Restate the question, target system, decision, outcome, and applicability constraints. Expand acronyms, synonyms, adjacent fields, and contrary terms.
Run the discovery CLI with several queries. Use --offline only to reuse a populated cache. Deduplicate preprints, revisions, and published records before selecting papers.
Open each potentially decisive original paper independently. Identify the version read, method, workload or dataset, baselines, measured result, and limitations. Inspect supplements and linked code when the claim depends on them. Do not infer conclusions from title, abstract, citation count, venue, or discovery rank.
Search explicitly for replications, contradictory findings, later revisions, and follow-up work. Separate peer-review status from relevance and evidence quality.
Synthesize only supported claims. Prefer a compact table:

Work/version Supported evidence Limitation Applicability
State the strongest supported conclusion, competing evidence, concrete engineering implication, remaining uncertainty, and the first measurement to run on the user's own workload.

The CLI queries keyless arXiv, Crossref, and OpenAlex routes, caches responses, and emits normalized metadata. Its output never substitutes for reading a paper. Read discovery and evidence for source semantics, acquisition, and reporting rules.

Validation

Trace every title, author, identifier, venue, and numerical claim to fetched metadata or the paper itself. Link the exact paper version used. Label unavailable full text and unresolved version linkage. Do not invent citations or present a local benchmark as a scientific conclusion.

Run python3 scripts/test_discover_literature.py after changing the CLI.

research-scientific-literature

Research Scientific Literature