Show the working.

Your input is first turned into a neutral research question and a literature search query written in the vocabulary a biomedical index actually uses. A second query is generated from a different angle — a different outcome, population or mechanism — so that research pointing the other way has a route into the results.

Both queries are then run against Europe PMC in several passes: one restricted to systematic reviews and meta-analyses, one restricted to trials, one unrestricted, and one sorted by publication date. Relevant registered trials are pulled from ClinicalTrials.gov. Crossref is used to fill in citation counts and funder disclosures that Europe PMC does not carry.

Europe PMC — the primary index, covering PubMed/MEDLINE records, PMC full texts, preprints and patents, with abstracts and publication metadata.

Crossref — DOI-level publication metadata, citation counts and funder registry entries.

ClinicalTrials.gov — registered interventional trials, used as context rather than as published findings.

Minkey does not scrape Google Scholar, search engine result pages, or paywalled full texts.

Candidates are scored on study design, topical overlap with the search terms, recency, citation count, reported sample size and whether the work was done in humans. Papers that share almost no vocabulary with the question are dropped, and preprints and study protocols are pushed down because they have not been peer reviewed or have no results yet.

Selection then applies quotas by evidence type, so that reviews, trials, observational work and mechanistic studies all get slots. This is deliberate: taking the top scores alone would produce a list of reviews that agree with each other.

A well-conducted systematic review or meta-analysis pools many trials and is harder to fool than any single study. Randomised controlled trials can establish that an intervention caused an effect. Cohort and cross-sectional studies can only show association, which may reflect something else entirely. Animal and laboratory work explains mechanism — it does not establish a human effect.

That hierarchy is a starting point, not a rule. A large, careful cohort study can be more informative than a tiny, poorly controlled trial, and Minkey shows you the design of every source so you can weigh it yourself.

Evidence posture describes what the retrieved research says about the question. Evidence strength describes how good that body of research is, regardless of direction. Research confidence is the model's confidence that the summary fairly represents the sources it was given.

None of the three is a probability that a claim is true. Minkey deliberately avoids a single truth score, because a number like that hides exactly the information that matters.

A language model does three jobs: it converts your input into a research question and search queries, it reads the retrieved abstracts, and it writes the synthesis. It never chooses which papers exist, never adds a citation of its own, and never sees the internet.

Every source shown to you was retrieved by the search layer before the model was called. The model refers to sources only by internal identifiers, and the server maps those identifiers back to verified metadata afterwards. Any identifier that does not correspond to a retrieved source is discarded rather than displayed.

The structural fields — study design, sample size, subjects, follow-up, funding, disclosures — are extracted from the actual record and its abstract by deterministic code, not written by the model. Where a field cannot be established, Minkey shows “Not reported” instead of guessing.

It cannot judge whether a study was conducted honestly, whether its statistics were applied correctly, or whether unpublished trials would change the picture.

It cannot read most full texts — abstracts omit a great deal, including many limitations authors describe in the paper itself.

It cannot know how a finding applies to you specifically. Population, dose, duration and baseline health all matter, and none of them are personal to the reader.

Language models can also state things fluently and confidently that are wrong. That is precisely why every claim here is tied to a source you can open.

Most interesting health questions do not have a settled answer. Compressing them into one anyway is how research gets distorted in the first place.

Minkey is built to say “mixed”, “limited” and “unclear” when that is the honest reading, and to show you the disagreement rather than resolve it artificially.

Minkey is not peer reviewed by humans, and it has no relationship with, or endorsement from, any scientific institution, journal or database provider. It reads their public APIs like any other reader.

It is an educational research tool. It does not diagnose, prescribe, or replace a clinician who knows your history.