Write the question as a claim with a scope, for example "X reduces Y by ≥ Z in setting S" (applicability).
Search at least two providers with native queries. Record each provider, query, date, and filter (search). Then run a disconfirming query (contrary).
For each candidate paper, pin its version (DOI, or arXiv vN) and look up its corrections and retractions (identity, updates).
Read the full text of every paper the answer depends on. Record the section, table, or figure in an evidence note (evidence level, location).
Run check_stats.py on the quoted results. Appraise the design, and record effect sizes with their uncertainty (stats, design, effect).
Synthesize across studies by comparing designs, not by counting significant results (conflicts).
Run check_evidence_note.py on every note. Answer with scoped conclusions, each linked to its notes, and list what was not read or not searched.

Route the task to a card

Task	Card
Find papers on a question	Provider and native query
Which version is this?	Version identity
Avoid one-sided results	Disconfirming search
Is it retracted or corrected?	Corrections and retractions
Scripted or bulk queries	Rate limits
"Nobody has studied this"	Zero results
Can I cite this abstract?	Evidence level
Summarizing findings	Claim to location, effect size
Do the numbers add up?	Statistical consistency
Is the study sound?	Design appraisal
Papers disagree	Conflicting studies
Does it apply to us?	Applicability

Rules

Conclusions rest on full text, or on code and data, never on metadata, abstracts, citation counts, or search rank.
Every source has a DOI, or an arXiv ID with its version. Every finding names a section, table, figure, or page.
Report effect sizes with uncertainty and n. "Significant" alone is not a finding.
Absence claims name the providers, queries, and date searched.
Do not run code from a paper without reviewing it. Never infer a mailto or credentials.

scripts/fetch_metadata.py --provider arxiv|crossref|openalex (--query Q | --id ID) [--param K=V] [--limit N] [--output FILE] queries one provider natively per call and returns the native response, with bounded retries. Tested by scripts/test_fetch_metadata.py.
scripts/check_stats.py FILE|- recomputes p-values from t, F, χ², z, and r reports.
scripts/check_evidence_note.py NOTE.md... checks evidence notes against the template's rules and refuses "supported" when only an abstract was read.
scripts/test_analysis_tools.py tests the two checkers.
assets/evidence-note-template.md; assets/examples/{note-supported.md, note-abstract-only.md, results-section.txt}; assets/metadata-fixtures/.
sh assets/examples/verify.sh [network]: offline by default; network adds live, read-only metadata queries.

The searches run: provider, query, date, and filters, including the disconfirming query.
The evidence notes, with check_evidence_note.py passing, and the corrections lookup for each source.
The check_stats.py output for the quoted statistics.
Scoped conclusions: supported, contradicted, or unresolved, each linked to its notes.
What was not read, not searched, or not accessible (for example, paywalled full text).

The decisive paper's full text is not accessible, and the conclusion would rest on its abstract.
The question's scope (population, setting, date) is too vague to judge whether a finding applies.
A paper's code would need to run with side effects or credentials.