UOR-R4 M1 model — the M1-closing 19-table discrete trunk (two seeds)
The M1-closing model of the UOR-R4 project: the trained 19-table integer discrete trunk
trunk-joint-pointer-v1-d11, seeds 1 and 2, released as self-contained UORLUT01
artifacts. These are the artifacts the project's M1 closing record and held-out-phrasing panel
were measured on (m1_close_pass: true, m1_milestone_status: CLOSED_PASS), and the model the
Studio's browser serving path loads as a next-token predictor.
- Schema
uor-r4.lut-discrete-trunk/1, 19 integer tables: the 4 discrete tables (discrete.table,discrete.cells,discrete.routes,discrete.gates) plus the 15 post-rotation joint-pointer trunk tables. - Shape: vocabulary 4,096 · blocks 4 · context 192 · kept_cells 8 · stages 5 · read_mode
relation· read_kept 8 · read_sweeps 1 · objectivenearest_address_margin_v1· read initializationactive_leaf_v1· epoch 420 · seed 1 / 2. - Sizes: 86,536,132 B (seed 1) and 84,943,328 B (seed 2).
- Runtime: native integer (
uor-r4-integer,IntegerDiscreteModel) and the project's WASM discrete serving path (loadDiscrete/predictDiscrete, top-1/top-8 next-token prefix prediction). No float is needed to load or step it.
This repo is separate from caseyallard/uor-r4-m1-discrete (the 4-table Stage-T base exports,
542 KB each): this is a different container family and runtime path, 160× larger, and it is the
milestone-closing artifact. Keeping the two releases apart keeps their provenance and download
sizes clear.
Files
| File | Seed | Bytes | SHA-256 |
|---|---|---|---|
t16-frozen-zero-seed1-trunk.d11 |
1 | 86,536,132 | 04652a3de8e966a23539b04fbd77c7d9872d9f13a541fcb224d48e766e2c147d |
t16-frozen-zero-seed2-trunk.d11 |
2 | 84,943,328 | 1b49e8bd823a1021e31de55c767111f9cf43807d3e6d007644e49281935877de |
tokenizer-seed-1.json |
1 | 265,393 | a9fb13f70c69ffe5afed40c8251a8e702adf0b00732d35ccff9457a3309ee377 |
tokenizer-seed-2.json |
2 | 265,393 | f804ecbee34e3c4e0001a737d003e3d620ba2c40722d416e7f3dd64843e2b30e |
SHA256SUMS lists all four. Tokenizer artifacts are stack-bound to their seed's artifact sha256
(not interchangeable) and were built from the project's pinned driver tokenizer (BPE, vocab
4,096, sha256 d36d3e8700a123e620012df77de195f244fbdb4d05d9e1aa7e77fa9407590f89) with the
project's extended make_tokenizer (accepts uor-r4.lut-discrete-trunk/1); each --check
reports TOKENIZER ARTIFACT MATCHES THE STACK.
Evaluation (verified from the project's committed evidence)
M1 closing, DEV panel, 14,287 rows (docs/labs/m1-close-saved-model/evidence/,
m1-close-grade.json = CLOSED_PASS):
| Seed | top-1 | top-8 | admitted | export gap (train vs D11) |
|---|---|---|---|---|
| 1 | 7,002 / 14,287 (49.01%) | 7,485 | 9,012 | 0 (14,287 / 14,287 rows) |
| 2 | 7,001 / 14,287 (49.00%) | 7,478 | 9,010 | 0 |
Floors: Unigram 1,226, T2 6,300, required 6,324; two-seed spread 1. Planted fault B4 PASS;
float/INT export parity exact (Δ = 0 on every row and aggregate; both GPU builds byte-identical).
M1 bar (e), Fresh V2 held-out panel, 14,479 rows (m1-fresh-panel-transfer,
trunk-joint-pointer-v1-d11-fresh-gate-receipt.json: outcome: PASS_BAR_E_FRESH_TRANSFER,
bar_e_pass: true):
| Seed | top-1 | top-8 | vs T2 | vs Unigram |
|---|---|---|---|---|
| 1 | 7,215 / 14,479 (49.83%) | 7,688 | +1,149 | +5,977 |
| 2 | 7,228 / 14,479 (49.93%) | 7,696 | +1,163 | +5,990 |
Floors: vs T2 ≥ 82, vs Unigram ≥ 156; both seeds pass; reference mismatches 0, refused rows 0.
M1 bar (f), held-out-phrasing panel (m1-heldout-phrasing-transfer, same two artifact
sha256s): saved-model integer top-1 7,850 / 7,872 (seeds 1 / 2), per that item's evidence.
Speed (committed evidence, no bar set): Apple M1 laptop d11-scorer toks (192-token prefill,
64-token decode, 1 warmup + 20 runs): seed 1 decode 11,026 tok/s median, prefill
2,596,805 tok/s; seed 2 decode 10,573, prefill 2,764,260. GPU device scoring (1×
RTX 5090, compute_120): 183,206 / 181,086 response tok/s; context 34,625,934 /
34,225,309 tok/s.
Provenance: m1-close-saved-model (antigravity-110, PR #3312) and
m1-fresh-panel-transfer (antigravity-111, PR #3318); pre-registration
docs/harness/prep/m1-close-saved-model-2026-10-26.md; 1× RTX 5090 pod vast-57903382,
source commit 0806df755c09c8fb0330e4457266c9f6f85ed874; checkpoints 0831900f… / 8bbb5950…;
sealed store caseyallard/uor-r4-store:results/antigravity/m1-close-saved-model-2026-10-26/outputs/build-a/
(archive m1-close-saved-model-20261026.tar, sha256 a758ca09…); this release is a
byte-identical copy of the stored build-a artifacts and the two builds A/B are byte-identical.
The 4-table base artifacts are the base_artifact_sha256 fc31bcc7… / a75ea7f5… of these
trunks (released separately).
How the Studio serves it (current status)
- The Studio's discrete serving path loads this container in the browser WASM engine
(
loadDiscrete/predictDiscrete) and returns top-1 and top-8 next-token predictions for a prefix of token ids. It is a next-token predictor, not a chat model. - First load downloads the artifact (~86.5 MB for seed 1); a user selects/loads it explicitly. The Studio page itself remains offline-first and makes no model request at page load.
- Greedy continuation repeats tokens. In the project's browser-path check, greedy stepping (always taking the top-1 token, appending it and predicting again) produced a repeating token sequence — the same behavior the native runtime shows. This is a next-token demonstration, not coherent text generation.
- The browser path is being landed in the project's Studio work in this release session; this model card describes the model artifact, and the Studio's current development branch is the serving code (not yet a public deployment claim).
- Prompt tokenization: the per-seed stack-bound tokenizers are included; token ids can be produced from text with them.
Limits
- Pre-alpha research model. M1's bars are the project's own narrow, sealed measurements (DEV, Fresh V2, held-out-phrasing) — not general language capability.
- Not an assistant, not chat-capable. No instruction following, dialogue, tool use, safety training or alignment is claimed. Output is next-token predictions.
- Greedy repetition (above) limits usable continuation; sampling may vary but no text quality is claimed.
- Size: 86.5 MB (seed 1) — a real download and memory cost in the browser; the seed-2 artifact is an alternate, not needed to run the default.
- Not a from-scratch chat baseline, and no claim is made that it beats any external model.
License
MIT — repository LICENSE: "Copyright (c) 2026 The Universal Object Reference Foundation (UOR)
... Permission is hereby granted, free of charge ... THE SOFTWARE IS PROVIDED "AS IS" ...".
Links
- Project: https://github.com/UOR-Foundation/uor-r4
- M1 close:
docs/loop/tasks/m1-close-saved-model/antigravity-110/REPORT.md,docs/labs/m1-close-saved-model/README.md - Fresh panel:
docs/loop/tasks/m1-fresh-panel-transfer/antigravity-111/REPORT.md - Held-out phrasing (bar f):
docs/loop/tasks/m1-heldout-phrasing-transfer/claude-211/REPORT.md