verse

A copy of the live deltaverse.pythai.net surface, plus the RAGE ingest stack that indexes it.

Two things live here, and they are licensed differently on purpose.

tree	what it is	licence
www/	The served docroot, byte-for-byte. Every path is a live URL.	Apache-2.0 OR GPL-3.0-or-later (LICENSE-APACHE / LICENSE-GPL) — except vendored subtrees, which keep their upstream licence
rage/, tools/	Server-side: pgvector/pgvectorscale ingestion, embedding, DDL, conversion-to-database-storage, the mirror sync script. Never reaches a browser.	MIT (LICENSE — the repository default)

The licence rule, in one sentence

www/ is the docroot. Everything under it is downloaded and executed by browsers, and is dual-licensed Apache-2.0 OR GPL-3.0-or-later. Nothing outside www/ reaches a browser, and is MIT.

Apache-2.0 is there for the explicit patent grant and the NOTICE trail an auditor asks for; GPL-3.0-or-later is there for copyleft and anti-tivoization. Downstream picks either. The dual grant exists to answer client-security review, not as a formality — this is code that touches the microphone (www/engine/frequency-source.js), the camera (www/ollywoo/faicey.js), and, in the case of www/engine/ngn/wordpress-reader.js, runs inside other people's WordPress installs.

MIT on the server side is a deliberate strategic release: take the ingestors, run them against your own corpus, keep what you build.

Three exceptions the one-sentence rule cannot express

Read LICENSING.md for the per-directory statement, NOTICE for the attribution trail, and .reuse/dep5 for the same map in machine-readable form. The short version:

1. www/vendor/ is upstream, not ours — three.js r184 (MIT, www/vendor/three/LICENSE) and MediaPipe tasks-vision (Apache-2.0). We redistribute them; we do not relicense them.

   www/coinlogos/ ships its addressing layer and NOT its images. The index/, manifest.ndjson, ipfs-map.json, resolve.py and SOURCES.md are here; the ~23,600 files under logos/ are not. Three independent reasons, any one of which would be enough:

   • Trademark. Every upstream repo is openly licensed, but the logos remain the marks of the projects they represent. SOURCES.md states it plainly: an open repo licence is permission to redistribute the file, not permission to imply endorsement. Restating 23,600 third-party marks under Apache-2.0 OR GPL-3.0-or-later would assert a grant we do not hold — precisely the defect this repository's licensing exists to avoid.
   • One store, not four. mindX, DeltaVerse, bankon and agenticplace all want these images. ipfs-map.json already addresses 1,859 of them by CID. Content-addressed bytes belong in one place that every consumer resolves from, not copied into every repository that renders them.
   • They break Windows. 1,917 directories and 617 files under logos/ contain a colon — they are CAIP-19 identifiers (logos/eip155:16688/ERIS.png). NTFS reserves :, so a full checkout fails with error: invalid path. The colons are load-bearing (manifest.ndjson keys every record by its CAIP-19 string and resolve.py resolves against exactly those names), so renaming them is not available. Leaving the images out removes the problem rather than pushing a sparse-checkout workaround onto every Windows reader.

   Nothing is lost: resolve.py and manifest.ndjson describe every asset, and the live host serves /coinlogos/ CORS-open today.

2. Two DeltaVerse-authored files hide inside www/vendor/. www/vendor/three/machine.mjs (the tier loader) and the 49 www/vendor/**/index.html folder-guard stubs are first-party and dual, not three.js MIT. A "everything under vendor/ is upstream" rule gets them wrong; LICENSING.md and .reuse/dep5 name them individually.

3. Six typeface files cannot be conveyed under GPL-3.0. www/vendor/three/addons/fonts/gentilis_*.typeface.json is SIL OFL 1.1, whose clause 5 forbids distribution under any other licence; helvetiker_* and optimer_* are MgOpen (MAGENTA Ltd, 2004), which adds a field-of-use restriction GPL-3.0 §7 does not permit. Both are freely redistributable as part of this larger work — which is what we do — but the dual grant explicitly carves them out rather than making a false claim.

Where it comes from

Nothing here is anonymous. Where a tree lives in a git repository we know, the repository address is cited rather than the bytes being presented as originless. Full tree-by-tree map with pinned commit SHAs and an honest account of where live has drifted from repo: PROVENANCE.md.

upstream	role
https://github.com/AgenticPlace/DeltaVerse.git	the site: pages, engine/, nGn/, ollywoo/, gfx/, vendor/
https://github.com/AgenticPlace/mindX.git	the RAGE ingestors, and four of the live pages (docsreader, docsplayer, codephreak, look)
https://github.com/GATERAGE/RAGEnet.git	sibling RAG federation layer — Apache-2.0, not a dependency of these ingestors, not vendored here
https://github.com/Professor-Codephreak/gitmind.git	MIT git monitor — noted, out of scope, not included
https://github.com/Professor-Codephreak/BinaryClock.git	www/nGn/binary-clock/
https://github.com/Professor-Codephreak/ultimate-input-field.git	ancestor of www/nGn/ultimate-input/ (clean-room rewrite)
https://github.com/mrdoob/three.js.git	www/vendor/three/ (r184)
https://github.com/google-ai-edge/mediapipe.git	www/vendor/mediapipe/

The mirror source is the docroot /home/deltaverse/www on the host serving deltaverse.pythai.net. tools/sync-from-live.sh refreshes it.

Size, and what to do about it

106,197,688 bytes   597 files   96 directories   (www/)

www/coinlogos      41M   the addressing layer only — index/, manifest.ndjson,
                         ipfs-map.json, resolve.py, SOURCES.md. No images.
www/vendor         31M   three.js r184 + MediaPipe (incl. 19M of .wasm)
www/audio          27M   40 .opus parts, machine-rendered narration
www/gfx           4.6M   first-party brand marks
www/engine        672K   the nGn substrate modules
www/ollywoo       468K   the face/voice organ modules
www/nGn           196K   binary-clock + ultimate-input
everything else  ~700K   12 root pages, one .css, the docs

It ships raw: no LFS, no submodules, no fetch-on-demand for anything that is here. The one thing deliberately not here is the coinlogos image payload, for the three reasons given above — and that is the difference between a 104 MB repository that clones anywhere and a 758 MB one that fails on Windows.

git clone https://github.com/cryptoAGI/verse.git

# code only — skips the 27M of audio and 31M of vendored wasm
git clone --depth 1 --filter=blob:none https://github.com/cryptoAGI/verse.git

No colons remain in any path, so this clones on Windows without a sparse checkout. That was not the reason the images were left out, but it is a real consequence of it.

Where to look

Of 597 files, roughly forty are the application. The rest is art, vendored libraries and generated audio.

• www/*.html — the twelve root pages are the app. Logic is inline, not linked: chainmarketcap.html carries 2,592 lines of inline JS, ollywoo.html 1,018, index.html 690.
• www/engine/ngn/*.js — the fifteen nGn substrate and expression modules, each opening /*! DeltaVerse nGn — …, each with a paired .js.xml doc sidecar.
• www/ollywoo/*.js — seven ES modules: the face rig, the webcam landmarker, the TTS envelope, the director console.
• rage/ingest/ — start at ingest_deltaverse_docs.py. It is the shortest path from a directory of Markdown to a queryable pgvector store.
• rage/schema/doc_embeddings.sql — the DDL. vector(1024), primary key (doc_name, chunk_idx), DiskANN with an HNSW fallback.

Because www/ is a mirror, curl is a legitimate way to navigate it: any repository path under www/ is the same path on https://deltaverse.pythai.net/.

The RAGE stack

rage/ is the ingestion and conversion-to-database-storage layer extracted from mindX, trimmed to what actually runs:

• store — PostgreSQL 16 + pgvector 0.8.6 + pgvectorscale 0.9.0, doc_embeddings as vector(1024) behind a StreamingDiskANN index.
• embeddings — bge-m3 (1024 dims, 8192-token window) by default, mxbai-embed-large as the alternate, resolved through a three-source cascade: vLLM /v1/embeddings → Ollama /api/embed → HuggingFace feature-extraction with mean pooling.
• ingest — walk, tier-prefix, chunk, upsert ON CONFLICT, prune tail orphans, write diagnostics.
• install — rage/install/install_pgvectorscale.sh builds the DiskANN index and proves it with an EXPLAIN showing Index Scan using *_diskann.

The database extensions are runtime dependencies of your machine, not vendored code — nothing from pgvector or pgvectorscale is redistributed here, so no licence obligation attaches to them.

Configure with RAGE_DB_DSN. No credential is baked into any file in this repository; the upstream copies of these scripts carried a hardcoded DSN and that is one of the things the extraction fixes.

Known gaps

Stated here because an honest gap beats a plausible silence.

• www/vendor/mediapipe/ needs its Apache-2.0 licence and NOTICE added. The vendored vision_bundle.mjs has had its header stripped and the two .wasm files (9.5M + 9.4M, the only WebAssembly on the site) carry no attribution. Apache-2.0 §4 requires both to travel with the binary.
• www/vendor/mediapipe/face_landmarker.task (3.7M model weights) and www/ollywoo/facecap.glb have unestablished terms. Google licenses model bundles separately from library source, and three.js example models carry per-asset credits distinct from the library's MIT. Recorded as unresolved rather than guessed.
• www/engine/ngn/audio_rack.js bundles dreamknob 1.1.4 (MIT, Dreamodus Software Inc.) but its bundled-licence footer names only the four React licences. NOTICE supplies the missing attribution.
• lil-gui 0.17.0 (MIT, George Michael Brower) and potpack (ISC, Vladimir Agafonkin / Mapbox) sit inside www/vendor/three/addons/ and are not covered by www/vendor/three/LICENSE. Both are in NOTICE.
• Five vendored three.js addons fetch from CDNs at runtime (RapierPhysics.js, JoltPhysics.js, 3DMLoader.js, the two WebXR input-profile files) and addons/inspector/ui/Style.js @imports Google Fonts. Nothing is redistributed, so no licence obligation — but it breaks the project's stated no-CDN posture and Style.js is the only third-party network beacon left in the docroot.
• www/coinlogos/SOURCES.md and report.md disagree on the source list (report.md names bgd-labs-web3-icons and polkadot-js-apps; SOURCES.md omits both). Both files ship verbatim; the disagreement is recorded, not papered over.
• www/docsreader-src/ contains one README and no source. The directory name promises sources it does not have.

Attribution required downstream

Two obligations travel with this repository and are not optional:

• "Data provided by CoinGecko" — required for the 27,977 cross-mapped asset records in www/coinlogos/. CoinGecko and CoinMarketCap imagery was deliberately not mirrored (marked rehostable: false in the manifest); only the cross-map is here.
• Trademark, as www/coinlogos/SOURCES.md puts it: "an open repo license is permission to redistribute the file, not permission to imply endorsement." Every logo remains the trademark of the project it represents.

---

Copyright © 2026 Professor Codephreak / PYTHAI / AgenticPlace, except where LICENSING.md, NOTICE and .reuse/dep5 record otherwise. Licence texts: LICENSE (MIT, the repository default), LICENSE-APACHE, LICENSE-GPL.