Mathodology Evidence Search

Use evidence to support a claim or model choice. Prefer primary work and official data. Mark assumptions and unresolved gaps; never fill them with invented facts.

Search and read

Inspect available tools. Normally combine built-in WebSearch with mcp__search__search: complementary discovery queries can find different sources. Record which channels actually contributed. If a channel is unavailable, continue with the working one and note the coverage limitation; an unavailable search is not evidence that a source does not exist. With no search, work from supplied material and state which external claims remain unverified.

The project .mcp.json registers the keyless free-search-mcp server and a staged download directory. Tool signatures vary by installed version: read the live schema instead of assuming the argument names in an old example.

Capability	Appropriate use
search / research	Focused discovery / an open research question
cache_search	Reuse material already read
fetch / fetch_batch	Read shortlisted web pages
read_doc	Read PDF and structured or office documents, with pagination
extract_structured	Confirm publisher metadata such as DOI and title
compare	Reconcile conflicting sources or definitions
paper_graph	Explore prior work and check available correction/retraction notices
download	Retain a file needed for reproduction or a licensed reference
engines	Diagnose thin results or unavailable sources

Use the narrowest category the live tool exposes. Typical groups include paper (with index, preprint, biomed, cs, openaccess, trial and math subgroups), dataset (repository, ml, gov), news, finance, github, forum and image. Subgroup availability depends on the server. A general web domain filter is not equivalent to a bibliographic search.

Retrieve before writing: find and read a source before drafting the sentence that cites it. A reference recalled from memory is a lead to check, not a citation. Merge results by DOI or canonical URL. Read an accepted source once, using the appropriate reader. When sources disagree, cite both, prefer the official or primary record, and record why the adopted value fits the model.

Verify citations and retain evidence

Resolve every DOI or arXiv identifier with a tool, including papers you are sure of, and take title, authors, year and venue from the retrieved record rather than from recall. Cite only URLs that were actually retrieved. Cite a preprint as a preprint; look for the published version and prefer it when it exists. State a finding no more strongly than its study design allows: an association, a simulation or a single case does not establish a cause or a general law. When an identifier or detail cannot be confirmed, record the gap instead of guessing; an omitted DOI is better than an invented one. For a load-bearing paper, use paper_graph when available and inspect publisher notices. A lack of a returned notice does not prove that a paper has never been corrected or retracted; record the coverage limitation when that matters.

Before submitting a paper, and on any bibliography an agent wrote, run the optional reference checker. It uses only the Python standard library, reads .bib, Markdown and LaTeX files or bare DOIs, and writes nothing:

python3 .claude/skills/mathodology-evidence-search/scripts/verify_refs.py refs.bib paper.tex

It resolves identifiers as Mathodology Studio does: Crossref first, otherwise the registration agency named by doi.org, with metadata from DataCite or from doi.org for agencies such as mEDRA, JaLC, ISTIC and CNKI. It compares each entry's title, first author and year with the registered work, reports retraction, correction and withdrawal notices, flags preprints and suggests a DOI for entries without one. Exit status 1 means at least one failure; --json gives per-item results. A DOI that the DOI system reports as non-existent is probably mistyped or invented. A timeout or rate limit is reported as inconclusive and says nothing about the reference. A clean run shows that identifiers exist and match their records, not that a source supports the claim it is cited for.

Keep a readable source note: supported claim, URL/DOI, version or date, access date, discovery channel, extracted quantity, credibility, and uncertainty. Record queries briefly so another researcher can retrace discovery. No fixed YAML handoff or separate ledger file is required. Verification belongs in these notes; the deliverable cites its sources without narrating the checking ("all DOIs verified").

Numerical inputs need a retained data file or reproducible acquisition procedure. For retained downloads, record source URL, license and SHA-256. MCP downloads are staged and may expire; copy needed files to the working directory promptly and verify that the retained bytes match the recorded hash. Do not leave the only copy in staging. Do not redistribute material without permission to do so.

If download is unavailable, use a permitted existing file-transfer tool or give the user a source link and explain the gap. Do not reconfigure custom MCP settings or bypass access controls. Avoid collecting credentials in the conversation.

Figure references

Use figure presets for visual work. Its reference collection distinguishes originals, counterexamples and synthetic previews. For additional references, verify the individual asset's license and any third-party credit lines. Open access alone is not a reuse license.

Save a small, relevant selection with attribution, version, original URL, access date and hash. Prefer an immutable source revision for code examples. Keep copied source as reference text, review it before adaptation, and do not execute it automatically. Extract design principles; do not present another paper's image or numerical results as the current model's output.