Reading a literature is the part of research that doesn't scale. Fifty papers, each making a handful of claims, half of them quietly disagreeing — and somewhere in there is the actual state of the evidence. The tools meant to help mostly make it worse: a chatbot will cheerfully summarize all fifty into a confident paragraph with no way to check a single sentence.
Evidence takes the opposite stance. It's a new Glyph workstation that turns a body of literature into a living, citation-grounded evidence graph — and its first principle is that nothing exists unless it's cited.
What it actually does
Ask a research question. Find the papers — type a topic or paste DOIs and Evidence pulls them from OpenAlex — or paste the text yourself. Then an agent reads them: it extracts each claim, verifies it against the exact sentence in the source, and only then lets it into the graph. Concepts become nodes; relationships become typed edges — supports, contradicts, increases, decreases, moderates. Click any edge and you land on the quote it came from.
The result is a map, not a paragraph. And a map you can interrogate.
It finds the things you'd miss
- Contradictions. When one paper says X increases Y and another finds no effect, Evidence flags it — and shows you both sides with their quotes, instead of averaging them into a false consensus.
- Gaps. Two concepts studied heavily, but never together? That's a research opportunity, and Evidence surfaces it.
- Robustness. A leave-one-out sensitivity pass tells you whether a conclusion rests on a single load-bearing study or survives dropping any one paper.
It reads the way reviewers read
The same graph renders as a concept map, a causal DAG, a GRADE evidence table and Cochrane-style Summary of Findings, a sortable evidence table, and a PRISMA flow for systematic reviews. Pick an exposure and an outcome and Evidence computes the confounders and a back-door adjustment set with do-calculus. Date the claims and it shows when the evidence accumulated — and flags evolving findings, where an early belief was later overturned.
A research assistant that can't make things up
Evidence has an Ask tab — a multi-turn assistant that streams its answers as it writes. But unlike a general chatbot, it's grounded in your graph's verified claims plus the analyses Evidence derived from them. It will summarize the evidence, draft the narrative for a review, explain a contradiction, or characterize how certain the literature is — always with inline citations you can click to jump straight to the finding. If the evidence doesn't cover your question, it says so. It does not invent a reference.
And when it spots a gap, it does something about it: it proposes concrete literature searches to fill it, each a single click to run. Analyze, suggest, search, re-synthesize — the loop closes.
It lives
Save an Evidence project and re-run it as new papers appear. The graph diffs — "+4 concepts, 1 new contradiction" — with a full run history. Then export what you need: a citation-grounded Markdown report, BibTeX or RIS for any reference manager, a claims-level CSV for your stats package, or a portable JSON graph you can re-import. No lock-in.
Why this matters
The promise of AI for research has mostly been "read it for you." That's exactly backwards — it removes the one thing research depends on, the ability to check. Evidence keeps the checking and removes the drudgery: the synthesis is deterministic and auditable, the claims are quote-verified, and every figure that comes out the other end is ready for a manuscript or a PRISMA appendix.
It's live now, on every tier. Bring a question and a stack of papers.