THIRD_PARTY_NOTICES.md

Third-party notices

No model weights or engine source trees are included in this release.

Benchmark selections

• GSM8K: OpenAI, MIT License, https://github.com/openai/grade-school-math.
• MATH: Dan Hendrycks et al., MIT License, https://github.com/hendrycks/math.
• MMLU: Dan Hendrycks et al., MIT License, https://github.com/hendrycks/test.
• HumanEval: OpenAI, MIT License, https://github.com/openai/human-eval.
• HumanEval+: EvalPlus authors, Apache License 2.0, https://github.com/evalplus/evalplus and https://github.com/evalplus/humanevalplus_release.
• The needle dataset was generated for this study and contains synthetic filler and invented facts.

The release includes the fixed subsets needed to identify and rescore the reported runs. The source download URLs used by the original selection script were not all revision-pinned. The released subset files are therefore frozen by their own SHA-256 values rather than represented as byte-for-byte regeneration from mutable upstream branches.

Model artifacts (not redistributed)

• Qwen/Qwen3.8-27B, unsloth/Qwen3.8-27B-GGUF, bartowski/Qwen3.8-27B-GGUF, and turboderp/Qwen3.8-27B-exl3: Apache-2.0 according to the tested repositories. Exact revisions and file hashes are in environment/model_artifacts.json.

Engines and tools (not redistributed)

• llama.cpp: MIT, https://github.com/ggml-org/llama.cpp.
• ExLlamaV3: MIT, https://github.com/turboderp-org/exllamav3.
• TabbyAPI: AGPL-3.0, https://github.com/theroyallab/tabbyAPI.
• Matplotlib: PSF-based license, https://matplotlib.org/stable/project/license.html.
• Tectonic: MIT, https://github.com/tectonic-typesetting/tectonic.

The bibliography cites scientific and community sources for the limited propositions described in the paper; citation does not incorporate those works into this artifact.