AI-Research: the original root research workspace

Before there was a Rust language model, there was a Python workspace called AI-Research. It held the first experiments on routing information by geometry instead of by dense attention, a benchmark for AI agents in a game world, and some small mathematics studies. This project is a snapshot of that workspace, kept so the history can be traced. The claims inside it are from the Python era and belong to that workspace; they are not claims about the current model, which is pre-alpha.

flowchart TD
    W["AI-Research workspace, Python era"] --> R["ai-router: routing research and paper drafts"]
    W --> M["MUDBench: agent benchmark and game world"]
    W --> Q["ramsey: Ramsey-style math experiments"]
    W --> D["deepermath and loose notes, plots, CSVs"]
    R --> N["Later: R4 prime router in Rust, then the UOR-R4 geometric model"]

The workspace's four parts. Only the router line leads to the current model.

What is in the project

The snapshot has 3,335 files and about 358 MB (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). The inventory counts about 10 MB of code across 678 code-like files (669 of them Python), around 207 MB of JSON, CSV, log and NPZ data, about 28 MB of images and about 119 MB of documents (1,037 Markdown files, 123 LaTeX files, 93 PDFs).

PartSizeWhat it is
ai-router/about 335 MB (router-research about 229 MB, preprints about 101 MB)Geometry-native routing research: transport laws, wheel-based experiments on prime admissibility, run outputs, a staged "kill list" tracker, and the paper 1 manuscript rebuilds (versions 1 to 4 and a submission bundle) with a claims_to_evidence_table.md
MUDBench/about 3.7 MBA benchmark and game-world runtime for evaluating agents: specifications, source in src/, tests, scenarios, a scoring spec
ramsey/about 2.2 MBScripts and result tables for recursive Ramsey-style studies
deepermath/about 1.8 MBEarly prime-router scripts and a few papers
loose filesthe restPlots and CSVs from wheel experiments (moduli 210, 2310 and 30030), base-fiber admissibility, and transport-error comparisons across three configurations

The workspace's own entry points are README.md, PROJECT_MAP.md and the router's docs/research/KILL_LIST_TRACKER.md and ACTIVE_STATE.md. Its paths in PROJECT_MAP.md point to a different machine and no longer resolve.

What it claimed

The workspace's EVIDENCE_SUMMARY.md lists what it had explored, each with "Observed" and "Unknown" lines. For example, geometric routing produced sparse activation in prototypes, with long-range stability and scaling listed as unknown. In its prime admissibility work, a shared transport law fitted on wheel sizes 2310, 30030 and 510510 is described as generalizing to the unseen size 9699690 with only small degradation, and the note itself says this is evidence for reusable structure and not for exact closure or a "prime oracle".

Treat these as the old workspace's statements. This page reports what the files say, not what was reproduced. Several nested copies of the same material exist inside the snapshot; the labels such as "canonical" apply to the old source repository.

Where it stands

StatementKindEvidence
The snapshot is 3,335 files and about 358 MBmeasurementfiles.sha256 (count); sizes from the audit inventory docs/labs/repo-renormalization-2026-10-19/audit/folders.json
It contains router experiments, an agent benchmark and small math studiesmeasurement (file listing)research/ai-research/README.md
The router results transfer to the Rust modelhypothesis. The model's own tests are the evidencesee the mechanisms
Any quality claim about languagenone. Nothing here trains or scores a language model for conversationproject README
Duplicates1,799 of its 3,117 distinct file hashes also appear in prime-analysis; 10 appear in research-archivesfiles.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/ai-research.tar (377,595,904 bytes, 3,335 files, SHA-256 dbdb1fa6f09712b9faba207a0188d23d0a591c9773cdcbb7ec6aeb21b8ddf1ce). Every file inside is listed in files.sha256, and the tar is listed in tars.sha256.

Source files brought into this folder

Only source and short notes come in; all data stays in the store.

  • README.md, PROJECT_MAP.md, and the loose *_note.md files
  • MUDBench/src/** and MUDBench/tests/**
  • ai-router/**/*.py, after removing files that duplicate research-archives, and ai-router/preprints/paper1_arxiv_rebuild_v4/**/*.tex
  • claims_to_evidence_table.md files
  • ramsey/**/*.py and deepermath/**/*.py

Stay in the store: all CSV, NPZ, PNG and PDF files, large run outputs under router-research/, preprint build artifacts, and 1608.04200v2.pdf (a third-party arXiv paper).

The owner's inventory suggests splitting this project into three folders later (router research, MUDBench, Ramsey). That is a proposal and is not decided; this page treats the snapshot as one project.

Go deeper