External code snapshots
Papers are in this directory as PDFs (provenance.tsv names the arXiv version and its sha256). The other kind of original is someone’s code: where a formula was checked against a reference implementation rather than against prose. Those are not vendored — the code stays where its authors put it, and this file records which revision we read, so a claim can name it (AGENTS.md §1.4).
Snapshot = URL + commit + date. No code is copied into this repo.
| what we read it for | repository | commit read | license | read on |
|---|---|---|---|---|
| the Newton–Schulz coefficient triple, and its printed table | NoahAmsel/PolarExpress |
71cc37943d99 |
MIT | 2026-09-30 |
GDN-2 reference ops (fla/ops/gdn2), KDA’s kda.py decay init |
fla-org/flash-linear-attention |
9f38d24980c4 |
MIT | 2026-09-29 |
the official chunked KDA forward (tests/torch_ref.py) |
MoonshotAI/FlashKDA |
7afb9f454f16 |
MIT | 2026-09-29 |
| GatedDeltaNet-2 erase/write gates (NVlabs) | NVlabs/GatedDeltaNet-2 |
a5552fe3c67e |
NOASSERTION | 2026-09-29 |
Muon+ ColRow normalization and its param routing |
K1seki221/MuonPlus |
8a9ace123afe |
not declared | 2026-09-29 |
| DSpark draft head, acceptance head, DeepSpec repo | deepseek-ai/DeepSpec |
005e03b81cec |
MIT | 2026-09-29 |
the official Engram oracle (engram_demo_v1.py) |
deepseek-ai/Engram |
fb7f84a21f91 |
Apache-2.0 | 2026-09-29 |
| Attention Residuals reference implementation | MoonshotAI/Attention-Residuals |
85e22310fe5e |
not declared | 2026-09-29 |
| modded-nanogpt Track 3 optimization record | KellerJordan/modded-nanogpt |
4ea6b937337a |
MIT | 2026-09-29 |
Two of these are not free to reuse, and neither is vendored:
NVlabs/GatedDeltaNet-2 declares no license (NOASSERTION — read, do not copy)
and K1seki221/MuonPlus and MoonshotAI/Attention-Residuals declare none
either. What we implemented came from the papers; these rows say what the
formulas were checked against.
71cc37943d99 is still PolarExpress’s HEAD as of 2026-10-01, which is why
AGENTS.md can cite it by short sha.