chengziyue1222/math-model-agent/research-model-literature

Research Model Literature

Build a traceable evidence base for modeling choices and parameters.

Workflow

1. Turn the topic into model, application, data, and validation search concepts.
2. Search current authoritative sources; prioritize original papers, official datasets, standards, and competition rules.
3. Search with an identifiable provider and fetch the source or authoritative metadata page. Save candidates explicitly in source_candidates; never inject a problem-specific bibliography from inside a reusable Skill script.
4. For every candidate record origin (SEARCHED, USER_SUPPLIED, or MANUALLY_ENTERED), retrieval_provider, UTC retrieval_timestamp, source_fetch_status, metadata-verification status/evidence, and content-relevance evidence. USER_SUPPLIED is not equivalent to VERIFIED.
5. Cluster findings by method and compare assumptions, datasets, metrics, limitations, decision coupling, uncertainty semantics, and validation design. Treat competition exemplars as style/method evidence only: never recover their hidden numerical solution, named entities, or tuned parameters.
6. Separate established findings, source-supported inference, and open questions.
7. Validate candidate identities and duplicates with algorithms.modeling_contracts.validate_source_records, then generate BibTeX with records_to_bibtex only from passing records. Record transferable design patterns separately from problem-specific outputs so that a later project cannot inherit an exemplar's answer.

Integrity Rules

Read references/integrity-rules.md before producing citations. Never invent a source, DOI, quotation, result, or BibTeX field. Mark inaccessible or uncertain records as unverified instead of completing them from memory.

Executable Contract

For repository-managed competition and audit work, run scripts/execute_skill.py with model_problem, selected_model, citation_requirements, and the explicit source_candidates file. Emit the contracted evidence roles; inspect them with python <skill-directory>/scripts/_runtime/skill_contracts.py --skill research-model-literature --profile competition (or audit). A bibliography without fetched-source provenance and relevance evidence, or generated from hard-coded task sources, is unverified. For an ordinary scoped literature lookup, keep only the evidence needed to support the answer instead of manufacturing empty registry artifacts.