Research work fails in a particular way: a number is taken from a paper, travels through three documents, and arrives somewhere it can no longer be checked. We build systems where extraction, citation, versioning and provenance are structural, so a claim carries its source, a source carries its version, and a reviewer can reach the original passage without having to ask anyone.
A figure extracted from a table, transformed in a script and rendered in a chart usually loses its origin somewhere in the middle. Provenance has to travel as data through each step, or the only way to verify a published number is to redo the work by hand.
Findings live in PDFs, in tables split across pages, in figures and in supplementary files. Automated extraction produces something plausible from almost any input, so the useful system is the one that reports its own confidence and routes the uncertain cases to a person.
A systematic search returns far more records than the finished review includes. Deduplication across databases, consistent inclusion criteria, double screening with a recorded resolution path, and an auditable trail of exclusions are what make the result defensible.
Papers are revised and retracted, datasets corrected, models retrained and criteria amended mid-review. A system storing only the current state cannot answer what was known at the point a decision was taken, which is usually the question that matters most later.
An analysis that cannot be re-run is a claim rather than a result. Pinned dependencies, recorded parameters, seeded randomness and stored intermediate outputs are the difference between a finding that can be defended in review and one that has to be repeated from scratch.
An AI workspace for search, research, writing and fact-checking.
nexusityai.com (opens in a new tab)Answers are drawn from the live web and your own documents, with numbered citations to check.
Each question goes to the model best suited to it rather than one model handling everything.
Preferences and facts established in earlier sessions carry into later ones.
The agent works a stated goal through background analysis and returns a cited report.
A pasted draft is broken into its claims and each one checked against live sources.
Search across a large corpus is close to solved, and a language model will summarise anything placed in front of it. What neither does on its own is keep the summary honest: which passage supports this sentence, is that passage from the current version of the paper, and does it actually say what the summary claims.
We build the citation link as a first-class record rather than a formatting convention. An extracted claim points at a document, a version, a location and the supporting text, and the interface makes the jump from claim to passage a single action. A claim that cannot produce its passage is shown as unsupported rather than quietly rendered like the rest.
Language models are genuinely good at the first pass: proposing candidate extractions, clustering near-duplicate records, drafting a screening rationale, and flagging where two papers appear to disagree. They are unreliable as the final authority on any of it, and confidently so, which is the part that costs time further down the line.
So the model proposes and a person disposes, with the system recording which is which. Every field carries whether it was extracted automatically or confirmed by a reviewer, and only the confirmed set feeds analysis. That also produces the data needed to tell whether extraction quality is improving, rather than an impression that it is.
Research groups already have a method, and it is usually sound. The first phase encodes that method rather than replacing it: the search strategy in use, the real inclusion criteria, the extraction fields already being filled, and the way disagreements get resolved. Running one completed review through the system end to end finds the places where the written protocol and the actual practice differ.
Automation is then introduced where volume justifies it and error is recoverable: deduplication first, then candidate extraction, then screening assistance. Manual override stays available at every step, because a method nobody can intervene in is one nobody will trust with a publication.
Yes. Federating over subscription databases, institutional repositories and internal document stores is a matter of connectors and credentials. Access rules follow the source, so a record can be discoverable in search while its full text stays restricted to those entitled to read it.
By not letting it produce them. Citations come from the retrieval layer rather than from the model, and every generated sentence is checked back against the passages actually supplied to it. Anything without a supporting passage is marked unsupported rather than printed alongside the rest.
It does not have to. Most groups keep the reference manager and citation styles they already use, and the system exchanges records with it. Replacing it becomes an option later, once the extraction and provenance layer is carrying more of the work.
Sources are monitored against retraction and revision feeds, and a change propagates to every claim citing that source. The claim is flagged rather than deleted, because the fact that a conclusion once rested on a retracted paper is itself part of the record.
That is the intent. Pipelines record their inputs, parameters, versions and seeds, and emit a re-runnable specification alongside the output. A reviewer with access can execute the same pipeline and compare results rather than taking the numbers on trust.
Share the workflows, records, exceptions and integration boundaries the system needs to support. We will help define the technical scope.