NovusGraph
Every researcher I've watched read a paper does the same three things, roughly in this order. First, skim the abstract to decide whether it's worth their time. Second, hunt for the specific claim they came for — usually a benchmark number, a method comparison, an ablation. Third, if the claim is interesting, verify the evidence for it: which table, which figure, which citation chain.
None of those three steps benefits from reading the paper front-to-back. What they need is a graph: which papers make claims of type X? Which ones back those claims with strong evidence? Which ones contradict each other, and on what? The linear PDF is a bad interface for any of that. But it's what we have — so we built the interface we actually wanted, on top of the PDFs we have to accept.
The three primitives
NovusGraph's graph has three primitive nodes: claims, evidence atoms, and method families.
A claim is any assertion a paper makes — "our method reaches 92.3% on ImageNet," "diffusion models converge faster with X," "we hypothesize the loss surface is Y." Claims are the units of scientific statement. Papers make dozens; some fields make hundreds of claims per paper.
An evidence atom is the specific bit of the paper that backs a claim: a row in Table 2, the third column of Figure 4, the sentence in Section 3.1. Atoms are byte-addressable — you can point at them, quote them, and re-render them independently of the surrounding prose.
A method family is a set of techniques that share structural DNA — all the attention variants, all the KL-regularized objectives, all the two-tower retrievers. Families are how we let you say "show me the neighborhood of technique X" without knowing every alias of X.
What breaks when you stop at retrieval
Most retrieval-augmented systems stop at "here are the top-K relevant documents." That's fine for a demo. It falls apart the moment you actually try to do literature review with it.
- You get back papers, but you wanted specific claims, and now you have to re-read the papers to find them.
- Two papers can be "relevant" to your query and yet contradict each other — the retriever doesn't know that.
- Neither the citation chain nor the method family is visible in the results, so you can't ask follow-ups like "what did this technique originally build on?"
A claim graph is not a magic solve for any of those, but it's the right substrate for asking them. Retrieval finds the neighborhoods; the graph gives you the structure inside the neighborhoods.
What's next
We'll keep this blog light and specific. Next up: how the ingest pipeline extracts claims and atoms from a raw arXiv PDF, why we picked claim-level extraction over full-paper embeddings, and where the approach still has sharp edges. If you want to see it in action, open the workspace and ask a question you'd normally spend an afternoon on.
The graph is not the point. The graph is what lets you finally ask the questions you actually had.