designfuchen/word-atlas · designfuchen

7,298 illustrations and over 20,000 audio clips, all given away free: how an IELTS vocab site earned 4,000 stars in 12 days

What it is A free, open-source IELTS vocabulary website. Each of the 7,298 words comes with an AI-generated illustration, British and American pronunciation, and read-aloud example sentences. No sign-up, and the assets can be downloaded and reused directly.

Why it took off The author is a Chinese AI content creator who self-reported 598k views on the first post, and big accounts in the same circle kept reposting it. The sheer workload (thousands of images, twenty thousand audio clips, all free) and a polished interface gave people reasons to share.

BHow you could ride this Don't build another IELTS version. Target a different audience: junior and senior high school, CET-4/6, Japanese, or HSK for foreigners. Use its open-source generation scripts and CC BY assets directly, launch on your own account, and line up a few creators in advance to repost on day one.

Grade B: the recipe is clear and both the assets and the scripts are reusable, and there's room in other audiences. But most of its stars come from the author's built-in 600k exposure and mutual-promotion circle. Followers without that account will struggle to replicate it, and the novelty of "the first open-source illustrated vocab book with audio" has already been used up. What the small circle is saying

Data (updated daily) · Repo created 2026-10-25 · 4,391 stars · 7 days +845: GitHub · stars
Data 2026-10-26 → 2026-11-08 (14 days, 14 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

Word Atlas (词光) is a free, open-source IELTS vocab website that needs no sign-up. Each word gets an AI illustration, British and American audio, and read-aloud example sentences. "One word, one picture" has been Baicizhan's approach for over a decade, so it's not new. What's new is one person using AI to produce the whole set of images and audio, open-sourcing the assets too, and handing it to a Chinese AI creator who already had hundreds of thousands of impressions to post it. The recipe can be copied, but the hardest part to copy is that account.

One post, 2,600 stars in two days

Oct 25, 2026, a little after 8 p.m. Beijing time, Chinese AI creator @yangleiocx3 ("Alei") posted on X with a promo video:

"I built what I believe is the best IELTS vocabulary website in the world. Completely free, completely open source."

十几分钟后,同一篇介绍发到了 V2EX 的「分享创造」节点。GitHub 仓库是当天下午才建的。按他 10-31 在置顶帖里自己报的数,这条首帖有 59.8 万浏览、6,500 人收藏(作者自报,未核实)。

The star curve that followed (GitHub stars, daily snapshots):

Date	stars	fork	What happened that day
10-25	Repo created		First X post, first V2EX post; @FiniYang quote-reposted it the same day
10-26	189	54	AI account 木马人 reposted: "With this tool, IELTS cram schools are going out of business~"
10-27	1,505	383	+1,316 in a single day, the peak; @geekbb reposted it, 268 likes
10-28	2,771	627	+1,266; Chinese reprint sites began mass-republishing it
10-29 ~ 11-06	3,078 → 4,244	681 → 908	+65 to +300 per day, with GitHubDaily and multilingual reprints relaying it

The fork-to-star ratio is about 21%, roughly double what's typical for new repos of this size. People who fork aren't just bookmarking. They're taking the assets and using them.

What it actually is

The README's opening line reads "Turn words into pictures, so review has a trail to follow":

7,298 IELTS entries, each with a stylistically consistent illustration, 22,905 British and American pronunciation and example-sentence recordings, and 14,596 example sentences;
Built-in FSRS spaced repetition that tells you each day what to learn and what to review;
A WebGL "vocabulary star map" with 19 topics, where learned words light up. Intro voiceover, ambient sound, interaction sound effects, and a 48-second promo video;
No account. Progress is stored in your own browser. Code is MIT-licensed; images, audio, and the word lists are open under CC BY 4.0, and the data/ directory can be downloaded directly.

Images and text were generated in bulk through an OpenAI-compatible API, voice through ElevenLabs, and the generation scripts are open-sourced in the repo too. In a V2EX reply, the author said: "The high-quality voiceover and images alone cost quite a bit."

How the stars came

The first push came from the author's own account. @yangleiocx3's bio reads "Mac + AI tool reviews | how to use them in China, save money, and get around restrictions | Build in public". He registered in June 2026 and has already posted 7,990 times. This is a Chinese AI content account with a steady posting rhythm and a mutual-promotion circle, not a cold start. By contrast, his GitHub account dates back to 2016 and belongs to an older iOS developer. His earlier repos mostly had 0 stars.

The second push came from big accounts in the same circle reposting. On 10-25 itself, OpenPencil's author @FiniYang quote-reposted it: "The voiceover alone is worth nearly ten thousand yuan! Bookmark it and you're ahead." Alei replied: "Thanks, boss, for quoting me." Over the next two days, 木马人, geekbb, and a string of Chinese AI accounts kept reposting. 10-27 and 10-28 each added roughly 1,300 stars over those two days, which is basically this wave.

The third layer was content farms. From 10-26 to 11-02, more than 15 Chinese reprints turned up in search: headlines like "GitHub's IELTS vocab tool that got 2000 stars in 2 days", plus Docker deployment tutorials, Sina Kandian, Taiwan's HSIAO NEWS, and a Brazilian Portuguese site. Videos were reposted on Bilibili, YouTube Shorts, and Instagram too. This layer carried it beyond the X circle and sustained the long tail of one to two hundred stars a day.

Fourth, the author kept the heat going himself. On 10-31 he posted "This time I actually pulled it off" with the data, saying people had made Korean, Russian, and Turkish versions. On 11-03 he posted "Open source can really make money, so many people are willing to pay voluntarily" with a donation screenshot (amount unverified). On the V2EX thread, he took the suggestions there (a font issue where e looked like c, hiding the Chinese so you guess the word first, dictation not being replayable). He replied "Changing it right now" that day and "Bros, I fixed every issue you raised" the next.

Channels that didn't work: no Hacker News, Product Hunt, or Reddit traction found. The repo was only 12 days old, too new to enter GitHub Trending. Whether anyone seeded it on Xiaohongshu or Douyin wasn't found.

Why this one got shared

The workload is visible. Almost every reposting caption repeats the numbers: 7,298 images, over twenty thousand audio clips, "the voiceover alone is worth nearly ten thousand yuan." AI let one person produce a volume of content that "looks like it needs a team," and people starred it partly to acknowledge that effort.
It looks good and is easy to share. The star map, the background music, and the promo video all make a screenshot worth posting. V2EX comments: "The aesthetics show real care", "UI and UX are both great, very comfortable."
It taps into frustration with domestic vocab apps. Products like Baicizhan and Bu Bei Dan Ci have memberships, ads, and required sign-ups, so "completely free, no sign-up, open it and start memorizing" is the selling point in itself. The most common commenters are parents: "Great stuff, I'll get my kid using it", "Please make junior and senior high versions, I'll happily send a WeChat red packet."
Forking it gives you something to work with. Some people deployed it to their own servers the same day, "had Gemini modify it to enable automatic sync". On Juejin, two posts already exist: "I re-skinned word-atlas and put it live" (10-26 and 10-31).

Who else is doing the same thing

Project	Scale	Compared with Word Atlas
Baicizhan	1.74M rating on the China App Store	The originator of "one word, one picture". Closed-source commercial product, requires an account, has memberships and ads
MoMo Dan Ci, Bu Bei Dan Ci	990k / 460k ratings	Focused on algorithms or real context, no illustrations, paid
qwerty-learner	23.7k stars	Open source, typing-based word learning, no images or recordings
earthworm	11.2k stars	Open source, word-in-sentence building, not a vocabulary illustrated guide
DictionaryByGPT4	6.4k stars	Also an AI-generated vocab book, but text only, no app
Anki + FSRS	32.6k stars	General review tool, you have to find your own decks

(Star counts are GitHub data as of 11-06; ratings are from the China App Store)

On the open-source side, no project before it combined a full set of illustrations with human-quality audio. Word Atlas now owns that gap.

Competitors are already following: 11-06, an iOS app called "Vocabry (词绘)" is recruiting TestFlight beta testers on V2EX. It has 7,644 words, an illustration for each, FSRS, and a star map where "every word you master lights up a star", nearly identical to Word Atlas, with added graded reading, original books, listening, and in-app Pro purchases. There's also a Persian fork and an Android offline fork. The author is building an iOS app too, and the README's opening section already funnels users there.

If it were you

Bottom line: The recipe is genuinely copyable: use AI to batch-produce a full set of multimedia learning content, open-source all the assets, package it as a polished piece of work, and quantify the labor in the copy. But most of Word Atlas's 4,000 stars came from the author's 600k-exposure account and his mutual-promotion circle. Without that distribution, the same thing would likely have stalled at 44 replies on V2EX. And the novelty of "the first open-source illustrated vocab book with audio" has already been spent. A second project in the same category won't get the "best in the world" kind of surprise.

Who it's for: Developers who already have a few readers on X, Jike, or Xiaohongshu; people who know a particular exam or language well enough to judge whether AI-generated images and example sentences are accurate; and people willing to front a few thousand yuan in generation costs.

Not for: people with no channel of their own who are counting on GitHub organic traffic; anyone trying to build "another IELTS version" (that spot is already taken); anyone hoping to make money from this. Open-source projects like this bring stars and followers, and sponsorships are not income.

How to do it:

Change the audience, not the skin. Commenters have already listed the requests: junior high, senior high and primary school textbook versions; CET-4/6, postgrad entrance, TOEFL; Japanese JLPT N1 to N5, Korean TOPIK for Chinese users; or the reverse, an HSK version for foreigners learning Chinese, which is an English-GitHub market where word-atlas doesn't reach.
Build directly on its shoulders: the generation scripts and the FSRS setup already exist, and the assets are CC BY 4.0, so you can use them with attribution. Put your effort into new word-library images and audio, not a frontend rewrite.
Fill the gaps it leaves: people want etymology and a root-word index that V2EX lacks, phrase pronunciation drills, and instant lookup of words inside sentences. Make "what word-atlas doesn't have" your hook.
Copy its README opening: a cover image, one line, a row of numbers (how many words, images, audio clips), then a 30 to 60 second video.

How to get users: launch from your own account, and the same day post on V2EX under "Share & Create"; line up two or three bloggers in the same circle ahead of time and have them quote the launch on day one. If it's an exam tool, post a "free open-source word memorization" video on Xiaohongshu and Bilibili, where parents and students are. For HSK or foreign-learner audiences, go to Reddit's language-learning subreddits and r/coolgithubprojects. Reposting sites will come to you.

When to move: commercial follow-ups like Word Painter already exist, and the author is building an app too. The novelty of "AI builds a full study kit, then open-sources it" probably has a month or two left (inferred). Ship before year end; whoever goes first gets the "first X version" headline.

Signs you've misread it: if a week after launch your own post has under 10,000 views and the repo is at three-digit stars, the hook isn't "open-source word memorization", it's the account. And if word-atlas gains only single digits a day after mid-November, it means this kind of project's stars came mostly from one wave of reposts, and the long tail is thinner than it looks.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report

Competitors

Name	Where	Launched	Scale	How it differs from this one
Baicizhan	iOS	2012-09	1.74M rating on the China App Store	Also image-based word memorization, but closed-source commercial, needs an account, has memberships and ads. word-atlas is free, open-source, no sign-up, with a consistent illustration style and a star map.
Bu Dan Danci (Don't Memorize Words)	iOS	2014-02	China store 460k ratings	Built on real context and video example sentences; no illustrations; commercial subscription
Mooncake (Momo) Vocabulary	iOS	2014-07	China store 990k ratings	Focuses on memory algorithms, mostly plain text, paid
qwerty-learner	Open source	2021-01	23.7k stars	Type-to-memorize; no illustrations or audio assets
earthworm	Open source	2024-01	11.2k stars	Learn English by making sentences from linked words; not a vocabulary atlas
DictionaryByGPT4	Open source	2024-06	6.4k stars	Also an AI-generated word book, but text only: no app, images or audio
Anki + FSRS / IELTS decks	Open source	2012 / 2022	Anki 32.6k stars	General-purpose tool; you have to find decks yourself. word-atlas works out of the box with images and audio
kajweb/dict, hefengxian/my-ielts	Open source	2019 / 2023	3.8k / 4.1k stars	Just word lists or materials, not apps

Where users come from

Verified Author's own launch on X: @yangleiocx3 10-25 first post "I built what I think is the best IELTS word-memorization site in the world", 10-31 self-reported 598k views and 6,500 bookmarks; the account was registered 2026-06 and has posted 7,990 times. It's a Chinese AI self-media account.
Verified Reposts from big accounts in the same circle on X: FiniYang quote-tweeted 10-25 (the author replied "thanks boss"), 木马人 10-26, geekbb 10-27 (268 likes), Gorden_Sun 10-29, GitHubDaily about 11-02
Verified V2EX Share & Create: v2ex.com/t/1251691,10-25 12:58 UTC, 44 replies
Verified Reposts by Chinese content farms / aggregator sites: Sanse Source Code, Zupu, Gelv Poetry, Feng Dao, Lao Su, Muyan, Youtu, APP Miao, Sina Kandian, AGI Hunt, HSIAO NEWS, novidadesia, and 15+ others (10-26 to 11-02), plus re-uploads on Bilibili, YouTube Shorts, IG, and FB.
Verified Author keeps posting build-in-public updates: 10-31 data-update posts (71K views), 11-03 tip-jar posts
Inferred Xiaohongshu / Douyin: nothing found

Why 词光 · Word Atlas (designfuchen/word-atlas) took off

Research date 2026-11-06. Data from our github database (repo_day, daily since 2026-10-26), the GitHub API, X post pages (read via Jina), the V2EX API, and Exa/Jina search.

One line

A Chinese AI blogger (X @yangleiocx3, "Ali") paid for generative AI and ElevenLabs to give each of 7,298 IELTS words an illustration, British and American pronunciation, and read-aloud example sentences, then built it as a visually polished, no-sign-up PWA. The code (MIT) and the full asset set (CC BY 4.0) were open-sourced together. It launched on X's Chinese AI circle and V2EX, then was amplified by reposts from big accounts in his circle and dozens of Chinese content-farm reposts. 12 days, 4,244 stars, 908 forks.

1. Data: what the curve looks like

Date (our database snapshot)	star	fork	Daily change
10-25	Repo created (UTC 08:27)
10-26	189	54
10-27	1,505	383	+1,316 (peak)
10-28	2,771	627	+1,266
10-29	3,078	681	+307
10-30 ~ 11-06	3,218 → 4,244	705 → 908	+65 to +200 per day

Source: GitHub database repo_day; live GitHub API values 11-06 4,310 stars / 921 forks.

Fork/star ratio is about 21%, roughly double the median for new repos of similar size (about 11%). That fits its use case, people taking the assets and reskinning them: the word library and media are in the repo, so you can fork and deploy. Our suspect flag is empty for this repo, so it doesn't look like star-botting.
The repo is only 12 days old, short of the roughly 60-day threshold for GitHub trending, so trending was not its source.

2. Where users come from (growth tactics)

① 作者自己在 X 发首帖,当天又发 V2EX。 首帖在 2026-10-25 约 12:45 UTC(按视频 ID 推出来的时间),文案是「我做了一个我自认为全世界最好的雅思背词网站,完全免费,完全开源。🔥🔥🔥🔥🔥」,配一段宣传视频。作者 10-31 在置顶帖里自己报数:59.8 万浏览、6,500 人收藏,GitHub 3,600+ 星。V2EX 同一篇文章 10-25 12:58 UTC 发在「分享创造」节点(https://www.v2ex.com/t/1251691),44 条回复,有人说「刚在 x 上看到,这里又看到」。

Author account @yangleiocx3: by ID it was registered in 2026-06, and has 7,990 posts so far. The bio reads "Mac + AI tool testing | how to use them in China, save money, unblock | Build in public", a typical Chinese AI self-media account that already had a posting machine and an engaged circle. The GitHub account is a 2016-year-old iOS developer account (Boxin Yang), and its previous repos mostly sat at 0 stars.

② Reposts from big accounts in the same circle. These were the main push behind the roughly 1,300 stars per day gained over 10-26 to 10-28:

@FiniYang (OpenPencil author, "Agent architect") 10-25 quote-tweeted the same day: "The voiceover alone is worth nearly 10,000 yuan! Bookmark it and you're ahead." The author replied "thanks boss for quoting me". The author also has a codex-openpencil repo; the two are in the same circle (https://x.com/FiniYang/status/2114392284305502615).
@cnyzgkc 木马人 (AI self-media) 10-26: "With this tool, IELTS cram schools are going to go under~" (https://x.com/cnyzgkc/status/2114688809107644737)
@geekbb 10-27:268 likes / 46 reposts (https://xx2.cn/post.php?id=2114979081158844600)
@knowledgefxg, @weiwei2018831, @BeardedPretty, @fhwofjow51260 and other Chinese AI accounts reposted in 10-25 to 10-27.
Long-tail follow-ups: @Gorden_Sun 10-29 (2,645 views), @GitHub_Daily about 11-02.

③ Chinese content farms and auto-aggregators reposted it in bulk (10-26 to 11-02). Over 15 articles are traceable: 三色源码 "IELTS word app, 2 days to 2000 GitHub stars", 祖普科技, 格律诗, 冯导 AI Tools Research Institute, 老苏's blog (Docker deployment tutorial), 沐颜小栈, 友兔, APP喵, Baffox, Sina Kandian, AGI Hunt (Chinese and English versions), HSIAO NEWS (Taiwan), novidadesia (Brazilian Portuguese), buildrun.io, frontiernow, thedailycommit and more. There were also Instagram, Facebook, YouTube Shorts, and Bilibili re-uploads titled "Productivity nerds, rejoice! This GitHub word-memorization tool". This layer carried it out of the X circle and drove the long tail.

④ The author kept posting to sustain the heat. 10-31 posted "This time I really pulled it off", sharing data and saying people had translated it into Korean, Russian and Turkish (71K views). 11-03 posted "Open-source projects can really make money, so many people are willing to pay voluntarily" with tip-jar screenshots. The README and the App landing page carried WeChat/Alipay tip codes, plus a WeChat contact for "ads, investment, partnerships".

Channels with no evidence found: paid ads, HN, Product Hunt, Reddit and TikTok turned up nothing. No direct Xiaohongshu posts were found (TikHub search returned nothing).

3. Why it took off (the causes)

Visible workload is itself the distribution hook. 7,298 illustrations, 22,905 audio clips, 14,596 example sentences. The reposts kept returning to "just finishing the images is a huge job" and "the voiceover alone is worth nearly 10,000 yuan". AI let one person ship what looks like a team's output, and viewers starred it for that sense of effort.
Looks great online. A star map (WebGL2 word starfield), a voiced splash screen, 6 ambient-sound tracks, interaction sound effects, a 48-second promo film: screenshots and video that travel well on X. V2EX comments: "the aesthetics are well done", "both UI and UX are great".
Free, open-source, no sign-up, hitting the backlash against domestic word apps with subscriptions and ads. Chinese users already resent the memberships and ads in products like Baicizhan and Bu Dan Danci, so "completely free, no sign-up, open and study" is a selling point in itself.
The assets are open (CC BY 4.0), so forkers have something to take. Others can reskin it straight away, build Anki decks, or ship other-language versions. Fork rate is 21%, and on Juejin there are already two posts from 10-26 and 10-31 titled "I reskinned word-atlas and shipped it" (https://juejin.cn/post/7700855774545526838 and juejin.cn/post/7702863198185586739).
The author brings his own distribution. A frequently posting Chinese AI self-media account, plus a circle of "boss" FiniYang and others who repost each other. A first post at 598k views is not what organic cold-start gets you.
IELTS has rigid demand, and the audience is concentrated on X and V2EX among Chinese users. Students abroad, programmers and parents all show up in the comments ("for my kid to use", "please make junior and senior high versions").

4. Competitors

Product	Form	Launched	Scale	How word-atlas differs
Baicizhan	iOS/Android	2012-09	1.74M rating on the China App Store	The originator of "image-based word memorization" and a direct competitor: commercial, closed-source, needs an account, has memberships and ads. word-atlas is the free, open-source take on the same idea, with a consistent AI illustration style plus a star map
Bu Dan Danci (Don't Memorize Words)	iOS/Android	2014-02	China store 460k ratings	Built on real context and video example sentences; no illustrations; commercial subscription
Mooncake (Momo) Vocabulary	iOS/Android	2014-07	China store 990k ratings	Focuses on memory algorithms, mostly plain text, paid
IELTS-specific apps such as Ielts Danci Zhan	iOS	From ¥2017	Thousands of ratings	Exam-oriented, commercialized
qwerty-learner	Open-source web app	2021-01	23.7k stars	Type-to-memorize; no illustrations or pronunciation assets
earthworm	Open-source web app	2024-01	11.2k stars	Learn English by making sentences from linked words; not a vocabulary atlas
DictionaryByGPT4	Open data	2024-06	6.4k stars	Same "AI-generated vocabulary book" idea, but text only, no app. It's the predecessor of Vocabulight.
hefengxian/my-ielts	open-source material	2023-07	4.1k stars	IELTS resource collection, not an app
kajweb/dict	open-source wordlist	2019-02	3.8k stars	raw word list, no images or audio
Anki with IELTS decks / FSRS	Open source	2012 / 2022	Anki 32.6k stars	general spaced repetition, you have to find your own decks, high barrier; Vocabulight has ts-fsrs built in, works out of the box

(star count from the github repo 2026-11-06; rating count from the iTunes Search API, China store)

Bottom line: "An AI picture for every word" is not a new idea; Baicizhan has been doing it for over a decade. What's new in Vocabulight: ① AI brings the cost of the whole set of images, audio and example sentences down to what one person can afford, and then all of it is open-sourced and free; ② the product is built as an "aesthetic piece" (star map, soundtrack), which makes it naturally shareable. On the open-source side, no project has both full illustrations and human-quality recordings. That gap is empty.

5. Differentiation, and whether that's why it broke out

The differentiator is "Baicizhan-level image-text-audio content" plus "open source, free, local-first, no sign-up" plus "looks good." That's why it took off, but it needed the author's own X distribution: the sense of labor and the aesthetic gave people a reason to repost, and the author's AI creator circle provided the first wave of 600k impressions. A good product without that account would likely have stalled at the few dozen replies on V2EX.

6. How many people can see it

It falls into "talked about in small circles, leaning toward widely discussed": the Chinese X AI circle, V2EX, Juejin and dozens of Chinese reprint sites have all written about it, there are English and Portuguese reprints, and GitHubDaily (a major Chinese GitHub account) also posted it.
No discussion found on mainstream English communities (HN, Reddit, Product Hunt). Revenue is not public; there are only screenshots of tips the author posted (amounts unverified).

7. Copycats

Vocabry (readori): 11-06 is recruiting TestFlight beta testers on V2EX (https://www.v2ex.com/t/1254964). 7,644 words, illustrations, FSRS, a star map that lights up a star for each learned word, ambient sound, typing practice: almost identical to Vocabulight, with reading, listening, shadowing and Pro in-app purchases added. This is a clear commercial follow-up.
Two re-skinned sites on Juejin; forks including a Persian version (appxa/word-atlas, 14 stars) and an offline Android version (H1DDENADM1N/word-atlas-android); the Korean, Russian and Turkish translations the author mentioned.
An App Store app named "词光-音乐歌词播放器" (Vocabulight Music Lyrics Player, id 6825042357) appeared; it's unrelated to this project and is probably riding the name. 11-02
The author is building an iOS app (TestFlight, requires iOS 26), and the README's first screen funnels people there.

8. Takeaways

"Large AI-batch-generated multimedia content" plus "fully open source" plus "aesthetic packaging" is a repeatable star recipe. Swap in TOEFL, CET-4/6, Japanese JLPT N-level words, high school entrance exam vocab, or classical poetry (someone on X already said they'd PR poetry in). The content generation scripts are open-sourced in scripts/atlas/ (OpenAI-compatible API for text and images, ElevenLabs for audio).
The copy should quantify the labor: "7,298 illustrations / 22,905 audio recordings / voiceover worth nearly ¥10,000."
Launch on X first (own account plus a mutual-promotion circle) on the same day as V2EX, then wait for content farms to auto-repost.
Monetization path: open-source site to drive traffic, then build an app, collect tips, and take on ad partnerships.

9. Still unclear

598k views and 6,500 bookmarks are self-reported by the author; the original first post couldn't be read directly.
Follower count and the view numbers for the reposts by FiniYang and others are incomplete.
Whether it was pushed on RedNote or Douyin was not checked.
No data on the site's real retention or DAU; tip amounts are unverified.
The GitHub stargazer timeline couldn't be pulled (API rate-limited), so star peaks can't be matched hour by hour to each repost; only a rough daily view is possible.
License shows NOASSERTION because the code is dual-licensed MIT and the content CC BY 4.0, not because it has no license.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.