Kanshiki takes a health claim and tells you how strong the published research on it actually is, and which way that research points. Kantei takes a single study and appraises it. Both show their working: every study is listed, every score is computed by rules you can read.
What it does
Kanshiki (้่ญ, forensic examination) takes a claim like "turmeric reduces arthritis pain", searches PubMed, sorts what it finds by study design, and produces two separate judgements.
- Strength grades how good the research on the topic is: systematic reviews and meta-analyses count for more than randomized trials, which count for more than observational studies, which count for more than case reports. Recency, sample size, and replication all factor in. This is arithmetic, not opinion, and the thresholds are published in the response.
- Direction grades which way that research points, weighted by the same quality tiers so one meta-analysis outvotes five case reports.
They are deliberately separate, because the interesting cases need both. "Ivermectin treats COVID-19" has strong evidence that contradicts it. A claim can be thoroughly studied and wrong.
Kantei (้ๅฎ, appraisal) looks at one paper: a plain summary, citation and venue statistics, author track record, funding disclosure, and retraction status. Paste a PubMed link, a PMID, or a DOI, or click any study in a claim result.
Two ways to run it
These are not two tiers of the same product. Self-hosting is the real thing. The public demo exists so you can try it before deciding.
| Self-hosted | Public demo | |
|---|---|---|
| Cost | Free, forever | Free |
| Rate limits | None | Yes, a few checks per minute |
| Where the model runs | Your machine, via Ollama | Cloudflare Workers AI |
| Where your claims go | Nowhere. They never leave your computer. | Cloudflare, and PubMed for the search |
| Setup | One command, plus a model download | None |
| Speed | Depends on your hardware | Around 10 seconds per claim |
If you are doing real work, self-host. A literature review means many claims and many studies, and a shared demo cannot support that for everyone at once. Running it yourself removes the limits entirely and keeps what you are researching private, which matters more than it sounds when the subject is health.
Accuracy is not the trade-off. On the bundled evaluation, a local
llama3.1:8b scores exactly the same as the hosted model.
git clone https://github.com/kitakitaaura/kanshiki.git && cd kanshiki
./setup.shSource and issues: github.com/kitakitaaura/kanshiki
The script checks what you have, installs nothing behind your back, pulls the model if it is missing, and starts the app. The model download is a few gigabytes and happens once.
Data sources
- PubMed, via NCBI E-utilities. Every study, abstract, publication type, funding statement, and retraction flag comes from here.
- OpenAlex. Citation counts, venue statistics, and author records in the study inspector.
- Semantic Scholar, used only when OpenAlex has no record.
If you build on this, cite those sources directly. Kanshiki arranges their data; it does not produce it.
Where AI is used, and where it is not
Most of what Kanshiki does is deterministic. The same studies always produce the same grade.
| Step | How it works |
|---|---|
| Evidence strength grade | Fixed rules. No AI. |
| Study design classification | PubMed publication types. No AI. |
| Timeline, funding mix, spotlight selection | Fixed rules. No AI. |
| Credibility signals and retraction override | Fixed rules. No AI. |
| Citations in every format | Pure formatting. No AI. |
| Turning your claim into a search query | Language model |
| Judging which way each study points | Language model |
| The written summary | Language model |
You can switch the model off entirely with the AI toggle in the header. Grades, timelines, funding, retractions, credibility scores, and citations all keep working. You lose the direction badge and the written summaries, and the search query is built by keyword matching instead.
Plenty of researchers would rather no language model went near their conclusions. That is a reasonable position: model judgements are not reproducible the way a fixed scoring rule is, they can be confidently wrong, and auditing them means re-reading every abstract yourself. The toggle exists so you can keep the deterministic half and leave the rest.
What it cannot tell you
It reads abstracts, not papers
Direction judgements and study summaries come from titles and abstracts. Full text is not fetched: NCBI does not serve it and most papers are paywalled. An abstract can omit or understate what the full paper shows.
The model judging direction is fallible
On the bundled evaluation, the current models get 7 of 8 studies right. That is useful, and it is not the same as correct. Per-study judgements are shown next to each study precisely so you can check them. If one looks wrong, it may well be wrong.
Citation and venue metrics measure attention, not truth
A heavily cited paper in a prestigious journal is a paper other researchers noticed. That is a signal about standing, not about whether the finding replicates. Kantei's credibility score says so on every response.
Retraction is the one hard override
A retracted paper is excluded from the evidence grade entirely, whatever its other signals look like, and is capped near the floor in the study inspector. It is still shown and still labelled, because knowing a retracted paper exists on a topic is useful.
The search shapes the answer
A grade only reflects the studies the query returned. A different query returns different studies. That is why the query is shown on every result and why you can edit it and re-run.
Not medical advice
Kanshiki grades published evidence. It does not grade your situation. It does not know your history, your medications, your risk factors, or anything else about you, and it is not designed to. A claim that is well supported in the literature can still be wrong for a particular person, and a claim with thin evidence is not thereby false.
Nothing here is a diagnosis, a treatment recommendation, or a reason to start or stop anything. Read the studies, then talk to a clinician who knows your case.