brandonhimpfen/ai-provenance-corpus：AI 溯源事件语料库 v0.1 发布

A machine-readable corpus of provenance events describing how AI-assisted artifacts are retrieved, generated, cited, edited, verified, reviewed and published.

The corpus records process metadata rather than artifact content. It is intended for research, benchmarking, standards development, auditing, RAG evaluation, trust modelling and provenance graph experiments.

Design principles

• Event-centred: meaningful actions are represented as append-only events.
• Content-minimizing: records describe how an artifact was produced without requiring the artifact body or private prompts.
• Graph-ready: stable identifiers and explicit edges support property-graph and RDF-style transformations.
• Validation-first: every core entity is governed by JSON Schema Draft 2020-12.
• Layered: synthetic, public-derived, reference and research records can coexist while remaining distinguishable.
• Reproducible: dataset metadata, versioning, checksums and citation files support formal releases.

Repository structure

ai-provenance-corpus/
├── data/
│   ├── artifacts/artifacts.jsonl
│   ├── events/events.jsonl
│   ├── sources/sources.jsonl
│   └── relationships/graph_edges.jsonl
├── docs/
├── examples/
├── schemas/
├── scripts/
├── tests/
├── CITATION.cff
├── LICENSE
├── metadata.json
└── pyproject.toml

Core entities

Quick start

Requires Python 3.10 or newer.

python scripts/validate.py
python scripts/build_manifest.py
python scripts/export_graph.py --format csv --output build/graph

No third-party Python packages are required for the basic structural checks. Full JSON Schema validation is available through the optional validation dependency:

python -m pip install -e '.[validation]'
python scripts/validate.py --json-schema

Example provenance chain

The included sample follows one synthetic research brief through:

1. semantic retrieval,
2. model generation,
3. citation attachment,
4. human editing,
5. automated and human verification,
6. editorial approval, and
7. publication.

The records intentionally omit the generated prose. Hashes and metadata may be used to bind records to external artifacts without redistributing those artifacts.

Dataset layers

Every artifact and event includes a dataset_layer value:

• synthetic: purpose-built test or benchmark records;
• public: records derived from openly observable workflows;
• reference: carefully curated canonical examples;
• research: records collected through a documented study protocol.

Trust and confidence fields

Confidence values are optional observations, not claims of truth. The schema supports retrieval, citation, verification, editorial and overall dimensions. Producers should document how each score was calculated and should not compare values across collection protocols without calibration.

Identifier policy

Identifiers are opaque strings with type prefixes:

• art_ for artifacts
• evt_ for events
• src_ for sources
• agt_ for actors or agents
• edge_ for relationships

Identifiers should remain stable within and across published versions.

Privacy and licensing

Do not include private prompts, personal information, proprietary source text, credentials or model secrets. Use hashes, controlled vocabularies and external references where content binding is necessary.

The repository code and schemas are licensed under Apache-2.0. Dataset records are dedicated under CC0-1.0 unless a release states otherwise. Source metadata does not override the rights attached to the underlying source.

Citation

Use the metadata in CITATION.cff or the DOI attached to a formal release. Before publication, replace placeholder author and repository fields in CITATION.cff, metadata.json, and .zenodo.json.

Status

This is a release-ready v0.1 foundation, not a claim of a comprehensive real-world corpus. The included records demonstrate the data model, validation workflow, and intended research use.