Research archives: the Gemini-era sandbox

This folder is a time capsule. In an early stage of the project (the "r3/r4 era"), the owner built a development sandbox with an AI assistant: text corpora, small Python tools that generate and index them, a first "prime router" prototype, and an early Rust checkout of a WebAssembly router. It is kept because it shows how the ideas started. It is not a live project, it does not run the current model, and it contains no result that the current model relies on.

flowchart TD
    S["gemini-dev sandbox, r3/r4 era"] --> C["Text corpora: about 63% of the bytes"]
    S --> T["Python tools: corpus generation, GloVe loader, indexing, server, CLI"]
    S --> P["prime-router prototype: Python, with a web page"]
    S --> R["Early Rust wasm-router checkout, vendored UOR standards"]
    S --> X["scratch notes and extra_reading PDFs"]
    R --> N["Superseded by crates/ in the current repository"]

What the sandbox contains. The Rust part has been replaced by the current workspace.

What is in the project

All of it sits under gemini-dev/. The folder is 99 MB in 128 files (research-archive/2026-10-19/files.sha256 in the store; sizes from the audit inventory docs/labs/repo-renormalization-2026-10-19/audit/folders.json). About 65 MB of that is data (JSON, NPZ, token files and text corpora) and about 37 MB is documents.

PartSizeNotes
wikitext2_proxy/about 62 MBA stand-in for the WikiText-2 language-modelling corpus
Corpora (complex_corpus.txt, universal_corpus.txt, wiki_corpus.txt, world-model corpora)about 12 MB, 11 MB, 3 MB and smallerText used by the prototype tools
prime-router/about 3.4 MBA Python prototype: "Evolving Hypersphere Brain World Model" with a web page and a CLI
extra_reading/about 2.9 MBResearch PDFs with text extracts
rust/about 0.6 MBAn early uor-r4-wasm-router checkout with vendored UOR standards; its engine file is uor_r4_router_engine.rs
Python tools55 .py files overallCorpus generation, a GloVe loader, large-corpus indexing, a server and a CLI
scratch/, zeta_data/smallWorking notes and zeta-zero data

The original README says the GloVe and manifold caches (about 490 MB) and a nested copy of the AI-Research workspace were left out of this snapshot.

One thing to know before reading the code. The prototype describes a "zero-weight" geometric text generator, and it also offers a mode that connects to a locally running transformer through Ollama for comparison. The current charter forbids neural comparators and controls, so nothing in this folder is run or revived; it is preserved as history only.

Where it stands

StatementKindEvidence
The sandbox is 128 files, about 99 MBmeasurementfiles.sha256 (count); sizes from the audit inventory docs/labs/repo-renormalization-2026-10-19/audit/folders.json
The prototype's "zero-weight" generation worked as describedhypothesis, not tested for this pagegemini-dev/prime-router/README.md is the only source
Any part of it is on the current model pathnoThe Rust checkout is superseded by the current crates; the folder is marked archive
Overlap with other projects10 file hashes are shared with ai-research and 10 with prime-analysisfiles.sha256 comparison, 19 October 2026

Where the originals are

The full project is preserved in the private Hugging Face dataset caseyallard/uor-r4-store under research-archive/2026-10-19/tars/archives.tar (103,435,776 bytes, 128 files, SHA-256 2553007578044fa4f5f7000714a633b0693d7714f80cb999e0671e7ee71e78b2). Every file inside is listed in files.sha256, and the tar is listed in tars.sha256. The tar is named archives; the folder here is called research-archives.

Source files brought into this folder

Only scripts and notes come in, as an index of what existed:

  • README.md
  • gemini-dev/*.py
  • gemini-dev/prime-router/**
  • gemini-dev/scratch/**/*.md
  • an index (a listing, not the files) of gemini-dev/rust/**/*.rs, since the current crates replace it

Stay in the store: wikitext2_proxy/, the corpora, all NPZ and token files, the extra_reading PDFs and zeta_data/.

Go deeper