Skip to content

docs(TODO.impl): recipe arms launched; engram double-step defect recorded #101

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
Sep 29, 2026
Merged

docs(TODO.impl): recipe arms launched; engram double-step defect recorded #101

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
12 changes: 9 additions & 3 deletions TODO.impl/03-headwise-muon.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,12 @@ not an architecture change — single-variable, cheap to ablate.

## Status

- [ ] Inspect muon.py for the built module's exact API
- [ ] Implementation + tests
- [ ] Smoke run
- [x] Module built (gpu/muon.py: add_headwise_group, per-slice
Newton–Schulz via _muon_step_headwise)
- [x] Implementation + tests (tests/test_muon_headwise.py: headwise ==
per-slice vanilla at h=1, differs for heterogeneous heads)
- [x] Recipe arm LAUNCHED 2026-09-29: ara-diac-small-2-1-hwmuon ->
run-014-hwmuon (2.1 recipe + headwise_muon; 40 headwise q/k
routed, 88 matrices stay whole-matrix; canonical r7 labels
e70ce991 pre-seeded, labels trusted). Watchdog armed. Gate on
landing: full-set CER vs 2.1's 4.5701.
17 changes: 14 additions & 3 deletions TODO.impl/06-sinkhorn-update.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,17 @@ carries to the row/col normalizations here if magnitudes get small).

## Status

- [ ] Implementation
- [ ] Tests
- [ ] Optional run decision
- [x] Implementation (gpu/sinkhorn_update.py: SinkhornUpdate +
sinkhorn_balance, log-domain — the mHC NaN lesson)
- [x] Tests green on real tensors
- [x] Recipe arm LAUNCHED 2026-09-29: ara-diac-small-2-1-skembed ->
run-015-skembed (2.1 recipe + sinkhorn_embed; the 4 2D
embedding-like tensors — tied table + heads — routed from AdamW
to SinkhornUpdate at 2.6e-4 via the _engram_table_opt side-opt
hook; canonical r7 labels pre-seeded). Watchdog armed. Gate on
landing: full-set CER vs 2.1's 4.5701.
- Routing defect found while wiring: split_parameters handed
_engram.table.weight to Muon AND the Sinkhorn side-opt
(double-stepped) in the completed run-013-engram. That verdict
stands as measured; fixed so future engram/sinkhorn arms route
each tensor exactly once.
7 changes: 6 additions & 1 deletion TODO.impl/10-engram-lexical-memory.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,4 +87,9 @@ depth ✗✗, memory ✗ flat, on-policy ✗, routing ✗, soup ✗. 2.1
- [x] Run (launched under the standing directive, resumed 7× through
preemptions; two eval traps caught: dropped-table load, wrong
harness, tied-alias strictness — all fixed in PR #222)
- [x] Verdict recorded in RESULTS.md (PR #223)
- [x] Verdict recorded in RESULTS.md (PR #223)
- Post-verdict routing defect (2026-09-29, found while wiring the
sinkhorn arm): split_parameters also handed the table to Muon —
double-stepped beside the Sinkhorn side-opt for this whole run.
The FLAT verdict stands as measured with that wiring; fixed in
interscript-ml PR #225 for future arms.
Loading