research-lineage

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Research Lineage

Build conservative, evidence-visible links between the current paper and cited prior work.

## Workflow

1. Parse numbered bibliography entries from the uploaded paper.
2. Retrieve every in-paper context that cites each entry.
3. Filter relationships by the requested concept or method when one is supplied.
4. Classify each relationship as `extends`, `modifies`, `adopts`, `compares`, `evaluates_on`, or `background`.
5. Require explicit citation-context language for `extends`, `modifies`, or `adopts`. A graph edge alone proves only that a citation exists.
6. Prefer the paper's own citation sentence over abstracts, recommendation similarity, or citation counts.
7. Enrich titles, abstracts, authors, citation intent, and influential-citation metadata through Semantic Scholar only when configured.
8. Return the exact source page and citation context with every relationship.
9. Mark ambiguous relationships as `background` instead of guessing a stronger dependency.
10. Keep newer citing work separate from earlier referenced work; direction matters.

## Output

For every connected paper, provide:

- bibliography number and title
- relationship label and confidence
- one-sentence explanation
- exact current-paper citation context and page
- external URL and metadata when available
- uncertainty or conflicting evidence

## Guardrails

- Do not call a paper foundational because it has many citations.
- Do not claim that the current paper extends another merely because both mention the same concept.
- Do not let external summaries override the uploaded paper's stated relationship.
- Separate architectural inheritance, training-objective inheritance, benchmark lineage, and general background.
- Cache lineage by paper and concept so repeated clicks do not repeat graph requests.

From Siris2314/deep-paper-reader, MIT.

More from Siris2314/deep-paper-reader