| name | sweep-research |
| description | Sweep tracked AI labs for significant new or late-arriving research artifacts, including models, papers, libraries, datasets, evaluations, agent harnesses, technical reports, and HuggingFace-only releases. Use for periodic arXiv sweeps, lab-by-lab discovery, team-name or researcher searches, benchmark discovery, and turning a deduplicated candidate list into verified Lab Index outputs. |
Sweep Research
1. Load the sweep policy
Read AGENTS.md → Periodic Arxiv Sweeps completely before searching. Treat it as the canonical source for:
- the current and previous month prefixes;
- internal team and collaboration mappings;
- HuggingFace API probes;
- researcher-name, evaluation, domain-expert, and agent-harness searches;
- late technical-report and Artificial Analysis worklists;
- inclusion and exclusion criteria.
Use add-output after discovery for representation, provenance, and filing details. Use sync-artificial-analysis for bulk score maintenance.
2. Define the sweep scope
Record the labs, month prefixes, and artifact classes in scope. Unless the user narrows the task:
- Cover at least the previous 4-6 weeks.
- Search the lab name and every known internal team or collaboration name.
- Probe the lab's research page, publications index, GitHub organizations, and recent HuggingFace model and dataset uploads.
- Search prolific researchers when lab-name searches are incomplete.
- Run independent searches for widely adopted evaluations, reasoning-process benchmarks, domain-expert benchmarks, and agent harnesses or execution stacks.
- Generate the late technical-report and newly scored-model worklists specified in AGENTS.md.
Parallelize independent lab or search-family passes when supported and authorized. Keep a shared candidate list with the discovery URL and reason each item may qualify.
3. Qualify candidates
For each candidate:
- Verify the artifact and release date from a primary source.
- Read the paper title-page affiliations or acknowledgements before assigning a lab.
- Apply the research-focus and exclusion criteria in AGENTS.md; do not turn a sweep into an exhaustive publication catalog.
- Decide whether the artifact is a model, paper, eval, library, dataset, blog, or announcement.
- Distinguish a new output from a late source or paper that belongs on an existing output.
Report rejected borderline candidates with a short reason when that helps the user audit the sweep.
4. Deduplicate before filing
Run the complete, untruncated dedup procedure in add-output for every candidate. Search the lab directory and name/version variants. If a near-match exists, enrich the incumbent entry instead of creating another file.
5. File verified findings
Follow add-output for every accepted artifact. In particular:
- trace structured model fields to primary sources;
- settle from-scratch versus derivative provenance whenever a model has an Intelligence score;
- attach a late technical report to the existing model rather than creating a duplicate output;
- add both news and output records when an announcement is also a canonical research artifact;
- preserve historical flagship markers.
Do not create or modify files when the user asked only for a sweep report.
6. Verify
After data changes, run:
npm run validate
npm run build
Summarize the search coverage, accepted outputs, updated existing entries, rejected candidates, and unresolved leads.
