Add sub-1.5bpe frontier + entropy-bpe kernels and result sidecars

[jagmarques]

Summary

• Lands four Kaggle reproducibility kernels for the sub-1.5bpe frontier and NIAH-cliff data cited in the paper
• Adds six result JSON sidecars in experiments/kaggle/results/ so numeric claims are traceable to runs

Kernels

Dir	Model	Protocol	Configs
nq_mistral_subbpe	Mistral-7B-Inst-v0.3	AQUA-iso PPL + live zstd-L22/Shannon bpe	FP16, K3V2→K1V1 pb=0
nq_llama31_subbpe	Llama-3.1-8B-Inst	same; rope_scaling via prepare_rope_scaling()	FP16, K4V2→K1V2 pb=0
nq_qwen3_subbpe_entropy	Qwen3-8B (NF4 weights)	AQUA-iso PPL + entropy-coded bpe	K3V2/K2V2 pb=0/pb=1
nq_mistral_niah_frontier	Mistral-7B-Inst-v0.3	chat-template NIAH, 4K+8K, 5 depths	FP16, K4V2→K1V2 pb=0

All kernels use ungated model mirrors; no HF_TOKEN required on secondary accounts.

Result sidecars

Six JSONs in experiments/kaggle/results/ covering Mistral, Llama-3.1, Qwen3, Yi and HQMQ comparison runs.

Sanitization

• Removed jagmardrop internal account name from comment lines (lines 8/33 Mistral, line 8 Llama)
• Removed # From CLAUDE.md: block with 3 lines of internal context from Qwen3 kernel header
• kernel-metadata.json id fields retain jagmardrop/ namespace (Kaggle push requires it)
• HF_TOKEN: optional env-only throughout (os.environ.get("HF_TOKEN")); None is fine for ungated models

Proof of work

VERIFY: git -C /tmp/land_kernels diff --stat origin/main...HEAD

 experiments/kaggle/nq_llama31_subbpe/kernel-metadata.json  |   20 +
 experiments/kaggle/nq_llama31_subbpe/nq_llama31_subbpe.py  |  602 +++++++
 experiments/kaggle/nq_mistral_niah_frontier/kernel-metadata.json  |   20 +
 experiments/kaggle/nq_mistral_niah_frontier/nq_mistral_niah_frontier.py  |  445 +++++
 experiments/kaggle/nq_mistral_subbpe/kernel-metadata.json  |   20 +
 experiments/kaggle/nq_mistral_subbpe/nq_mistral_subbpe.py  |  597 +++++++
 experiments/kaggle/nq_qwen3_subbpe_entropy/kernel-metadata.json  |   20 +
 experiments/kaggle/nq_qwen3_subbpe_entropy/nq_qwen3_subbpe_entropy.py  |  611 +++++++
 experiments/kaggle/results/nq_hqmq_qwen3_v3.json  |  621 +++++++
 experiments/kaggle/results/nq_llama31_subbpe.json  | 1003 +++++++++++
 experiments/kaggle/results/nq_mistral_niah_frontier.json  |  621 +++++++
 experiments/kaggle/results/nq_mistral_subbpe_frontier.json | 1094 ++++++++++++
 experiments/kaggle/results/nq_qwen3_subbpe_entropy.json  | 1031 ++++++++++++
 experiments/kaggle/results/nq_yi_subbpe_niah.json  | 1736 ++++++++++++++++++++
 14 files changed, 8441 insertions(+)

Secret scan: SCANNER: gitleaks - clean
Token grep: CLEAN - no token patterns found
No .company/, CLAUDE.md, *.tex, *.bib, paper/*, .planning/ staged.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.