Benchmark source attribution

Original IFEval input prompts: Google Research, instruction_following_eval,
commit e6890f85757dd84e27ca6df2dd30651dafad28e0.
https://github.com/google-research/google-research/tree/e6890f85757dd84e27ca6df2dd30651dafad28e0/instruction_following_eval
Apache License 2.0: IFEval-Apache-2.0.txt in this directory.
Prompt excerpts are shortened and formatted for this recorded visualization.
The original checker is used separately after generation.

Tau2 runner and public benchmark source: Sierra Research,
commit fc0055dc4e0a316c3f83133267fbd6faaa770992.
https://github.com/sierra-research/tau2-bench
MIT license, Copyright (c) 2025 Sierra Research: tau2-MIT.txt.
Visible banking/tool excerpts are shortened and formatted; fictional benchmark
accounts are used. Private simulator instructions and grader targets are omitted.

ITBench-AA: Artificial Analysis, derived from IBM ITBench.
https://huggingface.co/datasets/ArtificialAnalysis/ITBench-AA
https://github.com/IBM/ITBench
Dataset license: Creative Commons Attribution 4.0 International.
https://creativecommons.org/licenses/by/4.0/
Selected source-derived observations and source/line references are indexed,
shortened and formatted for replay. Local root-group matching is separate from
the official normalizer and does not constitute the full official evaluation.

Generated candidate responses, critiques and revisions are recorded model
outputs. Crystal positions and camera motion are illustrative, not embeddings,
KV caches or hidden model reasoning. Source links identify provenance.