Companion to the WorldLibs research report and the realistic dynamic worlds report
You arrive at a real-looking Earth, built from NASA satellite imagery, with day and night drawn where the sun was in the year on the slider. Give it a nudge and it unrolls into a flat atlas where every library sits as a tiny floor plan of its own room. Pick one and the camera flies there, the plan rises into walls, and your device builds the room while you watch.
A working mock of the atlas the globe unrolls into. Drag the year or press play, switch to Europe, and pick a library to see its card. MVP libraries are in green.
The earlier plan dropped the globe because the European cluster piles up on it and half the pins face away. We keep the fix and add the spectacle back: the globe is what you see first, and picking happens on the flat atlas it unrolls into. The sketch below is drawn from the same land data as the atlas, with the sun where it stood at noon over Panama.
Static sketch. The real globe uses NASA Blue Marble and Black Marble textures; the night side glows with city lights.
| Option | Good at | Breaks on | Verdict |
|---|---|---|---|
| Realistic globe that unrolls into the atlas | A real first impression, then place and time in one flat view for picking | Needs a poster frame and a lazy 3D chunk to stay fast | Pick |
| Flat atlas only | Instant paint, every pin in view, works as a plain list on phones | Little spectacle on landing | Folded in as the picking layer |
| Google 3D Tiles as the base | The most realistic Earth on the web | Costs $6 per 1,000 loads past 1,000 a month, Google's attribution rules, heavy on phones | No, free MapLibre view instead |
| Timeline only | The story of libraries rising and burning | Loses place, so you never see that the lost libraries ring the eastern Mediterranean | Folded in as the scrubber |
| 3D hub hall (a Borges-style hexagon with a door per library) | Immersion from the first frame | A heavy load before you have chosen anything | Later, as a Borges side-room |
Deep zoom uses OpenFreeMap or EOX tiles. Ancient sites take coordinates from Pleiades. Historical borders come from OpenHistoricalMap (CC0), not historical-basemaps, which is GPL.
The glyph is the library's footprint reduced to its type. The same typology decides how each room gets built: a hall is one bay module repeated down an axis, so the Long Room, Admont and Strahov share most of their geometry and differ in ceilings, wood and paint.
Five screens, one continuous camera. A poster frame paints first, then the globe's lazy 3D chunk. The fly-in and the extrude animation cover the download of the room.
01
Scrub through time, let the globe unroll, zoom to Europe, tap a glyph.
Poster under 60 KB first; NASA KTX2 globe, 3 to 4 MB; unrolls to Equal Earth for picking.
02
The glyph tilts into perspective and its walls extrude into the room shell.
Shell and far-end impostor first; a Worker builds bays from the seeded kit behind the door.
03
First-person walk at real scale, with a hotspot on each signature object.
Baked day and candle lightmaps, three-mesh-bvh capsule collision, joystick and gyro on mobile.
04
Readable spines carry a thin bright band. Tap one and it slides out into your hands.
InstancedMesh spines from a KTX2 atlas; the hero book is one real mesh.
05
The book opens to a two-page spread and turns page by page to the end.
StPageFlip on a CanvasTexture, text from Standard Ebooks or Gutenberg, scans through IIIF.
Rooms are not downloaded as finished models. Each one ships as a hall-kit file of a few KB plus a seed, and your device builds the shelves and books in a Worker while the walls rise. Light and fine detail are too slow to make on a phone, so the build server bakes them once and ships them as textures. Roblox splits work the same way: it streams what is near you, shows a baked stand-in (SLIM) for what is far, and can pause a player at the edge of the loaded area until the next part arrives.
Never downloaded and never cached. The same seed always rebuilds the same room.
KTX2 and Meshopt glTF with hashed names, 25 MB or less per room. The browser keeps the last 8 rooms.
Total Library, a Godot project, generates a whole library in real time from a 17 to 18 MB build, so the approach holds up in practice. Generation time on a mid-range Android is still unmeasured and is the first thing Phase A tests.
A faithful room is only half of it. People remember a place for what it held, who worked there and what happened to it. The Digital Library Cave for Dunhuang put 21 zoomable documents and real historical figures inside a 3D hall and logged 14 million visits in a week. Each WorldLibs room gets the same layers, told from sources.
| Layer | What you get | Example |
|---|---|---|
| Cases hold real items | Every display case holds a readable item or a catalogue card, with a source and licence line | Gulliver's Travels on a Long Room shelf; the Book of Kells as a catalogue card under glass |
| Short tours | Two or three five-minute tours per room, plus free roam | "Who built the gallery", "What the vault changed" |
| People as silhouettes | Figures with a footnote each. No invented faces | Philodemus at Herculaneum, Xuanzang at Nalanda, the Mamma Haidara family in Timbuktu |
| Phase slider | Switch the room between building phases | Long Room: the original flat ceiling against the barrel vault of 1859–61 |
| The place today | What is happening to the real building | The Old Library closes in 2027 for restoration to about 2030: "visit while it is closed" |
| Sound | Convolution reverb from an OpenAIR heritage space | No library hall is in that set, so it is labelled "proxy acoustic, not this room" |
| Sun-true windows | The time slider sets the sun angle at the library's real latitude | Scrub to a winter afternoon in 1750 and low light crosses the Long Room floor |
Following the London Charter and the Seville Principles, the evidence tag drives the renderer: invented elements draw dashed or ghosted, a "show evidence only" toggle hides them, and AI-assisted assets carry a badge with the model and date. No AI-made text, faces or inscriptions appear anywhere.
Every Joule runs each idea as a loop: predict it, test it, prove it, earn a stamp. We borrow that loop as one shared component and put it on a few books whose central idea you can try with your hands. Each page takes about 2 to 4 days to build.
| Book | Room | Predict | Reveal | Phase |
|---|---|---|---|---|
| Byrne's Euclid (1847) | Long Room | Which coloured shape comes next in the proof? | The proof unfolds step by step, then a slider sandbox | A |
| Plan of St Gall | St Gallen | Where do the scriptorium and infirmary sit? | The plan lifts into 3D. Cod. Sang. 1092 is CC BY-NC 4.0 on e-codices | B |
| Darwin's divergence diagram | A later room | Which lineage splits first? | The branching animates through time | B |
Gamification lifts engagement more than learning (Souropetsis et al. 2023), so stamps stay cosmetic. They live on a library card in localStorage under wl.stamps, with no streaks and no leaderboards. The card can be saved as a PNG to share.
The feel of a game, with no score. Each pattern below comes from a game that already proved it.
| From | What it proved | What we take |
|---|---|---|
| The Uncensored Library (Minecraft) | One conceit, real texts and a minimal UI beat a long feature list. About 300 books and 25M+ visitors reported | Walk in, pull a real book, read it |
| Gone Home | Objects carry the story better than markers do | Left-behind objects (marginalia, a ticket slip, a plan fragment) that show up when you walk close |
| Return of the Obra Dinn | A ledger the player fills in by deduction | Sort claims about Alexandria or Nalanda into attested, inferred or legend |
| Play the Past on ruins | Destruction is testimony, never a timer or a win state | A sensitivity pass on Alexandria, Timbuktu 2013 and Weimar 2004 |
| Assassin's Creed Discovery Tour | Short narrated tours with combat removed | Guided mode with a sourced script per stop |
Three modes ship in Phase A: a guided tour, free roam, and a reader list with no 3D at all. Reduced motion swaps the fly-in and page turns for cross-fades.
AI works at build time and only makes proposals. A person signs off, and everything ships as static files. Nothing generates a world while you wait.
| Tool | Output | Cost | Catch | Use |
|---|---|---|---|---|
| World Labs Marble | Gaussian splats plus collider meshes; renders in Spark | About $1.20 per world | Export and commercial rights need the Pro plan; terms not yet read | One backdrop, tagged "invented" |
| TRELLIS.2 | GLB with PBR materials from an image | Hours on a rented 24 GB GPU | MIT licence | Props from licensed photos |
| Meshy, Tripo, Rodin | Props by API | Under $50 for a few dozen (estimate) | Output licences conflict across sources | Fallback after a terms check |
| AnySplat, WildGaussians | Splats from a few casual photos | Research code | Unproven on sparse Commons interiors | Watch |
| Genie 3 | Streamed interactive worlds | Ultra subscription | US only, prototype, no export | None |
| HunyuanWorld, Hunyuan3D | Worlds and meshes | Self-hosted | Licence excludes the EU, UK and South Korea | None |
| In-browser LLM guide | A narrator that answers from the evidence rows | About 450 MB download | WebGPU only, battery, can still make things up | Later, opt-in |
The most useful AI job is reading sources. A quote-to-JSON step takes a cited quote and has an LLM propose hall-kit parameters. A schema check runs next, then a check that the quote really contains each number. A person signs off and a paradata record is written. Any element without a quote id is tagged "invented" on its own, and the JSON is committed with the code.
Drawn to one scale. Every room here takes under about a minute to walk at 1.4 m/s, which keeps each one small enough to model and bake by hand, and sets the camera's walking speed. Dashed edges mark a dimension taken from background knowledge.
The report grades every reconstruction as known, probable or speculative. The app shows that grade in the rendering itself, so nobody mistakes a guess about Alexandria for a survey of Ephesus. Every lost room gets a toggle between the fact layer and the guess.
Lost libraries need a sense of time as much as a sense of place. Each one gets four layers: a lifespan with honest, blurry ends, a then-and-now view, the evidence shading above, and shelves that show what survived and what is known to be gone. Tap a row to see its dated events.
Every book on a shelf gets a record, but only some can be read. The shelf shows which is which before you reach for it: readable spines carry the bright band, catalog books stay closed, and lost works appear as outlines with their title.
We store pointers, not copies. Our files hold the checked facts and, for every book, its ID at the source. The pages themselves come from the holder's public API. There is no database: everything we host is static files on a CDN, under 1 GB at full size (facts under 2 MB, a globe of 3 to 4 MB, 30 rooms at up to 25 MB each, a few hundred pre-built texts). Copying the scans is what would make it grow, since the First Folio alone is 920 pages, about 440 MB even at screen size, so we never copy them.
Each step loads only what that step needs, with file names that change when the content changes. The browser keeps files forever and asks again only for the small per-library manifests. A second visit to a room costs one tiny request.
| When | What loads | Budget | Cache rule |
|---|---|---|---|
| Open the site | One HTML file with the atlas, the land outline and all 30 library rows inlined, plus the globe poster frame. No API calls. | ≤ 150 KB compressed, poster < 60 KB | Service Worker, cache-first, hashed name |
| Globe goes live | The lazy three.js chunk, then the NASA day, night and cloud textures as KTX2. | About 170 KB gzip, then 3 to 4 MB | Cache-first, immutable |
| Hover or tap a glyph | That library's manifest: dates, events, sources and its book list with tiers and IDs. | 5 to 10 KB | Stale-while-revalidate, so it opens instantly and refreshes quietly |
| Enter the room | The hall-kit JSON and seed, the far-end impostor, then baked lightmaps, atlases and props. Geometry is built on the device and never downloaded. | A few KB, then 25 MB or less per room | Cache-first, immutable; last 8 rooms kept, geometry never cached |
| Look at a catalog book | Nothing new: its details are already in the manifest. One cover thumbnail. | About 15 KB | Cache-first |
| Open a readable text | The book's table of contents and its first chapter, then the next chapter in the background as you read. | 20 to 60 KB per chapter | Cache-first; whole book kept when saved offline |
| Open a scan | A slim page list from our CDN, then images straight from the holder's IIIF server at spread size (1024 px), the next two spreads prefetched, 256 px thumbnails for the page strip. | A few KB, then 270 to 480 KB per page (measured) | Runtime cache with a size cap, oldest pages evicted first |
| Come back later | Reading position, bookmarks and offline books live on the device. | 0 requests | IndexedDB, with persistent storage requested when you save a book |
Gutenberg is the one source a browser cannot read directly: its files send no CORS header, and it asks people not to fetch its site by script. So a build step pulls each text once from Standard Ebooks or a Gutenberg mirror, cuts it into chapter files of plain paragraphs and puts them on our CDN. Gilgamesh is 250 KB as plain text.
The text arrives as paragraphs, not pages. The reader paginates for the current screen and font size, so a phone and a desktop share the same files and changing the font never refetches anything.
Every book also opens in a plain reader without the 3D room: adjustable type, light, sepia and dark, and page turns that switch to a fade when reduced motion is on. The 3D page turn is the treat, the plain reader is the workhorse.
e-codices and the Bodleian both let any site load their IIIF images, so the browser asks them for exactly the size it draws. The only thing we pre-process is the manifest. The First Folio's is 1.2 MB, so a build step cuts it to a page list of a few KB. If a holder ever blocks browsers, we mirror the few pages a room shows, never the whole book.
Wikidata, Wikimedia Commons and Open Library all answer browsers directly, and we use them. They fill IDs, cover images and author dates at build time, where the checks compare them with our sources before anything ships. They don't serve facts live: anyone can edit Wikidata, and Gutendex returned 503 while this page was being checked.
Snappy means something is always on screen and every wait is covered by motion. Budgets: 5 MB before the first room, LCP under 3 s, p95 frame time of 20 ms or less. WebGL is the default; WebGPU ships only after an A/B test wins.
| Library | Field | Was | Now |
|---|
Each phase proves one thing before the next one spends money on it.
Phase A
Done when someone can scrub to 1733, fly to Dublin, walk the Long Room, pull Gulliver's Travels and read it to the last page.
Phase B
Done when a new hall room ships in under a week.
Phase C
Done when the atlas has no greyed-out glyph without a date for its room.
Seventeen similar projects were checked, along with Every Joule, Total Library and Roblox streaming. None combines real rooms, readable books and evidence tags, but each one settles a question we would otherwise test ourselves.
| Project | What it settles | What we do |
|---|---|---|
| Every Joule | One hard idea per page, which the reader predicts, tests and then proves | Living pages for Byrne's Euclid, the Plan of St Gall and Darwin |
| Roblox streaming, SLIM | A shipped engine pauses the player at the edge of what has loaded instead of showing holes | The door holds until the next bay exists, and the far end shows an impostor |
| Total Library (Godot) | A whole library can be generated in real time from a 17 to 18 MB build | Seeded hall kits built by a Worker |
| The Uncensored Library | Curation and one clear idea carry a project | Every room has a reason to exist, stated in one line |
| Assassin's Creed Discovery Tour | Short narrated tours work | Guided mode |
| Pergamon VR with AI (ISPRS 2025) | Academic VR with AI guides, but no evidence shown per element | Evidence tags on every element |
| Wonderful Ephesus | Celsus rebuilt from a laser scan; a scan is the strongest evidence a ruin has | Celsus from the excavation plan and licensed photos |
| MyShelves, 3D Bookshop | The take-a-book-and-read loop works in a browser, and Foliate JS reads Standard Ebooks EPUBs | Same loop, with pre-built text parts |
| Mirador, StPageFlip | Apache-2.0 and MIT, both maintained | Mirador for the scan panel, StPageFlip for page turns |
| IA BookReader, Turn.js | AGPL, and a licence that restricts commercial use | Copy the BookReader UX only; skip Turn.js |
| Trinity College Dublin | The official page already links to virtual tours and 3D floor plans | Link out and credit it, and claim no endorsement |
| World Digital Library | Retired in 2020–21; its address now redirects to a Library of Congress page | Keep every data file exportable |
| Library of Babel sites | The Borges name is crowded | A small Borges side-room, no Babel branding |
| Wikidata WikiProject Libraries | Library coordinates, labels and images are already modelled | Live labels and images by QID through SPARQL |
From the brainstorm in the dynamic worlds report. The tag shows how likely each one is to work.
| Idea | What it is | Odds |
|---|---|---|
| Sun-true rooms | Window light follows the real sun for the place, date and hour on the slider | high |
| Far-end impostor | A baked picture of the hall's far end shows before the bays exist | high |
| The address is the shelf | The URL names the library, bay, shelf and book, so any page can be shared | high |
| Evidence only | One switch hides every invented element | high |
| Left-behind objects | Marginalia, a ticket slip or a plan fragment appears when you walk close | high |
| Borges side-room | A hexagon room that leads into the reading list | high |
| Library card | Stamps from living pages, saved as a PNG | high |
| Evidence ledger | Sort claims about Alexandria or Nalanda into attested, inferred or legend | medium-high |
| Quote-to-JSON | An LLM turns a cited quote into hall-kit parameters for a person to sign off | medium-high |
| Mistranslation page | Two translations side by side where they disagree | medium |
| Unrolling scroll | A papyrus reader that unrolls sideways, for Herculaneum | medium |
| Splat hero zone | One photoreal Gaussian splat corner per room, the rest built from the kit | medium |
| Marble backdrop | An AI world seen through a window or doorway, tagged invented | medium |
| See it today | A click-to-load 3D street of the real site from free OpenStreetMap data | medium |
| Local guide | An opt-in narrator running in the browser that answers only from evidence rows | low-medium |
| Streaming by rumour | Load a room before anyone asks for it, based on guesses | not viable |
Yes, for Phase A. The open items below each have a default, so none of them blocks the first commit.
Planning checks
18 / 19
Each check is a command in data/planning-gaps.json that re-proves its status. The one failing check is listed next.
Open, with a default
Decided
Needed from the owner to start
5 / 5
Measure first in Phase A