Evidence
Does the text point toward sources, studies, data, documents or named organizations that a reader could investigate?
This guide explains how to use the tool, how to read its result, and how to move from the indicator to the underlying evidence. It is written for readers who want a practical starting point — not a technical description of the scoring model, which is covered in the Methodology.
The Supience Information Check is designed for one situation: you have a paragraph of information — an AI answer, an article, a recommendation, a business claim — and you want to know what to look at more carefully before relying on it.
The tool does not tell you whether the information is true. It looks at observable characteristics of the text that can make a claim easier or harder to investigate.
The indicator is built from four dimensions. Each one asks a different question about the text.
Does the text point toward sources, studies, data, documents or named organizations that a reader could investigate?
Does the text contain concrete details — names, dates, quantities and conditions — or does it rely on vague statements?
Does the wording acknowledge relevant limits and conditions, or does it present complicated claims with excessive certainty?
Does the text give a reader something specific to check independently?
The tool is designed to perform its core analysis in your browser. Its purpose is to provide an immediate assessment of observable characteristics in the text you submit.
The indicator is currently optimized for English-language text. Results on text in other languages may be less reliable.
The result shows four component scores and an overall indicator. The overall indicator is a summary, but the four components usually tell you more than the total.
For example, a paragraph might score relatively well on evidence because it mentions a study, but poorly on specificity because it does not say what the study measured, when it was done or what the result actually was. In that case, the evidence signal is present, but the claim may still be difficult to check.
Two properties of the scoring model matter when reading a result. First, the components are anchored to base values near the middle of the range, so short passages and passages with few detectable signals will score near 50 in each component — that is a signal of missing information, not a verdict on the text. Second, the same weights are applied to every subject, so a medical claim and a historical claim are judged by the same surface features even though they require different standards of evidence.
The score bands describe the range the number falls into. They do not represent fixed thresholds of reliability.
A high indicator does not mean the information is true. Text can contain citations, dates and specific numbers and still be wrong — for example, if the citation is inaccurate, the date is misleading or the numbers are presented without important context.
The indicator measures signals that make a claim easier to investigate. It does not measure whether the claim survives that investigation.
It also does not judge whether the signals are meaningful. Numbers receive credit even when they are decorative, cherry-picked or applied to the wrong population. A passage full of precise-looking figures can score well without any of them being relevant to the claim.
A low indicator does not mean the information is false. Short, informal or conversational text may simply not contain the observable signals the model is designed to recognize.
For example, a short factual statement such as "The library closes at 6 PM" may not contain enough text for meaningful signal detection. That does not make it wrong.
A low score can also reflect a limitation of the tool rather than a problem with the information. The tool cannot open links, resolve named authorities or compare the text against any external source. If the text is quoted from a source the tool cannot access, or is in a language the tool is not optimized for, the score may be lower than the information deserves.
The tool is most useful when:
AI-generated answers can be fluent and specific without necessarily providing enough information for a reader to independently verify every important claim.
When checking an AI answer, pay particular attention to important factual claims, dates, numbers, citations, named studies, recommendations and statements that could affect a decision.
For a more detailed AI-specific process, see How to Verify an AI Answer.
Do not use the indicator as the basis for a high-stakes decision. Medical, legal, financial, safety and security questions require authoritative sources and, where appropriate, qualified professional advice.
The tool should also not be used to make claims about the honesty of a person, the reliability of a company or the accuracy of a specific website based solely on a score.
You can use the tool as part of a short verification routine:
For a longer version of this process, see the Information Verification Checklist.
Paste the answer, or upload a supported text-based document, and run the Information Check. Review the four component signals and then verify the most important claims against appropriate sources.
A lower indicator means the supplied text contains fewer of the observable signals the tool looks for. It does not mean the information is false; it means further checking may be useful.
A higher indicator means the text contains more observable signals associated with evidence, specificity, uncertainty and verifiability. It is not proof that the underlying claims are correct.
The standard Supience calculator works offline using transparent rules. An optional enhanced mode can add browser-based semantic similarity analysis. The model runs locally in the browser and does not turn Supience into a chatbot or fact-checking service.
Because the model may need to be downloaded the first time, enhanced mode is not guaranteed to work without an internet connection. If it cannot load, the normal calculator continues to work.
The standard calculator runs entirely in your browser with no server dependency. A few technical conditions affect what is available:
The site is also intentionally small. There are no user accounts, no public API, no paid tier, no native apps and no browser extension. Supience is a single web tool, accessed through this site.
Supience is a single-pass calculator. It returns a result for the text you submit and does not hold a conversation. Three practical consequences of that model are worth knowing.
The tool does not ask for more detail, request a source, or prompt you to clarify a claim. It analyses exactly what you paste and returns a result based on that. If the text is short or contains few detectable signals, the result reflects that — the tool will not ask you to try again with something longer.
Your last check is stored on this device using browser local storage. Opening the site in a different browser, a different device, or a private session gives you a clean slate. There is no account, no sync, and no list of past checks.
The Compare previous action compares the current submission against the single saved result. If you have checked several passages in sequence, only the most recent one is available for comparison.
When you share a result, the link contains the four component scores and the total, encoded in the URL. It does not contain the text you submitted or the uploaded file. This protects privacy, but it also means the shared link shows a score without the context that produced it.
For batch checks, each paragraph can be copied individually, but there is no combined export. If you need a record of several checks, copy or share the result before submitting new text.
A practical process for checking whether a source actually supports a claim.
Read the guide →How to check whether a citation actually backs the statement it appears to support.
Read the guide →Reading percentages, comparisons and averages critically.
Read the guide →Why apparent contradictions in research are normal, and what they mean.
Read the guide →What an abstract tells you, and what it leaves out.
Read the guide →Common patterns of reasoning that feel persuasive without supporting a claim.
Read the guide →