Trainer delta — timing-cheap split-QKV

Sealed candidate kit (claim-stack v1) is public here under kit/ (byte-frozen copies on this site). Corresponding trainer tree: rainbowpuffpuff/nocap/kit ↗ · sealed blobs train_gpt2_split_qkv_cheap_v1.py / muon_opt_split_qkv_cheap_v1.py · evidence results/claim_stack_v1/ ↗ · reproduce notes ↗. Muon attribution: Keller Jordan. Timing-cheap split is this campaign’s banked treatment — not a claim that we invented Muon or Ortho/NS.

Ecosystem context only: Modded-NanoGPT.

Files

File	Role	SHA-256 (PREREG freeze)
`kit/train_gpt2_split_qkv_cheap_v1.py`	Sealed candidate trainer	`becfce387c3f82ba1a98d2c83d69b8382a99cb406e283e19882a6072a7ac9cbc`
`kit/muon_opt_split_qkv_cheap_v1.py`	Cheap-split Muon wrapper (same NS×5)	`9086531ff1c70e1eb30473a020f5ec555b91cd489900d8ee8c5e7f3bf06b1b20`
`kit/train_gpt2_split_qkv_v1.py`	Joint / dual-polar lineage trainer (contrast)	not the seal freeze
`kit/muon_opt_split_qkv_v1.py`	Split Muon with default `match_joint_frobenius=True`	not the seal freeze

SHA freeze source: results/claim_stack_v1/PREREG.md ↗.
Verified byte-identical to Claim-W1 and Claim-W2 volume kit_override/ (2026-07-17 pull).

What “timing-cheap” changes

Candidate seal flags:

--muon_split_qkv 1
--muon_split_qkv_match_joint 0

Mode	Flags	Behavior
Joint stock control	`muon_split_qkv=0`	Fused QKV; historical joint Ortho/NS
Dual-polar / match-joint	`split=1`, `match_joint=1`	Three block NS5 maps plus a second joint polar to rescale Frobenius — extra work on the clock (~1.026× step tax → NO-GO)
Timing-cheap (sealed)	`split=1`, `match_joint=0`	Three block NS5 maps only — no second polar, no extra closed-form scale

Same kernel: zeropower_via_newtonschulz5 (Keller Ortho / NS×5). Split ≠ inventing Ortho — see split-vs-ortho-explained.html.

Claim boundary

Sealed wall-clock evidence is the full stack (Muon B32 + timing-cheap split-QKV) vs F1_seal, not Muon-alone and not matched % vs joint Muon stock (N/A). Numbers: NUMBERS_CANON.html.