鑑識

Kanshiki

Forensic examination of health claims.

AI is off. Evidence grades, timelines, funding, retractions, and citations still work. Direction and written summaries do not.

Paste a health claim, headline, or passage

Edit the PubMed query

Leave this empty and Kanshiki builds the query for you. Fill it in and your text goes to PubMed exactly as typed, including field tags and boolean operators, with no rewriting.

Try:

Paste a PubMed link, a PMID, or a DOI

Try:

Saved on this device only. Nothing is uploaded and there is no account. Export the collection to a file to move it to another browser or machine.

Export

Nothing saved yet. Use "Add to collection" on a claim result or an inspected study.

Claim examined:

PubMed query: ·

Two separate judgements. The first badge grades the strength of the research that exists: study designs, how many, how recent. The second grades which way that research points, weighted by the same study quality. A claim can be well studied and wrong.

The evidence is split. Here is the strongest study on each side.

Strongest evidence for

Strongest evidence against

How the evidence changed

Export

Kanshiki grades published evidence, not your situation. It is not medical advice and it does not tell you what to do. Read the sources, then talk to a clinician.

This paper has been retracted.

What this paper says

Credibility signals

Stats

Cite this

Export

These are credibility signals: retraction status, study design, citation traction, venue, author record, funding disclosure. They do not tell you whether the paper's findings are correct. Read the paper on PubMed.

This is a public demo and is rate limited. For unlimited use, run Kanshiki on your own machine for free. It is one command, your data never leaves your computer, and there are no limits. How to self-host

About and methodology · API · Sources from PubMed via NCBI E-utilities. Evidence grade is computed by a fixed hierarchy; the written summary is model-generated.

Turning AI off

Plenty of researchers would rather no language model went near their conclusions. That is a reasonable position, not a superstition: a model's judgements are not reproducible the way a fixed scoring rule is, they can be confidently wrong, and checking them means re-reading every abstract yourself. This switch exists so you can keep the parts of Kanshiki that are deterministic and auditable, and leave the rest.

Nothing here is hidden from you either way: the evidence grade never used AI, and every model-driven figure is labelled as one. But two things Kanshiki does are model-driven and will stop, so this is a real change in what you get back, not a cosmetic setting.

What you lose

Direction of the evidence. Whether studies support or contradict a claim is judged per study by the model. With it off, the direction badge reads "Direction not assessed" and the contradiction spotlight never appears.
Plain-English summaries. A claim gets a sentence assembled from study metadata instead of a written verdict. A single study shows its raw abstract rather than an explanation of it.
Search quality. Your claim is turned into a PubMed query by keyword stripping rather than by the model, so the search is broader and cruder, and you may be graded on a different, less relevant set of studies.

What still works, unchanged

The evidence strength grade, its quality tiers, gates, and score. None of it ever used AI.
The evidence-over-time timeline.
Retraction detection, and the funding breakdown.
Study credibility scoring, citation and venue statistics, and all four citation formats.

It is also considerably faster: roughly 2 seconds for a claim instead of 20.