← All papers · prometheus7.com

Wander Around — A Walkable Cartography of the Forgotten Web

State paper, 2026-05-04

Abstract

Wander Around is a single-server substrate that turns large public corpora into walkable 3D environments. Eleven walkable surfaces run on one Hetzner CPX42 node (89.167.7.54), spatializing roughly 720,000 items across nine distinct corpora — long-tail Common Crawl (6.83M useful cards in DB, 197K placed in /world), all of English Wikipedia (~6.8M articles, 200K placed in /wiki), Project Gutenberg (~22.9K books, 17.9K placed), Internet Archive Books (33.7K placed), arXiv papers (30.7K placed), DOAB open-access books (6.6K placed), plus an interactive 3D sphere view (/4d) and a thumb-only walkable subset (/rendered, ~33K cards with real WebP screenshots).

The thesis is simple: the virtual realm is a dimension. We just never had a projection of it. SEO turned the open web into a flat ranked list optimized for clickthrough — but the web is actually a vast city of hand-built homesites, weird hobbyist pages, irregular blogs, raw HTML. Most of it is invisible to the algorithms that decide what you see. Wander Around is a different projection: a cartographic instrument that lets you walk through what the algorithms threw away.

What’s running, end of 2026-05-04

Surface Corpus Items placed Status
/world long-tail web (Common Crawl) 197,512 / 6,830,239 useful full chrome, lazy thumbs, gold real-thumb markers
/game 2D mandala over web 200,000 (categories_flat) 8 graph paradigms, Set A/B/C panels
/rendered web subset with real thumbs 33,811 placed tighter pinwheel, populates as thumbs render
/wiki English Wikipedia 200,000 / 6,841,288 walkable, slow thumb pipeline (0.02/s)
/4d sphere of any corpus configurable 12 longitudinal continents, OrbitControls
/feed swipeable card feed 4M unchanged baseline
/library/gutenberg Project Gutenberg 17,959 / 22,876 covers fixed, 99.6% have real thumbs
/library/ia_books Internet Archive Books 33,711 placed covers partial
/library/arxiv arXiv papers 30,674 placed procedural facades only (papers, no covers)
/library/doab Open-Access Books 6,650 placed covers untested
/library hub default → gutenberg redirect dropdown switches corpora

Background services on Box C

Combined web-thumb throughput: ~1.5 successful renders/sec ≈ 130,000/day. At this rate the top 200K cards by score will all have real WebP screenshots within 36 hours.

Architecture

The substrate is corpus-agnostic. The same code path serves Common Crawl, Wikipedia, Gutenberg, arXiv, DOAB, and IA Books — the only difference is the source database and a small per-corpus adapter (scripts/library/library_layout.py).

The shared pipeline:

  1. Ingest — corpus-specific scripts pull metadata into a SQLite DB. Each row gets at minimum id, host/title, category, score.
  2. Layoutscripts/10_layout_ring.py produces a 12-sector pinwheel. Per-category radial arms are sized proportional to card count (personal/community arms reach 25-60K, sparse cats stop at 5-10K). Within each sector, era band widths are proportional to per-cat per-era distribution. Within each leaf cell, polar wedge fill places buildings as concentric rings with walkable streets between rows and angular alleys every 4 buildings. Slot size 28 ensures buildings (max width 16 in WORLD_JS) never overlap inside a row.
  3. Thumb pipeline — Playwright + wkhtmltoimage in parallel. Score-prioritized queue. Real WebP saved per card; procedural SVG facade as fallback.
  4. Serve — FastAPI streams chunked card data to the client based on player position (/api/game/region?cx=&cy=&r=).
  5. Render — Three.js scene loads only the ~9 chunks around the player at any time. The ~10 nearest in-frustum buildings get individual Mesh objects with real thumb textures + EdgesGeometry outlines + a gold cone marker if the card has a real screenshot. The bulk InstancedMesh holds the rest.

What changed in the past 72 hours

2026-05-01 evening session (the layout fight)

2026-05-01 → 2026-05-02 session (the surface explosion)

2026-05-04 (this session)

The Toon Town palette

The world’s permanent visual rule: floor and sky must never be the same color. Currently:

The MeshToonMaterial + 4-step gradient ramp gives all built geometry a banded cel-shaded look. Add new props with the same material to keep the visual language consistent.

The corpus numbers

What’s pending and why

The user has flagged three polish priorities for the current 3-day window:

  1. Covers / lazy thumbs: largely DONE for Gutenberg (99.6%). IA Books cover URL pattern still needs investigation (__ia_thumb.jpg vs services/img/). DOAB needs the expand= parameter. arXiv stays procedural.
  2. Toon Town aesthetic: largely DONE. Palette overhauled to bright cartoon, MeshToonMaterial throughout, clouds drifting, trees/lamps/benches added this session, cartoon shop-front facade replacing the web-2-formal look. Optional next: cel-shading outline pass on the bulk building InstancedMesh (technically harder due to Three.js InstancedMesh limitations; detail meshes already get EdgesGeometry).
  3. Gutenberg completeness: 99.6% (17,959 / 22,876). The remaining 4,917 are likely below the score threshold; could be addressed by lowering --max-cards floor in library_layout.py if user wants 100%.

Beyond the three priorities, deferred work includes:

Phantom players + novel marks

The presence layer makes Wander Around inhabited rather than just navigable:

These are operational but lightly used; visible polish (walking animation, distance fade, name-tag toggle) is deferred until traffic warrants it.

The substrate-template repo

C:\Users\BenHo\Desktop\ClaudeCode\wander-substrate\ is staged for public release. v0.1.0 tag prepared. Pending user actions:

  1. Create the GitHub repo
  2. git config user.email "..." (replace placeholder)
  3. git commit --amend --reset-author --no-edit
  4. git remote add origin <url>; git push -u origin main; git push origin v0.1.0
  5. Cut a Release from v0.1.0 with CHANGELOG content

Known issues / debt

Box C state at this writing

uptime              4 days, 1:23, load avg 17.46 / 18.12 / 18.73
disk free           116 GB of 301 GB (60% used)
wander-api          active
wander-thumbs       active (86% success, 0.80/s)
wander-wkhtml       active (82% success, 0.73/s)
wander-wiki-thumbs  active (slow, 0.02/s)
wander-warc         stopped (paused 2026-05-01 20:40 UTC)
wander-refresh      active timer

The load average is high but stable — running for 4 days continuously with all pipelines under demand. RAM has not OOMed. The system is comfortable.

Why this matters

The open web has not been mapped meaningfully since the early 2000s. The page-rank graph is mapped, but the content of pages — what they look like, what era they’re from, who made them, how hand-crafted they are — has been invisible to the tools that decide what you see. Wander Around makes that visible at city scale. Walking through 197K hand-made sites in /world feels like walking through a city of vanished neighborhoods. Walking through 17K Project Gutenberg books in /library/gutenberg feels like a public-domain library you can browse with your feet. The same substrate generalizes to any corpus with (id, category, score).

If wanderaround.io launches with all four library corpora at 80%+ thumb coverage and the Toon Town aesthetic settled, the surface area for going viral is substantial. The thumbnail-on-every-building moment (“wait, this is a real site I’ve never heard of”) is the conversion vector. We’re 36 hours from full thumb coverage at current pipeline rate, three days from Toon Town being polished, and one user-action from the domain being live.


Author: psiloceyeben (Prometheus7 Institute) with Claude Code. State of the project as of 2026-05-04 08:15 UTC. All deployment lives at 89.167.7.54. The substrate template lives at wander-substrate/ ready for public release.