research-model-literature

Research Model Literature

For national-award design references, cross-check official showcase identity with award evidence, record the exact PDF/printed pages actually viewed, and exclude inaccessible or unverified papers from completed samples. Keep design notes separate from model facts and prohibit reference-solution numbers from entering the independent solve.

Verify document completeness separately from award identity: compare page order and missing pages when a mirror differs from the official display. An abridged journal version may connect authorship but cannot replace unread contest pages. End a candidate search when independent routes add no identity evidence, then seek another eligible sample; never fill the count from a provincial award or a repository's unsupported label.

Attach external parameters and supported claims to existing evidence pointers, keeping source identity verification separate from content support. Ordinary chart types do not require invented citations; read references/figure-source-handoff.md when an external method, scale or parameter changes the plotted interpretation.

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. For each intended citation, record the exact supported claim, source locator, reading depth and transfer limits. Bibliographic identity and content support are separate checks; source-service failure is not proof of a fabricated reference.
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.