How papers are chosen and scored

Stage 1 · Gathering

Each morning the pipeline pulls the newest papers and preprints from OpenAlex (which indexes most scholarly journals), arXiv, SocArXiv and PsyArXiv, searching the research areas learned from your seed papers. Each candidate is then compared to your seeds by semantic closeness (meaning, not keywords), and the closest form your daily shortlist.

Stage 2 · Selection by the judge

A large language model reads every shortlisted title and abstract against your Selection Criteria: the description you wrote, plus your topics, exclusions and example papers. It scores fit on this scale:

9–10 squarely inside one of your topics; the topic is the paper's central question
7–8 clearly within a topic, alongside other aims
4–6 matches one component (your area alone) without the combination
1–3 tangential; shares vocabulary, not the research question
0 off-target or on your exclusion list

Every score comes with a one-sentence rationale, shown as the quoted line on each card. If a rationale is wrong, edit your criteria.

Your briefing is ranked by fit score. Equally fitting papers are ordered by semantic closeness to your seed papers.

Your description versus your topics ("flavors")

Your Selection Criteria have two parts that do different jobs.

The description is the few sentences about your research from setup. The judge reads it first, as context: what you study, the angle you take, and what only sounds related. It orients the judge, but it is too broad to score against on its own. A description like "I study political communication and AI" matches thousands of papers a week.

The topics (called flavors in the scoring rules) are what the judge actually scores against. Each one names a specific intersection of things you care about, such as "narrative persuasion: narrative as a persuasion mechanism, in any domain" or "LLMs as persuaders in online political contexts". A paper squarely inside one topic scores 9–10; it does not need to touch the others. A paper that matches only one component of a topic, for example AI alone or politics alone without the combination, scores 4–6. The tags on each card's chip show which topics the judge thinks a paper engages.

In short, the description says who you are; the topics say exactly what counts. Most of the precision comes from the topics, so write them as combinations rather than single subjects. Edit both in Settings → Selection Criteria.

How votes are used

Thumbs-up and thumbs-down on any card become boundary examples shown to the judge from the next run onward, labeled as papers the researcher accepted or rejected. A few votes on borderline cases measurably improve selection. Clicks through to a paper are also recorded, as a signal of which briefings were useful.