claim-graph

A citation graph tells you which paper pointed at which paper. It cannot tell you which claim supported, extended, or contradicted which other claim — and that is the relation a researcher actually wants when looking for a gap.

This is a working vertical slice: real papers in, typed edges between extracted claims, an explorer over the result.

ingest (OpenAlex / S2) → classify and type edges → graph (Kùzu) → explore

Where the slice stands

The pipeline runs and the graph is queryable. The unresolved part is whether claim extraction holds up on papers outside the domain it was tuned on.

Working

End-to-end, on real papers with provenance kept

Open

Extraction quality out of distribution

Status

Slice working

The explorer over the built graph. Nodes are claims, edges are typed by relation, and every node keeps the sentence it came from.Open this on its own page

The honest gap

Extraction was developed against a computer-science corpus. Applied outside it, accuracy has not been measured, so nothing about coverage should be believed yet. That is a measurement task, not a modelling one, and it is the next thing.

Provenance is kept per sentence, which means a wrong edge can be traced back to the sentence that produced it. That property is worth more than a higher score from a pipeline you cannot audit.

What it is a slice of

The longer aim is a claim-level layer over the literature: papers as inputs, ideas as nodes, and lineage that survives one of the papers being retracted. This slice exists to find out which parts of that are hard, cheaply.