Choose the paper workflow first
| Project | Job it documents | What remains your responsibility |
|---|---|---|
| JEV-Paper-Radar | Read new feeds and rank papers against plain-language interests | Set interests, thresholds and feedback labels; inspect the original paper |
| jev-paper-judge | Give structured feedback on an uploaded paper | Verify extracted text, citations and every recommendation against the PDF |
| jev-paper | Mark claims and caveats in an arXiv abstract inside a browser extension | Keep the authors' wording and read the full paper before relying on a mark |
| jev-paper-screener | Extract records and route them through Keep, Review, Drop or Incomplete decisions | Define inclusion criteria and complete the human screening required by the review method |
These tools are complementary. A radar is a discovery queue, a judge is a document-level review aid, a claim marker stays close to the abstract, and a screener formalizes eligibility. None of them replaces literature retrieval, provenance checks or a researcher's final decision.
Use Paper Radar for discovery
The pinned Paper Radar README describes one Noul per interest, a feed or RSS output, optional short summaries for selected papers and a record of probabilities and model versions. It also describes a cost measurement for one dated run; that number is the author's observation, so run paper-radar check for your own profile before budgeting.
Write one idea per interest and put exclusions in their own rules. A query that asks for several unrelated topics makes the threshold hard to interpret. Keep the title, abstract and source URL with each decision so a near miss can be audited later.
Use a screener for explicit inclusion gates
The Chinese-language Jev Paper Screener README documents PDF, URL and structured-file inputs, a title-and-abstract pass, and separate hard gates for topic, method transfer and evidence compatibility. Its routing labels are application rules. They need to be reviewed against the protocol for your own systematic review.
If the abstract is missing or the extraction is incomplete, keep that record visible as incomplete. A score computed from a truncated document is not evidence that the paper is irrelevant.
Keep document feedback and claim marking grounded
jev-paper-judge and jev-paper address different units. The first works from a paper document; the second keeps a claim and its caveat next to the source sentence. In both cases, preserve the paper URL, extracted passage and model version rather than exporting only a headline verdict.
A defensible review loop
- Retrieve records from a source you can name and deduplicate them before judging.
- Freeze the criteria, model ID and threshold for the run.
- Sample automatic Keep, Drop and borderline cases for human review.
- Store the original title, abstract or passage with each decision.
- Report how many records were incomplete, reviewed and changed by a person.
For the underlying retrieval pattern, see Jev search and RAG. For calibration and coverage tradeoffs, use the decision-model evaluation guide; a paper tool's output is not a universal accuracy score.