# Source Text And Evidence Policy

The Human Value Observatory stores metadata for all research sources, but it only stores source text where doing so is legally and practically appropriate.

## Rules

- Store metadata for all sources.
- Store extracted text only where reuse is clearly permitted, practically stable, or provided by the Observatory/user.
- Prefer public-domain, openly licensed, government, institutional, or user-provided text.
- For paywalled, metered, restricted, or uncertain sources, store metadata and short curator notes only.
- Do not reproduce full copyrighted articles.
- Use short excerpts only where allowed and necessary.
- Always preserve source URL, publisher, access notes, chunk ID, and section/page references where known.
- Avoid fake page numbers. If page is unknown, cite section and chunk ID only.
- Label keyword retrieval as keyword retrieval. Do not imply semantic or AI understanding where only keyword matching is used.

## Current Phase 4 Status

The first Phase 4 build uses a static chunk store under `data/source-text/`. Current examples are short curated/source-grounding notes, not a complete full-text archive. They exist to test retrieval, citation discipline, evidence cards, and source-grounded prompt generation before larger ingestion work.

Generated prompts must instruct AI tools to use only the provided metadata and passages, distinguish direct evidence from interpretation, and state uncertainty when passages are insufficient.