Built to be checked, not believed.

Five scoring dimensions

Each paper is scored on topical fit, methodological fit, relevance to your active positions, impact on ongoing projects, and actionability. The first three carry standardized definitions that are identical for every user; the last two are shaped by your profile.

Challenges are checked twice

A “possible challenge” flag means the paper may be in tension with one of your stated positions. Before you ever see that warning, a second, stricter model pass tries to refute it: only findings or direct arguments count, not topical overlap. Cases suppressed by the second pass are not deleted — they appear in your weak-signals list.

We deliberately say possible challenge, never “contradiction”. The judgement whether something truly challenges your position remains yours.

Evidence binding

Every assessment must quote a fragment of the paper's title or abstract verbatim. The quote is verified by code, character for character — an assessment whose quote is not actually in the metadata is rejected before it reaches you. If no abstract is available, the card says so and the confidence label is capped.

Reproducibility

Every evaluation permanently records the exact configuration that produced it:

profile_version   · criteria_version · prompt_version
model + version   · metadata_version · timestamp
raw structured response · your feedback

When you edit your profile, a new version is created; past evaluations keep referencing the version they were scored against. Nothing is rescored silently.

Validation before claims

Before the open beta, precision and recall are measured on a gold standard built with real research profiles, judged blind by the profile owners themselves. A second rater checks only whether each justification is actually supported by its quoted evidence. Targets and their pre-committed consequences publish together with the validation report.

What your feedback does — exactly

Marking hits relevant or not relevant does three documented things: it updates the precision statistics of your monitor; it can add the judged paper as a visible, deselectable example to that monitor's evaluation prompt; and it generates suggestions for your profile or queries that you explicitly accept or dismiss. Nothing retrains silently — with a few dozen labels per quarter, any claim of “the model learns you” would be dishonest.

Sources, privacy, honest limits

Your profile is private: it is never part of any shared cache, and only the parts needed for evaluation are sent to the model provider — under a data processing agreement, excluded from training, hosted in the EU.

Honest limits: PaperCurrent reads titles and abstracts, not full texts. It does not guarantee complete coverage of the literature and is not a tool for systematic reviews. Scores are priorities, not verdicts — read the papers that matter.