Releases and artifacts

A release is how a finished model reaches people: a numbered version, a file you can download, and a fingerprint (a checksum) that proves the file is exactly the one that was tested. For UOR-R4 the file is meant to be one integer-only model file in a format called UORLUT01, released at milestone M8. That release does not exist. Two old software releases are on GitHub from the summer, plus one pinned-corpus data release, but both software releases belong to an earlier engine and neither is a qualified model. The release automation still targets that older engine too. This page explains the file format, the release machinery as it is, the Hugging Face repositories, and how checksums bind them together. Nothing ships until a trained model passes M1 to M8.

Labels used below: proof (derived from definitions or code), measurement (a recorded run with its artifact), hypothesis (untested). See the glossary.

flowchart LR
  P["Trained model<br/>(GPU, offline)"] --> E["Export to UORLUT01<br/>+ export gap"]
  E --> H["SHA-256 of the file"]
  H --> R["Versioned release<br/>vX.Y, notes bind scores"]
  R --> W["Web client<br/>verifies, caches"]
  R --> D["Desktop client<br/>verifies, stores"]
  R --> C["uor-chat CLI"]
  S[("Private HF store<br/>MD5SUMS / SHA256SUMS")] -.inputs and results.-> P

The planned path from a trained model to the clients. Every hop carries a checksum. The release box exists only as a plan for this model; the old pipeline served a different engine.

The model file: UORLUT01

UORLUT01 is the container the integer engine reads. Its layout, from the reader format.rs (proof by reading the code):

• The 8-byte magic string UORLUT01, then the header length as a little-endian 64-bit integer.
• A JSON header with the schema name, the model's shape, the group size, numeric settings, a list of matrices and tables, the exporter's provenance (never read by the serving arithmetic), and an optional record of a trained-in rotation snap, with the SHA-256 of its root set.
• Zero padding to a 64-byte boundary, then the data sections at 64-byte-aligned offsets.
• Each weight matrix stored as 4-bit codes, two per byte, with one scale byte per row and group. The scale byte holds a 4-bit mantissa \(m\) and a 4-bit exponent step \(d_e\), meaning a group scale of \((16+m)\,2^{\,e_{\text{base}}+d_e-4}\).
• Small integer tables (16-bit or 32-bit) for the few nonlinear steps.

Three schemas exist: lut-stack/1 (a softmax attention read), /2 (adds a pointer head) and /3 (a softmax-free flock rank read). The first two serve exp-table softmax reads, which the no-squashing rule bans. No schema fits today's rules as it stands: all three also carry SiLU and GELU tables and stream dense 4-bit weight maps through a table-based product, which the charter excludes, so a new table-native schema is likely needed (lut-serving dossier).

A built-in fingerprint. When the reader parses a file, it computes the SHA-256 of all its bytes and keeps it with the parsed model (format.rs, line 557). A release can therefore pin that hash, and a client can refuse any file that does not match.

Safe loading. Every offset, length and product of sizes is checked, so a malformed file returns an error rather than crashing. A file that declares a product-key memory is refused outright, because the integer kernel for it is not built (format.rs).

Saved sessions use a separate format, UR4SESS1, which records a conversation's state so it can be restored byte for byte (session_codec.rs).

The release pipeline as it is

Source: reading release.yml and the repository tree. This audit did not confirm that a tag build passes today.

The two software releases on GitHub are historical. v0.1 (19 August 2026) attached r4 binaries, a WebAssembly frontend and a model bundle compiled from a pinned SmolLM2-360M-Instruct snapshot, a converted transformer that the current serving rules exclude. v0.1.0-alpha (8 September 2026) is marked a pre-release: its notes say the alpha qualification was withdrawn after a recovery audit found the claims unsupported, and it is kept only for historical reproducibility. Neither is a release of the geometric model. A third GitHub release, "D3 pinned corpus: simple-wiki-20231101-sample" (29 July 2026), is a data release, not a model (release pages on UOR-Foundation/uor-r4, read 19 October 2026).

The owner has decided to repair the legacy crates, not archive them. The root crate and four others are to be repaired so that they build on the current core. Repairing restores the build only; their old mechanisms stay parked (charter, section 2.4). A repaired build would let release.yml run again, but it would still package the wrong format until it is changed.

What an M8 release must contain

The M8 bar is "one versioned, checksum-pinned UORLUT01 artifact" that meets the frozen bars of M1 to M7, with "reproducible installation, release notes binding the exact measured scores and known limitations, and independent verification on consumer hardware" (project track, M8; issue #2036). Read as a checklist, it is a plan, not existing work:

1. The model file with its SHA-256, plus the tokenizer it was trained with.
2. Release notes that bind the code SHA, the file hash, the tokenizer, the training data and seed, every measured score with its panel, the export gap, and the known limits. The charter requires every artifact to bind its data, seed, configuration, parameters and format (section 3.12).
3. Binaries: uor-chat for each platform and the WebAssembly bundle, each with a .sha256, and each re-passing the zero-multiply audit (native 44 symbol ranges, WebAssembly 7 of 7 at last count; project track, M4).
4. The Studio, published to /studio on the documentation site (charter, section 2).
5. Independent verification on consumer hardware.

Missing today: a release workflow for UORLUT01, WebAssembly and desktop packaging of the integer engine in that workflow, and one inference contract. Two contracts still coexist: docs/inference_contract.md for the older R4G1 runtime, and D11 rules R1 to R5 for the product (D11).

Hugging Face repositories

Both public repositories were published on 9 October 2026. Sources: the two Hugging Face repository pages; compute guide. The public model repository's visibility is inferred from the listing, which showed no private flag.

How checksums tie it together

A checksum proves integrity: the bytes are the ones that were pinned. It does not prove who made them, and it says nothing about model quality.

Go deeper

• Language model for the M1 to M8 path; products overview.
• Serving and export for the engine and the export gap.
• Integer serving dossier.
• Compute guide for the private store and uor-store.
• Charter, sections 2, 3.6 and 3.12.