# Hub compatibility and provenance

This is a research-artifact dataset repository. The dataset card explicitly
selects `data/experiment_runs.jsonl` as the five-row `experiment_runs` subset with
a `test` split. The remaining files are research artifacts, not implicitly merged
into a heterogeneous table. `datasets.load_dataset` loads the result-summary table;
the CPU agent interface is executed locally as documented.

The root [README](../README.md) uses YAML language/tags metadata and manuscript
navigation. Agent files are ordinary repository files, read only by agents that
explicitly support or are instructed to read them. They do not activate an agent
on Hugging Face. Expert materials are review tools, not an endorsement.

The uploader uses `HfApi.create_repo` and `HfApi.upload_folder` with an explicit
`allow_patterns` list from the verified manifest. It does not delete remote files.
If an old repository contains obsolete files, an owner can review them separately.
Repeated uploads can create more than one commit after a transient failure; no
exactly-once publication guarantee is supplied. `--private` is a creation option
and does not change an existing repository's visibility.

Current upstream documentation used for compatibility checks on 7 October 2026:

- [Hugging Face dataset cards](https://huggingface.co/docs/hub/datasets-cards)
- [Explicit dataset configurations](https://huggingface.co/docs/hub/datasets-manual-configuration)
- [Uploading files with the Hub client](https://huggingface.co/docs/huggingface_hub/guides/upload)
- [Hub client API](https://huggingface.co/docs/huggingface_hub/package_reference/hf_api)

The exact client/dependency versions used for this edition are recorded in
`results/agent_expert_qa.json`. Client signatures are checked locally without
authentication. Live publication is not part of local packaging validation.

The original prior-art report is retained as historical source-release evidence.
This support edition does not redo scientific priority clearance or assert that
the original bounded search is exhaustive. The manuscript and research kernel
remain byte-identical to the incoming release.