Refcheck

refcheck — reference integrity for the scientific literature

Checks whether the references in a paper actually exist, actually say what they are cited for, and actually match the work they claim to cite.

The premise is simple and it is the same one that governs the rest of my work: **never let a document validate its own claims.** A reference list is a set of assertions about the outside world, so every assertion is resolved against an independent registry (Crossref, OpenAlex) before anything is believed, and every flag ships with the evidence needed to overturn it.

**Status: work in progress.** The deterministic half of the pipeline (ingest → parse → registry resolution) runs end to end on real arXiv papers today, the evaluation harness is built and tested, and the numbers below are real. The retrieval and model-judgement stages are next. Numbers below are measured, not projected; where a stage is not built yet, this README says so.

---

Why this and not another RAG demo

Retrieval-augmented question answering over a document set is a solved-looking exercise that every portfolio now contains. The hard part of applied AI is not getting an answer out of a model; it is **knowing whether the answer is right**, and being able to show your work when someone disputes it.

So the centre of this project is the evaluation harness and the failure taxonomy, not the agent. The agent is the cheapest part to write.

The cascade

Every reference is assigned exactly one verdict, by the cheapest mechanism that can establish it. Expensive stages only see what the cheap ones could not settle.

Verdict	Meaning	Established by
RESOLVED	Exists, and metadata agrees with the citation	Registry lookup
METADATA_MISMATCH	Real record, but authors / title / year disagree	Registry lookup
UNRESOLVABLE	No matching record in any registry	Registry lookup
UNSUPPORTED_CLAIM	The work exists but does not support the citing sen