Paper Sweep

Paper sweep

Answer “what recent work is worth reading?” Use the Research Radar CLI for collection and the host model for interpretation. The keyword/regex prior is a cheap ordering signal, not a claim that a paper is strong or SOTA.

Runtime

Run research-radar doctor first. If the command is unavailable in this checkout, use .venv/bin/python -m research_radar; elsewhere use the Python environment in which the user installed Research Radar. Do not hardcode a user's home directory or create an unrelated vault. The repository README contains installation instructions.

Collect

bashCopy code
research-radar sweep --source both --days 2 --data-dir data
research-radar sweep --source arxiv --days 7 --cluster World-Models --all --data-dir data
research-radar sweep --source file --input conference-papers.json --data-dir data

Use --date YYYY-MM-DD to anchor a historical window. The default is yesterday in UTC. arXiv submission dates and HF featured dates have different meanings; report which is used. --source file accepts a JSON paper list or an existing paper bundle; do not describe an imported list as a fresh online search.

Read the JSON summary printed by the CLI to find this run's papers.json, run.json, and report. Read the full machine records when assessing papers. All collected records, including papers that do not match keywords, remain in the bundle. --cluster, --top and --all control the reading view, not the completeness of that bundle. Check the source status and truncation flag before describing coverage.

Triage and deliver

• Review the title and full abstract for relevance; correct the primary cluster and keep useful secondary labels in your analysis. Sample the unmatched pool before asserting that a topic has no new work.
• Distinguish the author's claims, the heuristic prior, and your assessment. For strong novelty or performance claims, read the relevant methods, experiments, limitations, and code before endorsing them.
• Write concise explanations of the problem, contribution, evidence, and remaining uncertainty. Link the original paper.
• Report new versus previously seen records, the date window, and any partial source failure. Never turn “fetch failed” into “no papers today.” Exit 0 means collection succeeded; exit 2 can mean partial/truncated collection or invalid arguments; read the JSON/error before deciding whether usable records exist. Exit 1 means failure.
• For team investigation, pass the same papers.json to talent-scout; do not refetch or parse a shortened Markdown preview.

Configuration and operation

Defaults are packaged with Research Radar. Pass --config /path/to/clusters.yml for a custom topic profile; the YAML beside this skill is an editable example. The same data directory maintains cross-run state. If a source fails or is truncated, rerun the same command at most once; if it still fails, deliver the partial result with its status instead of looping. Broaden the window only when it addresses a real coverage gap, not to fill a quota. Scheduling is optional host orchestration and is not installed by this skill.