Skip to content

docs(TODO.impl): 03 separated-negative, 06 flat — recipe ledger complete #103

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ronaldtse merged 1 commit into from
Oct 1, 2026
Merged

docs(TODO.impl): 03 separated-negative, 06 flat — recipe ledger complete #103

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions TODO.impl/03-headwise-muon.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,14 @@ not an architecture change — single-variable, cheap to ablate.
- [ ] Flag-on path exercised end-to-end in a smoke run
- [ ] TODO.arabic/18 updated to point here

## VERDICT (2026-10-01): NEGATIVE — separated. 4.8164 vs 2.1's 4.5701

Full-set windowed protocol, teacher 2.2890 reproduced exactly, labels
e70ce991. Paired between-students bootstrap on final_preds: delta
+0.2267pp, CI95 [+0.016, +0.411], p=0.017. Head-wise Muon (positive
at DeepSeek's 671B) is significantly HARMFUL on a 300M byte-student
distillation — scale-boundary recorded. 2.1 stands.

## Status

- [x] Module built (gpu/muon.py: add_headwise_group, per-slice
Expand Down
7 changes: 7 additions & 0 deletions TODO.impl/06-sinkhorn-update.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,13 @@ carries to the row/col normalizations here if magnitudes get small).
- [ ] Spec entry (off by default) for the optional measurement run
- [ ] No change to any shipped run's reproducibility

## VERDICT (2026-10-01): FLAT — 4.5547, not separated from 2.1's 4.5701

Paired delta -0.0903pp, CI95 [-0.267, +0.069], p=0.858. The
Sinkhorn-balanced embedding update is indistinguishable from AdamW on
the tied tables at this scale. The module stands (Engram side-opt);
the recipe lever is closed.

## Status

- [x] Implementation (gpu/sinkhorn_update.py: SinkhornUpdate +
Expand Down
Loading