mizuraGet Mizura

Development snapshot · Reviewed 2026-11-08

Development docs
Use this page

Evidence & literature

Two checks matter to a reader. Were the numbers actually computed? Was the cited literature actually read? Mizura checks both, with limits described below; a passing check is evidence about the recorded output, not proof that the analysis or interpretation is correct.

Number grounding: every statistic traces to executed code

The checker looks for numbers with three or more significant digits that are absent from tool results in this session. It accepts the same value rounded or truncated to the answer's precision, or a percentage derived from a produced value. Numbers in your question, standard normal quantiles, and integer parameters the agent chose are excluded.

Some numbers do not count as results. Numbers in your own messages. The standard normal quantiles (1.96, 2.576, 1.645, 1.282). Integers the agent chose as parameters, a seed or a grid size. Hashes.

When a number has no source

  1. The answer goes back. Run code that computes and prints each one, then write the reply again using only printed values. If a number is not needed, leave it out.
  2. The agent runs the code, prints the values, rewrites the answer. At most twice.
  3. After that the reply can stand with a warning flag. The check does not establish that the computation or interpretation is sound; do not treat a flagged answer as verified.

Reports face a higher bar. One ungrounded number rejects the whole document, with the numbers named. The agent fixes them and resubmits. See Reports & exports.

Warning. Grounding proves a number was printed by executed code. It does not prove the code was right, the method was appropriate, or the comparison means anything. A grounded number can still be a wrong conclusion. Read the diagnostics in the run.

Mizura searches through its own search service. If that service is down, scholarly and package-documentation search still work. Only the general web search is lost, and the workspace warns you under the composer. Until it recovers, answers rely on the model's own knowledge.

Search intentSources
LiteratureOpenAlex, Semantic Scholar, PubMed, Crossref, CORE (open-access full text), Unpaywall, arXiv for preprints.
DocumentationOfficial package documentation, including CRAN.
Edge casesPackage documentation and Stack Exchange, for the how-do-I-do-this corner.
GeneralAll of the above, plus a web search.

Economics questions also reach grey literature: NBER, SSRN, IMF, World Bank material. Fetching tries a direct request first, then a browser. PDFs get per-page text recognition when the font map is broken.

The privacy screen

Every search query and fetched page is checked against identifier-like terms from your data before anything is sent. Column names. File names. A plain word like age passes. A distinctive identifier does not. The search does not run, and nothing is sent. URLs you add as sources get the same check.

What is in your dataset stays out of your searches.

Follow a paper's citations

Once you have a paper's DOI, ask Mizura to follow newer work that cites it, or the references it cites. A topic phrase can narrow the results. This is different from searching for an exact title: the citation lookup starts from a known paper.

For the paper with DOI [insert DOI], find later studies that cite it, focusing on replication or critiques of its method. Open the relevant sources before summarizing. Distinguish full text from abstracts and list anything you could not retrieve.

For backward tracing, ask for the paper's references instead. Returned links are candidates until opened. A missing result or failed lookup does not establish that no follow-up research exists.

Citation discipline

Search results are candidates, not sufficient evidence by themselves. A report reference must have been retrieved and opened in this session, either through search and fetch or as a source you added. The report checker enforces specific minimum text lengths and citation-key matches; it cannot prove that a paraphrase is faithful.

The checks and their minimum text requirements are listed below. For browsing, filtering, sorting, and downloading sources, see Find and manage outputs.

  • Every citation needs a reference. Every reference needs two things: retrieved this session (in a search result, or added as a source) and opened (fetched as a page, or added as a source). A search hit alone fails.
  • An opened page must be substantive, 120 words or more. Or a record with an abstract, 30 words or more. The title must match. The author's surname and year must appear on the opened page.
  • Quotes must be verbatim in the opened text. A quote the checker cannot find is a rejection. Not a footnote.
  • Method cards are not literature. They cannot be cited as if they were.

Chat is softer. The agent states only what the text it opened says, and abstract-only sources support abstract-level claims only. Want a citation checked? Ask for the report. That is where verification happens.

Sources tab and sources file

Everything the agent retrieves, and everything you add, appears in the Sources tab. Papers, search results, fetched pages, your files. Searchable, sortable, downloadable as a Markdown source list. Removing a manual source removes it from the session context. It does not delete your original file.