Citare
One person processes a paper with AI; everyone should inherit the result.
Core trick
Converts papers into claim graphs storing verbatim source quotes, page numbers, and boundary conditions, served over an MCP endpoint.
Description
One person processes a paper with AI; everyone should inherit the result.
LLMs can read papers, but PDF reading is not claim infrastructure. PDFs are long, messy, and parsed inconsistently: one run may drop a mediator or boundary condition, while another turns “associated with” into “causes.” And the more important the paper, the more often this unstable process is repeated. For field-defining papers, it is plausible that a million people will ask AI to read the same paper over the next decade. That wastes energy and amplifies citation drift.
Citare uses Claude Opus 4.7 to convert papers into reusable claim graphs. Each claim stores source text, page numbers, evidence type, causal strength, safe citation verbs, boundary conditions, and integrity warnings. Other tools return metadata; Citare returns the actual claim and makes it reusable through an open database and CitareMCP.
Try it now at https://citare.dev/
README excerpt
# Citare — Citation Integrity for AI Research > **PDF reading is not claim infrastructure.** Citare turns papers into reusable claim graphs with source quotes, page numbers, safe citation verbs, boundary conditions, and integrity warnings. > > Other tools return metadata; **Citare returns the actual claim** — verbatim, with the design-basis context an AI needs to cite it without lying. > > *DOI is for papers. Citare is for claims.* Live MCP endpoint: **`https://citare.dev/mcp`** (Streamable HTTP, no auth required). Built during the *Built with Opus 4.7* hackathon (April 2026). ## The problem Modern AI applications cite academic work at scale. Three failure modes recur: 1. **Silent causal upgrade.** A paper says *"X is associated with Y (cross-sectional, N=150)"*. The citing AI says *"X causes Y"*. The paper's own caveats about design basis and temporal precedence evaporate. 2. **Mediation dropout.** A paper shows *"X → M → Y, direct effect X→Y becomes null with M controlled"*. The citation extracts only the bivariate headline and loses the attenuation. 3. **Restatement propagation.** A paper introduces a complex theorem and later restates it in simpler form for a Discussion section. Downstream citations chase the simplified form and lose the theorem's actual conditions. None of these are hallucinations of non-existent content — they are selective readings of real content. Which makes them harder to detect than pure fabrication, and lets them propagate at internet scale.