literature-scout

name	literature-scout
description	LLM guidance for designing literature search profiles and queries.

Literature Scout Guidance

Use this guidance before calling search or paper-processing tools.

LLM Responsibilities

• Infer the research domain from project.yaml, seed papers, seed ideas, constraints, user-provided resources, and the researcher-confirmed literature/retrieval_scope_plan.json when present.
• Treat the scope plan as three distinct inputs. core_lines define normal mainline retrieval and must never be relabelled as a Bridge. adjacent_lines are selective theory/method or comparison leads. Only cross_domain_bridges can use a Bridge ID or the costly must_explore reservation semantics.
• Cover every must_cover core line through the smallest non-redundant set of queries. Combine closely related lines into one query where that preserves the research distinction; do not mechanically issue one search per line.
• Build a domain_profile with inclusion concepts, exclusion concepts, ambiguous terms, target venue/category hints, dataset or benchmark names, and related subfields.
• Design diverse queries from multiple angles: core mechanism, task/application, evaluation setting, baseline family, adjacent field, and recent terminology.
• Choose the source portfolio from the project-derived domain profile. Use broad scholarly indexes for every project, add informs_search for OR/MS, management science, information systems, operations, supply chain, queueing, or optimization questions, and favor computer-science sources for CS-heavy topics. An empty specialized source is acceptable and must not block T2.
• Decide whether a candidate is genuinely relevant. Tools may rank, deduplicate, verify metadata, and persist records, but they cannot make final domain judgments.

Tool Boundary

• expand_queries only combines your query ideas with seed-title phrases and date windows.
• filter_by_domain only applies the domain_profile you provide. If the profile is weak, keep uncertain papers instead of deleting them.
• enrich_papers can apply your llm_annotations; without them it only performs conservative schema completion.

Output Discipline

• Do not invent paper metadata.
• Do not let venue names alone decide relevance.
• When uncertain, preserve the paper with a review note rather than filtering it out.