Royalties*, July 2022 (118,000 artists: avg 1.5% of interactions → 29% of income; "typically 1–2%"; FPR winners 1.9% → 42%) ([PDF](https://vi.be/files/2022-08/soundcloud-x-midia-building-a-fan-economy-with-fan-powered-royalties.pdf)). Lower bound 0.6% from the 97-2-1 participation rule (CRITIC §7), not from the white paper. See D5.                                                                                                                                              | breakeven range (18 combinations); fig1, fig6                       |
| 9  | Superfan revenue share (retraction input)   | 29% of artist income                                                                      | Same SoundCloud × MIDiA white paper — share of *royalties* under FPR, not donations; this category error is why "0.42 donations/yr" was retracted (CRITIC §1)                                                                                                                                                                                                                                                                                                                                                                                           | retracted P6 inversion; Retracted & bounded                         |
| 10 | Donation check                              | $3.1–6.9 (range; lognormal $5 base)                                                       | Range anchor: Waskow, Markett, Montag et al., *Pay What You Want! A Pilot Study on Neural Correlates of Voluntary Payments for Music*, Frontiers in Psychology 7:1023, 2016 — corrected mean PWYW €3.10, refusal 24.4% vs 17.3% (128/525 vs 91/525), N=25, real Bandcamp albums ([link](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4933710/)). The lognormal($5, σ=0.8) *shape* is an **assumption, not measured** (RESULTS, honest limitations).                                                                                                     | direct-economy check; fig1/5/6/7, breakeven                         |
| 11 | Subscription price × pool share             | $11.99/mo × 70% to the rights pool                                                        | 70%: "around 70%" per Bergantiños & Moreno-Ternero 2023 (citing Meyn et al. 2023) ([PDF](https://arxiv.org/pdf/2310.11861)). $11.99 (US price) **not in vault** (vault holds UK £ tariffs from the CMA report). See D6.                                                                                                                                                                                                                                                                                                                                 | user-centric wallet model; fig1/5/7                                 |
| 12 | Twitch paying-fan cadence, sub price, split | k=12+ structural; $5/mo; 50/50 split                                                      | $5 monthly sub: Sjöblom & Hamari, *Computers in Human Behavior* 2017 ([PDF](https://www.utupub.fi/bitstream/handle/10024/171134/Why%20do%20people%20watch%20others%20play%20video%20games.pdf?sequence=1)); 50/50 split: only as the *title* of ref. 44 (Clancy 2022 letter) in the vault — percentages not in any note body. See D7.                                                                                                                                                                                                                   | Twitch mechanics row (MVA 2,353); fig5/7; RESULTS §2                |
| 13 | Twitch income inequality                    | α=−2.13, Gini 0.57 (top-10k), ≈0.93 platform-wide                                         | Houssard, Pilati, Tartari, Sacco & Gallotti, *Monetization in online streaming platforms*, Scientific Reports 13:1103 (2023) ([PDF](https://www.nature.com/articles/s41598-022-26727-5.pdf))                                                                                                                                                                                                                                                                                                                                                            | RESULTS §2 benchmark context                                        |
| 14 | Twitch motivation model                     | explains 3.7% of subscription variance; N=1,097                                           | Sjöblom & Hamari 2017 (same paper as #12)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               | RESULTS §2 (why cadence, not motivation, is the dial)               |
| 15 | Tencent gifting multiplier                  | ARPPU ¥175.1 vs ¥8.5 = 20.6× (4Q21); 6.4× after squeeze (4Q24); segment −66.3% over 3 yrs | TME 4Q/FY2021 results ([link](https://ir.tencentmusic.com/2022-03-21-Tencent-Music-Entertainment-Group-Announces-Fourth-Quarter-and-Full-Year-2021-Unaudited-Financial-Results)); TME 4Q/FY2024 results ([link](https://www.prnewswire.com/news-releases/tencent-music-entertainment-group-announces-fourth-quarter-and-full-year-2024-unaudited-financial-results-302404220.html))                                                                                                                                                                     | RESULTS §2 benchmark; PAPER §4 (regulatory risk in §9)              |
| 16 | Patreon cadence and scale                   | k=12 (monthly billing arithmetic); \~25M paid vs \~100M free (Dec 2025)                   | Water & Music, *Why superfan subscriptions are dying out*, Dec 2025 ([link](https://newsletter.waterandmusic.com/archive/why-superfan-subscriptions-are-dying-out/)). k=12 is billing arithmetic, not a vault number.                                                                                                                                                                                                                                                                                                                                   | RESULTS §2 benchmark                                                |
| 17 | Telegram Stars retention and payout         | desktop 96.5% / mobile 67.5%; min withdrawal 1,000 Stars (\~$13); 21-day hold             | Withdrawal minimum 1,000 Stars: **Telegram primary** (`stars_revenue_withdrawal_min` in [api/config](https://core.telegram.org/api/config), [api/stars](https://core.telegram.org/api/stars)); retention and $-conversion: Tribute vendor guide 2026 ([link](https://tribute.top/blog/telegram-stars-creators-guide)) — vault holds ranges 95–97% / 65–70%; the point values 96.5/67.5 are midpoints. See D8.                                                                                                                                           | fig4 rails; $13 threshold logistics (Finding 4)                     |
| 18 | TON transaction fee                         | \~$0.0005 used in the model                                                               | Vault holds TON-denominated figures: \~0.00039 TON (ton.org marketing, [link](https://ton.org/en/100000-transactions-per-second-ton-sets-the-world-record-on-its-first-performance-test)) and 0.000540370 TON measured ([TON docs benchmark](https://docs.ton.org/contracts/standard/wallets/performance)). The **$ figure is a conversion, not a vault number** — see D9.                                                                                                                                                                              | fig4 rails (0.1¢ of $1)                                             |
| 19 | Hamster Kombat collapse                     | ×25 in 6 months (300M → 12M), calibration anchor                                          | 300M+ users and post-airdrop decline to 12M MAU: AInvest ([link](https://www.ainvest.com/news/telegram-play-earn-ecosystem-high-growth-investment-opportunity-2025-2512/)), CryptoPotato Sep 2024 ([link](https://cryptopotato.com/hamster-kombat-announces-details-of-airdrop-2-3m-accounts-banned-for-cheating/)). The former "Caladan, Apr 2026" attribution is **not in vault** and has been removed from captions/SPEC (2026-08-19). See D10.                                                                                                      | sim3 calibration T1; fig11–13                                       |
| 20 | Streaming fraud scale (context)             | up to 85% of fully-AI-track streams fraudulent (2025); \~$2B/yr industry losses           | **Two sources, not one:** 85% — Deezer Newsroom, 29 Jan 2026 ([link](https://newsroom-deezer.com/2026/01/ai-generated-music-deezer-selling-detection-tool/), [Jul 2026](https://newsroom-deezer.com/2026/07/ai-music-exceeds-50-percent-daily-uploads-deezer/)); $2B/yr — Beatdapp via MBW, 9 Jul 2024 ([link](https://www.musicbusinessworldwide.com/streaming-fraud-costs-the-global-music-industry-2bn-a-year-according-to-beatdapp-now-its-partnering-with-beatport-to-combat-the-trend/)). See D11.                                                | motivation for fig3 (the model's dilution curve itself is analytic) |
| 21 | Payout-rule core theorem                    | any stable rule divides a listener's fee only among artists that listener streamed        | Bergantiños & Moreno-Ternero, *Revenue sharing at music streaming platforms*, arXiv:2310.11861 \[econ.TH], Oct 2023 ([PDF](https://arxiv.org/pdf/2310.11861)) — full text in vault                                                                                                                                                                                                                                                                                                                                                                      | PAPER §1 (World-A ceiling argument); README Related work            |
| 22 | Bottom-tier market share                    | bottom two popularity categories ≈ 2% of market (2.2B remunerated streams)                | Frederik Juul Jensen, PhD dissertation *Alternative Payment Systems on Music Streaming Platforms*, Université Sorbonne Paris Nord, defended 5 Sep 2025 (Deezer data: 160,747 users, 2018–2020, 3.3B → 2.2B streams) ([PDF](https://www.musikindustrin.se/wp/wp-content/uploads/2025/09/Frederik-Juul-Jensen-PhD-dissertation-Alternative-Payment-Systems-on-Music-Streaming-Platforms-2025.pdf))                                                                                                                                                        | sim1 validation ("bottom 90% hold 0.9%" cross-check)                |

### Discrepancies (vault vs repo attribution) — kept visible on purpose

The table above records what the vault actually verifies. Where the repo's running text attributes differently, the difference is listed here rather than silently harmonized — the same policy as *Retracted & bounded*.

* **D1 — "87% of artists (Luminate)".** The vault's Luminate 2023 number is 86.2% of **tracks** ≤1,000 plays, not artists. The artist-side figures in the vault are Chartmetric's (86% under 10 monthly listeners). Running labels now say tracks (2026-08-19); T1 remains a stylized *artist-world* share calibrated to that track statistic.
* **D2 — "CMA/Last.fm" for top-0.28% ≈ 50%.** The 0.28%/50% pair and Gini 0.72 are Celma's 2007 Last.fm crawl. CMA's own measurement is top-0.4% → 63–65% (2014–2020). The repo's combined label "CMA/Last.fm" should read "Last.fm (Celma 2008); cf. CMA".
* **D3 — the two per-stream anchors are not in the vault.** Neither Duetti $4.43/1,000 (independent) nor $0.0003/stream (signed pocket) has a vault note. They are calibration anchors / assumed external rates, not measured quantities (PAPER §3, CHANGELOG v0.5.1 precision caveat); nearest independent vault points: AEPO-ARTIS ≈£0.00065/stream, CMA ≈£0.001/stream, E\&Y/SNEP 2015 waterfall. L1–L4 do not depend on the dollar level; every monetary figure, $d^\*$ and MVA does.
* **D4 — "CNM \~6.8%".** No CNM note in the vault carries 6.8%. The vault's pass-through estimates span 8–20% (Rose 2024) and \~10.6% (CMU/AEPO-ARTIS). In the code the pass-through is *derived* from the two anchors (6.772%), so the model does not depend on the CNM figure; the corroboration claim does.
* **D5 — superfan lower bound 0.6%.** The SoundCloud white paper supports 1.5% average (→29% of income), 1–2% typical, 1.9% winners; 0.6% comes from the 97-2-1 participation rule (CRITIC §7), not from SoundCloud.
* **D6 — $11.99.** The US subscription price is not vault-sourced (vault holds UK £ tariffs via CMA); 70% pool share is sourced (B\&MT 2023).
* **D7 — Twitch 50/50.** Only the title of Clancy's 2022 letter on revenue shares exists in the vault's reference list; the 50/50 number itself is not in any note body.
* **D8 — Stars 96.5/67.5.** Vault (vendor guide) gives ranges 95–97% and 65–70%; the model uses the midpoints. The 1,000-Star withdrawal minimum is confirmed by Telegram's own API config; the \~$13 conversion uses the same guide's $0.013/Star (the guide is internally inconsistent about $/Star at one point).
* **D9 — TON fee.** Vault numbers are 0.00039–0.00054 **TON** per transfer; the model's \~$0.0005 is a dollar conversion not present in the vault (at TON $3–5 the measured fee is \~$0.0016–0.0027).
* **D10 — "Caladan, Apr 2026".** The Hamster 300M→12M collapse is vault-confirmed (AInvest, CryptoPotato, Decrypt), but no vault note names Caladan. Captions and SPEC now cite AInvest / CryptoPotato (2026-08-19); the vault wording is MAU decline; the SPEC's "DAU, 6-month window" framing is coarser than the sources.
* **D11 — "fraud up to 85%, $2B/yr (Beatdapp)".** Two different sources: the 85% share of AI-track streams is Deezer's; the $2B/yr industry loss is Beatdapp's. They should not be cited as one.

*Verified from the tonify-research vault (1,152 notes) on 2026-08-09; 132 search/note-show queries, nothing cited from memory.*


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://tonify-1.gitbook.io/tonify-docs/reference/sources.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
