Find out what the evidence actually says
Bring a claim you are unsure about. Epistry gathers the real research and reporting on it, weighs each source by how good the evidence is, and returns a verdict the arithmetic produced — with the strongest argument against it, and an honest account of where the answer is weakest.
Free to start — 5 analyses every month, no card required.
How it works
Nothing here is a black box. The same five things happen to every claim, and the report shows you the result of each one.
You ask in plain words
A claim you have heard, a question you are unsure about, or a link to an article. Epistry finds the testable claim inside it and shows you what it is about to check before it runs.
It writes down what would change its mind
Before any evidence is gathered — so the test cannot be written to fit an answer that already exists. Each condition is checked against what was actually found, and the result is on the report.
It fetches real sources
Academic databases, news archives and the open web: real documents with real citation records, not references recalled from memory. A source whose link is not in the actual search results is discarded.
Every source is weighed, not counted
A meta-analysis outranks an anecdote, and each piece is then discounted for weak replication, age, small samples, retraction and the GRADE downgrades. Ten outlets reprinting one wire story count as about one voice, not ten.
Arithmetic decides the verdict
The weighted balance of evidence sets the direction. The number of genuinely independent sources sets the confidence. Same evidence in, same verdict out — every time, for everyone.
Language models read the evidence here. They do not get a vote on the answer.
They find and summarise sources, and argue both sides of the case for you to read. But the verdict itself is computed — a published formula over the weighed evidence, with no model in the arithmetic. That is why the same question asked twice gives the same answer, and why you can check the work rather than take it on faith.
How to ask
No syntax, no operators. Write the claim as you heard it, or the question as you would say it out loud — Epistry works out what is testable and confirms the wording with you before spending anything.
Claim to fact-check
A yes / no question
Open research question
The more specific, the better the answer. “Is seed oil bad for you?” is a topic; “Seed oils raise inflammatory markers in healthy adults” is a claim that evidence can settle. A question with no testable claim in it gets a survey of the landscape instead of a verdict — which is the honest answer to that kind of question.
Reading a report
A report is long because the work is shown. These are the parts that answer the question you came with.
The verdict
Likely true, contested, or not enough evidence — always with the confidence stated, never bare certainty.
The competing case
The strongest argument against, and whether it held up. If opposition was searched for and none was found, it says that too.
Where it is weakest
The single largest vulnerability in the verdict, in one sentence, at the top rather than buried at the bottom.
Every source
What was found, what it was worth to the verdict, and why — each title links out to the original.
The stakes
Falsehood travels six times faster than truth on the same platform. That is a measured finding, not a lament — and it is the reason a claim needs to arrive with its evidence attached.
The field guide
The field guide covers the scientific method, calibrated skepticism, reading a research paper, spotting the arguments designed to bypass all three — and the terms this site uses, defined once. None of it needs an account.
The open corpus
Every public analysis joins a corpus anyone can browse — the claim, its verdict, the sources it rests on, and the papers behind them. Reading it is free and needs no account; the meter only ever sits on running something new.
The price of diesel fuel in the US in 2026 is dominated by the interruption of supply in the Middle East and not by Ukrainian strikes on Russian production.
Google is scanning it’s user’s data to train AI models, and you can stop this by disabling “smart features” in the google account settings.
LG TVs are recording audio and sending it back illegally
Heavy rainfall increases the probability of earthquakes occurring in California.
The Shroud of Turin is an image of a person named Jesus.
One California-grown almond requires water equivalent to the amount needed to process 38,000 average LLM queries.
Is the Dunning-Kruger effect cited as an explanation for increased polarization in online political discourse since 2016?
LLM AI is more effective for severe dyslexia than traditional non-AI spellchecking.
Character building in role-playing games effectively facilitates therapeutic self-awareness.
What does the evidence say about intermittent fasting?
Is there a measurable psychological benefit to being able to regularly see the horizon, vs living in an urban setting with occluded sightlines.
The proportion of cash businesses in UK town centres that are suspected money-laundering operations increased between 2015 and 2024.
Regular engagement with generative AI tools leads to measurable gains in analytical reasoning compared to non-users.
Protein and creatine consumption is effective at combating the effects of inattentive ADHD,
The ocean has absorbed 90% of the heat we created. It's boiling, acidifying, and dying. No ocean, no us.
Turmeric consumption helps alleviate symptoms or progression of pre-existing arthritis
To what degree is Generative AI causing a loss of work or earnings in the creative industries?
Pesticides contribute to the development of ALS (amyotrophic lateral sclerosis / Lou Gehrig's disease.
Health effects of microplastics?
tylenol causes autism