Chhlafiu4312/citeguard

CiteGuard

English | 中文

CiteGuard is a citation linter and bounded metadata verifier for DeepSeek Harness. It extracts DOI, arXiv, URL, and Markdown citations from drafts, checks what can be checked mechanically, and labels every conclusion at the right confidence level.

It never turns “the link responded” into “the claim is true.”

Why it exists

AI-generated citations fail in several different ways: malformed identifiers, invented papers, wrong titles, dead links, or a real paper placed beside a claim it does not support. CiteGuard catches the mechanical failures while preserving a clear boundary around semantic review.

draft ──> extract + normalize ──> offline validation ──> bounded provider check
                                       │                        │
                                       └── proximity labels ────┴──> evidence receipt

What you get

• DOI, arXiv, HTTP URL, and Markdown-link extraction with normalization and deduplication.
• Stable line, column, context, and same-sentence proximity associations for every citation.
• DOI metadata from the fixed Crossref API and arXiv metadata from the fixed arXiv API, with redirects locked to the original provider host.
• Explicit title-similarity mismatch detection for descriptive Markdown links.
• Arbitrary URL checks only when networkMode=full is explicitly enabled.
• SSRF defenses: HTTP-only schemes, no URL credentials, canonical IPv4/IPv6 private and special-purpose range rejection, DNS answer validation, per-redirect validation, redirect caps, one end-to-end deadline, prompt cancellation of rejected response bodies, and response-size limits.
• A model-callable citeguard_check tool, standalone CLI, stable JSON receipt, and reusable TypeScript API.
• Citation and input caps that prevent a draft from causing unbounded network work.

The precise evidence contract is documented in docs/design.md.