evidence-first-research（证据第一研究）

Evidence First Research

Overview

Use a research-before-starting workflow. Search existing evidence, tools, datasets, and established patterns before drafting an analysis plan, recommending a method, or producing scientific content.

Default to reuse or adaptation of validated approaches. Only introduce a novel method, pipeline, or claim of novelty after establishing that the gap is real.

Core Workflow

1. Define the task precisely.

• Restate the objective, target deliverable, domain, and decision stakes.
• Identify whether the task is literature synthesis, study design, data analysis, protocol drafting, manuscript support, tool selection, or interpretation.
• For medical work, identify the population, setting, intervention or exposure, comparator, outcomes, and time horizon when applicable.

1. Search before acting.

• Search for prior papers, systematic reviews, guidelines, benchmark datasets, existing tools, libraries, protocols, ontologies, and reporting standards before proposing work.
• Prefer primary and authoritative sources over tertiary summaries when accuracy matters.
• Search for negative results, contradictory evidence, failure modes, and replication attempts instead of only supportive results.
• Inspect local project artifacts before suggesting a new workflow when the task depends on an existing codebase, dataset, or protocol.

1. Evaluate evidence quality.

• Rank sources by relevance, rigor, recency, and direct applicability to the question.
• Distinguish peer-reviewed papers, preprints, guidelines, textbooks, package documentation, and informal discussion.
• Prioritize strong syntheses and well-matched study designs over isolated or weakly related findings.
• Flag uncertainty explicitly when evidence is indirect, outdated, conflicting, underpowered, or drawn from a mismatched population.

1. Choose the action mode deliberately.

• Adopt an established method when a strong, well-matched pattern already exists.
• Adapt a validated method when the problem is similar but not identical.
• Benchmark multiple credible approaches when the field lacks a dominant standard.
• Design a new approach only after documenting what was searched, what already exists, and why it is insufficient.

1. Execute with traceability.

• State the chosen approach and why it was selected over alternatives.
• Cite the papers, tools, libraries, datasets, or standards that informed the decision.
• Separate evidence, inference, and speculation.
• Preserve assumptions, inclusion criteria, exclusion criteria, and unresolved questions.

1. Re-check before finalizing.

• Verify that the strength of the final claims matches the strength of the underlying sources.
• Re-open the search if a key assumption is unsupported or if a stronger source is likely to exist.
• Perform an extra review for harms, contraindications, bias, and guideline consistency when the output is medically relevant or otherwise high stakes.

Search Targets

• Search for literature first:

systematic reviews, meta-analyses, guidelines, seminal papers, recent high-quality studies, protocols, replication studies.

• Search for research infrastructure:

benchmark datasets, registries, ontologies, reference implementations, software packages, analysis pipelines, laboratory or clinical standards.

• Search for methodological patterns:

study designs, statistical approaches, outcome definitions, preprocessing conventions, validation schemes, reporting frameworks.

• Search for practical constraints:

data availability, licensing, regulatory context, ethical constraints, reporting expectations, reproducibility requirements.

Decision Heuristics

• Prefer "adopt" when the question is standard and the field already has a stable method.
• Prefer "adapt" when the method exists but the data, population, or setting differs.
• Prefer "benchmark" when several plausible methods compete and no clear winner exists.
• Prefer "invent" only after showing that existing methods, tools, or study patterns do not adequately solve the problem.

Medical Emphasis

• Give extra weight to clinical practice guidelines, systematic reviews, meta-analyses, and pivotal trials when the task involves patient care, diagnostics, treatment, prognosis, or safety.
• Treat preprints, conference abstracts, animal models, in vitro studies, single-center retrospective studies, case reports, and expert opinion as lower-certainty evidence unless the task specifically requires them.
• Avoid patient-specific recommendations without current sources, clear scope limits, and explicit uncertainty.
• Flag when geography, formulary availability, regulatory status, standard of care, or population differences may change the answer.
• Distinguish mechanistic plausibility from clinical effectiveness, and surrogate outcomes from patient-important outcomes.

Output Pattern

Before doing deep work, produce a concise research checkpoint when useful:

• Objective.
• Search targets.
• Best existing papers, tools, or patterns found so far.
• Evidence strength and important gaps.
• Chosen path: adopt, adapt, benchmark, or invent.
• Main risks, assumptions, and next step.

Anti-Patterns

• Do not start building, analyzing, or writing as if the problem were novel without checking the literature and existing tools.
• Do not equate "published" with "reliable" or "recent" with "best."
• Do not overgeneralize from weak evidence, surrogate endpoints, or mechanistic arguments.
• Do not rely on abstracts alone when methods or limitations matter.
• Do not ignore population mismatch, confounding, missing comparators, or sample size limitations.
• Do not present speculation as consensus.
• Do not skip contradictory evidence just because it complicates the answer.

References

• Read references/evidence-evaluation.md when prioritizing sources, appraising medical evidence, or checking for common biomedical failure modes.