Trading model catalog: Qwen3.8-27B row native_proven from li26 (#342) — micro-F1 0.9909, C2 26.28 tok/s, bounded

[seathatflowsinourveins]

What

This records the li26 result merged in #342 on the Qwen/Qwen3.8-27B row of the trading model catalog (catalogs/us-equities/models.json and models.md). The row moves from source_review to native_proven, for one host and one task:

• Measured: the unsloth/Qwen3.8-27B-GGUF UD-Q4_K_M quant at revision 4ca7207, served by llama.cpp b11146 at 8192 context. On the frozen 360-filing 8-K item-extraction task (2020-03-02 cohort) it scored micro-F1 0.9909 against the filer-declared header items, and every reply was valid JSON.
• Production now serves the C2 profile: --gpu-layers 99 --n-cpu-ffn 20 --spec-type draft-mtp --spec-draft-n-max 3. It decodes at a median 26.28 tokens/s with a median TTFT of 4,355 ms, against C0's 5.25 tokens/s and 11,378 ms at the same F1.
• Limitations kept on the row:
  • The labels are header ITEMS, not hand-checked.
  • Decoding was greedy with thinking off, about 17 tokens per filing, one run per arm.
  • The MTP gain holds only for this protocol.
  • There is no free-memory guard, and the rollback to C0 was not exercised.
  • MiMo-9B failed non-inferiority (lower bound −0.0560). B and X are runtime_unsupported on b11146.
• decision stays conditional, and the manifest-bound checked_at stays 2026-09-19. The update is dated inside the row.

Evidence

• evidence/receipts/local-inference-c2-serving-switch-20260926.json (kind native_model_e2e) satisfies the catalog validator's native_proven rule. results/decision.json is also cited.
• scripts/validate_catalogs.py and scripts/validate.py passed. tests.test_catalogs: 47 OK. tests.test_host_receipts: 180 OK.
• Cross-family review (GPT-6 gpt-6-astra, effort max, read-only, of a35316e): VERDICT: accept, with no findings. "Numbers, identifiers, pins, flags, and outcomes match the merged evidence", and the limitations "adequately bound native_proven and the production claim".

SOTA sources

• li26 protocol and results: blueprints/convergence-practice/local-inference-latest-20260926/ (Local-model qualification results on the RTX 4090: C2 (Qwen3.8-27B fully on GPU + MTP) 5x decode at equal F1, now serving #342)
• unsloth/Qwen3.8-27B-GGUF at the pinned revision: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/4ca720788d1e01f1bff70c033e0d0028fd02e502
• Base model card: https://huggingface.co/Qwen/Qwen3.8-27B/blob/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0/README.md
• llama.cpp b11146 (commit 7fe450e): https://github.com/ggml-org/llama.cpp/releases/tag/b11146

🤖 Generated with Claude Code

[chatgpt-codex-connector]

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