Public sealed kit

Claim-stack v1 timing-cheap sources. SHA freeze and flags: TRAINER_DELTA. Challenge repo: kit/ on GitHub ↗ · results/claim_stack_v1/ ↗ · reproduce notes ↗ · repository root ↗ (access may be required until opened).

• train_gpt2_split_qkv_cheap_v1.py — sealed trainer · GitHub ↗
• muon_opt_split_qkv_cheap_v1.py — cheap-split Muon opt · GitHub ↗
• train_gpt2_split_qkv_v1.py — joint/dual-polar lineage (contrast)
• muon_opt_split_qkv_v1.py — match-joint default (contrast)