How we built journal matching without a black-box percentage.

When we started building the journal matcher, the obvious thing was a percentage. Paste an abstract, get "Nature Neuroscience: 94% match." Clean, satisfying, easy to act on. We built that version. Then we showed it to researchers, and watched them not trust it.

The problem with a percentage isn't that it's wrong. It's that it's opaque. A researcher reading "94%" has no way to interrogate the claim — no idea whether they should override their own judgment in its favor, no signal to disagree intelligently. So they don't trust it. And not trusting it means not using it.

Every match is a vector of evidence

Internally, every match is a vector of evidence. Keyword overlap with recent papers. Editorial scope from the journal's stated aims. Special-issue alignment. Citation neighborhoods. Author affiliations. A percentage compresses all of that into one number — and loses everything that made it useful.

So we stopped compressing. Every match now shows three things: the overlapping keywords (verifiable), recent papers the venue published on your topic (interrogable), and the data source the metric came from (citable). The match score is still there — but it's not the headline. The reasoning is.

Optimizing for justified confidence

Most AI products optimize for conversion — get the user to act. Research tools have to optimize for something stranger: justified confidence. The researcher needs to be able to defend the decision, to themselves and to their advisor. That means every claim needs to be traceable to a source, every recommendation needs a counter-argument the user can examine, and nothing should ever feel like it's hiding the reasoning.

It's slower. The interface has more text, less hero animation, fewer "magic moments." But the users we built this for don't want magic. They want a tool they can verify.

The number on the screen still says 94%. But when you click on it, three reasons unfold. That's the product.