Methodology

How the Supience Information Indicator Works

The Supience Information Indicator is designed to provide a structured way to examine observable characteristics of a piece of information before relying on it. It is not a truth detector, fact-checking database, source certification system, or probability that a statement is true.

Instead, the tool looks for signals that can make an information item easier or harder to investigate. The result is intended to help a person decide what deserves closer attention.

What the tool evaluates

The current Supience model considers four broad dimensions: evidence, specificity, uncertainty and verifiability. Each dimension represents a different property of the submitted text.

1. Evidence

Does the text point toward supporting material such as sources, studies, documents, records, named organizations or other evidence that a reader could investigate?

2. Specificity

Does the information contain concrete details such as names, dates, quantities, locations, conditions or clearly defined claims rather than relying primarily on vague statements?

3. Uncertainty

Does the wording acknowledge relevant limitations, conditions or uncertainty where appropriate, or does it present complicated claims with excessive certainty?

4. Verifiability

Does the submitted information contain clues that make its important claims easier for a reader to investigate independently?

Why these four dimensions?

Information quality is multidimensional. A paragraph can contain a citation but still be too vague to evaluate. Another statement may be highly specific but provide no way to determine where its numbers came from.

The four dimensions therefore complement one another. Supience does not treat one signal as sufficient by itself.

How the score is constructed

The calculator examines the submitted text sentence by sentence. It looks for observable patterns associated with evidence, specificity, uncertainty and verifiability, then combines the sentence-level signals using signal-density weighting: sentences containing more detectable signals contribute more than sentences containing none. The resulting four component scores are combined using fixed weights.

Current formula: Overall indicator = (Evidence × 0.28) + (Specificity × 0.22) + (Uncertainty × 0.20) + (Verifiability × 0.30). Each component is constrained to a 0–100 range and the final result is rounded to the nearest whole number.

Sentence signal density

Sentences are weighted by detectable signal density rather than treated as identical units. A sentence containing one or more evidence, verification, uncertainty, date, number, entity or related signals contributes more than a sentence with none. This is a weighting heuristic, not a measure of sentence importance or truth.

Evidence — 28%

Base value: 50. Source, study, research, data, report, citation and related signals can raise the component. Negated evidence phrases such as “no evidence” or “without data” reduce it. Absolute wording can also reduce the signal.

Specificity — 22%

Base value: 52. Numbers, years, dates, named entities, measurable units and comparison phrases add signal. Bare numbers receive less weight than numbers paired with units, comparisons or named entities. Vague wording such as “things” reduces it.

Uncertainty — 20%

Base value: 52. Useful hedges such as “may,” “depends” and “suggests” add more signal when they occur near claim/result terms. Weak conversational hedges such as “I guess,” “maybe” and “sort of” receive a stronger penalty. Strong certainty language can reduce it.

Verifiability — 30%

Base value: 48. Dates, sources, names, methods, links, citations, authority names and concrete measurements can add signal. Vague time phrases such as “recently” reduce it. Older detected years can receive a small age-related penalty when the text provides no explicit time context.

These are heuristic rules, not statistical estimates. The presence of a signal does not mean the underlying claim is correct. The calculator is deliberately transparent so users can inspect what it noticed.

Signal detection

The launch version of Supience is designed to operate from characteristics that can be examined directly in the submitted text. Depending on the implementation version, these may include patterns associated with:

These signals are useful because they can make an assertion more inspectable. They are not evidence that the assertion itself is accurate.

Date-relative reasoning

Supience compares detected years with the current year. If an older cited year is found and the sentence does not explicitly establish a historical context, verifiability may be reduced slightly. The result may also say that the oldest cited date is N years old and suggest checking whether the situation has changed. This does not mean older information is false or unusable.

What happens when a signal is missing?

The absence of a detectable signal does not automatically make information false. A short statement may simply lack enough context for the tool to identify a particular characteristic.

For example, a sentence such as “The meeting starts at 3 PM” contains a specific time, but without knowing which meeting, date or location, there may still be important context missing.

The indicator should therefore be interpreted alongside the original information and the circumstances in which it will be used.

Why the tool does not produce a truth percentage

A statement can score well on observable information characteristics and still be wrong. Conversely, a correct statement can be brief, poorly sourced or lacking in details that an automated text analysis can recognize.

For that reason, Supience does not present results such as “87% true” or “92% reliable.” Such numbers could easily be misunderstood as a statistical probability when the tool does not have the evidence required to make that claim.

Example of the distinction

Consider this statement:

“Company X reduced processing time by 35% during a six-month pilot conducted in 2025.”

The statement contains several concrete details: a company, a measurable result, a time period and a defined activity. Those characteristics can make the claim easier to investigate.

However, the details do not establish whether the 35% figure is accurate. A reader would still need to locate the underlying report, study, announcement or other appropriate evidence.

Interpretation bands

The displayed result is intended to help users decide how much attention to give the information's observable characteristics. Broad interpretation bands may be used to make the result easier to understand.

Stronger signals

The text contains several characteristics that make its claims relatively easier to inspect. Verification may still be appropriate depending on the consequences of being wrong.

Mixed signals

The text contains useful characteristics alongside areas that deserve additional inspection. The user should identify which important claims remain unsupported or unclear.

Weaker signals

The text provides fewer observable characteristics that support straightforward investigation. Additional context or independent evidence may be particularly useful.

Cases where the calculator is likely to be wrong

The calculator can recognize useful textual patterns without understanding the full meaning of a claim. That creates predictable failure cases.

Design principle: Supience tries to make its mistakes visible rather than hiding them behind a complex-looking score.

Important edge cases

Short statements

Very short text may not contain enough information for meaningful signal detection. A low or incomplete result should not be interpreted as evidence that the statement is false.

Opinions and recommendations

An opinion does not necessarily need the same evidence structure as a factual claim. For recommendations, users should consider the criteria, assumptions and circumstances behind the recommendation.

Creative writing

Stories, fiction, slogans and other creative material are not necessarily intended to make factual claims. The indicator is therefore not designed to judge creative merit or fictional accuracy.

Technical and scientific material

Technical claims can require domain-specific knowledge and primary literature. Observable text signals cannot substitute for qualified review of the underlying technical evidence.

AI-generated information

AI-generated text can be detailed, confident and well structured while still containing incorrect or unsupported claims. Supience can help identify characteristics worth examining, but it does not independently verify an AI system's output.

Optional semantic analysis

Supience can add a browser-based semantic layer to the deterministic calculator when the browser is online and supports WebAssembly. It is enabled automatically by default, but users can turn it off. The browser represents the submitted text as vectors and compares them with fixed prototype descriptions of factual-style claims, opinions, predictions, qualified claims, evidence-supported statements, and common medical, financial, product, news, and legal/regulatory claim types.

This layer does not generate answers, browse the web, verify sources or determine whether a claim is true. It is a similarity-based signal that can help the calculator distinguish different kinds of language even when they use different words.

Semantic influence: The semantic layer only tilts the deterministic result. Its Evidence adjustment is capped at ±5 points, so semantic similarity cannot define the score by itself.
Offline behavior: The standard calculator remains fully usable offline. Enhanced semantic analysis requires its model and runtime to be available in the browser. On first use it may need an internet connection to download them. If they cannot be loaded, Supience automatically uses the standard calculator.

Semantic limitations

Scoring behaviour and its limits

Several properties of the scoring model are worth stating explicitly. They follow from how the calculator is built, and they affect how a result should be read in different situations.

Short text stays near the base values

Each component starts from a base value — 50 for Evidence, 52 for Specificity, 52 for Uncertainty, 48 for Verifiability — and moves only as signals are detected. On short text, few signals are available, so the components stay near those base values and the overall score sits near the middle of the 0–100 range regardless of the quality of the information.

A middle-range score on a short passage usually means “not enough text to analyse,” not “moderately reliable.” The score bands describe the range of the number, not a fixed standard.

Absence of a signal is treated as a negative

If a text contains no evidence words, no dates, no numbers and no hedges, every component drops toward its base value and the overall score moves downward. This is reasonable for text that should contain such signals, but it also penalises correct, concise statements that do not need them.

A factual sentence such as “the meeting starts at 3 PM” may score low even though it contains no problem. The tool treats missing signals as missing evidence, not as evidence of absence.

Different fields require different signals

The same four dimensions are applied to every type of text. Medical claims, legal claims, financial claims, historical claims, product claims and personal recommendations all use the same weights. Different fields require different standards of evidence, and the calculator does not adjust for that.

A medical claim supported by a single small trial may score higher than a legal claim citing an official statute, because the surface signals happen to favour one over the other.

Numbers receive credit even without context

The calculator gives more weight to numbers paired with units, comparisons or named entities than to bare numbers. But even a bare number contributes some specificity signal. “42 people agreed” still includes the number 42, and the sentence still gains a small amount of credit.

The tool cannot judge whether a number is meaningful, decorative, cherry-picked or correctly applied. It can only detect that a number is present.

The opinion-share figure is a rough indicator

The “opinion/recommendation markers” figure is calculated by counting opinion-marking words and dividing by the number of sentences. It is a rough indicator, not a measurement.

A text with several short declarative sentences and one opinion marker can look more opinionated than it is. A text with a single long sentence containing many opinion markers can look less opinionated than it is. The figure is intended as a hint, not a verdict.

Weights and base values are not derived from a dataset

The current weights — 0.28, 0.22, 0.20, 0.30 — and the base values are heuristic. They reflect an editorial judgement about how the four dimensions should combine. They are not the output of training on a labelled dataset, and they are not calibrated against any external baseline of information quality.

The scoring is designed to be transparent and inspectable, not statistically optimal. A different choice of weights would produce different numbers and, in some cases, different band assignments.

Interaction model

Supience is a single-pass calculator. It does not hold a conversation, ask clarifying questions, or refine a result over several exchanges. Understanding what that means in practice helps set the right expectations.

No clarification dialogue

The tool analyses exactly what you submit. It does not ask follow-up questions, request additional context, or prompt you to enter more text when the input is short or ambiguous. If the submitted text does not contain the signals the tool looks for, the score will reflect that — it will not ask you to try again with something more specific.

This is a deliberate choice. The calculator is designed to give an immediate, single-shot response with a stable methodology, not a conversational assessment that changes based on what the tool decides to ask next.

No history across sessions

The tool stores only the most recent check on your device, using browser local storage. Opening the site in a new browser, a different device, or a private session gives you a clean state. Closing the browser tab does not clear the saved result, but clearing browser data does.

There is no account, no sync, no list of past checks, and no way to review more than one previous result at a time. If you want to keep a record of several checks, use the copy and share actions before submitting a new one.

Comparison uses only the most recent saved result

The Compare previous action compares the current check against the single most recently saved result on that device. If you have checked three passages in sequence, only the third is available for comparison.

To compare against a different previous check, save the current result elsewhere (using copy or the share link), then run the earlier comparison before submitting new text.

Share links contain the score, not the text

The share link contains the four component scores and the total, encoded as URL parameters. It does not contain the original text, the uploaded file, or any part of the analysis beyond the numbers.

This is intentional: sharing a result should never reveal the text that produced it. But it also means anyone opening a share link sees a score without the context that produced it. The share link is a reference to a result, not a way to reproduce it.

Batch mode is per-paragraph, not combined

Batch mode analyses each paragraph independently and shows one result per paragraph. There is no combined summary of the batch, no average-quality verdict, and no way to export the batch as a single document.

Each paragraph's individual result can be copied one at a time. If you need a consolidated record, copy each result as you go or capture them from the result panel before submitting a new batch.

Batch paragraphs cannot be revised individually

Editing a single paragraph in batch mode requires editing the entire textarea and re-submitting. There is no way to revise one paragraph in place and re-run only that one, and there is no way to remove a paragraph from a batch and re-run the rest.

The design reflects the calculator model. Each submission is a new analysis of the whole text you provide.

No persistent settings

The semantic-analysis toggle, batch-mode toggle and any other preferences are not remembered between sessions. They reset to their defaults on each visit. If you prefer a specific configuration, you will need to set it again each time.

File handling

The tool can extract text from a small set of file types, all processed locally in the browser. The extraction is intentionally simple, and each format has known limits.

PDF extraction is text-based only

PDF extraction works by reading the text layer embedded in the file. It handles most modern PDFs produced by word processors, layout tools and print-to-PDF functions. It does not handle:

Where the extraction fails, the tool reports an error rather than returning meaningless output. A PDF that extracts only part of its text will produce a shorter passage than the full document.

DOCX extraction handles body text only

DOCX extraction reads the main body text of the document. It does not extract or preserve:

The extracted text is the visible body content. Important content that lives outside the body — for example, a footnote containing the source of a statistic — may not appear in the analysis.

Plain-text formats extract directly

TXT, MD, Markdown, CSV, JSON, HTML, HTM and XML files are read as text. For HTML and XML, the extraction is simple: tags are not stripped, so the raw markup may contribute to the analysis. This can artificially increase the specificity signal (HTML tags contain numbers, dashes, etc.) and is not recommended where clean text is available.

Image OCR is not available in the offline build

The tool accepts image file types in the file picker, but the current offline build does not include optical character recognition. If you select an image, the tool returns a clear error and asks you to paste the text instead.

Only one file at a time

The file picker accepts a single file per selection. There is no way to upload several documents and analyse them together, and there is no way to combine the extracted text from more than one file.

File size is capped at 12,000 characters

Extracted text is truncated at 12,000 characters before analysis. Longer documents are analysed up to that limit, and the remainder is ignored. The file status line reports when truncation occurs.

Technical requirements

The standard calculator runs in the browser with no server dependency. Several technical conditions affect what is available.

Enhanced semantic analysis requires a model download

The optional semantic layer loads an embedding model from a third-party content delivery network on first use. Depending on the connection, the model may take 30 seconds to several minutes to download. The model file is around 30 MB in its quantised form.

Once the model has been downloaded and cached by the browser, subsequent uses do not require another download. Clearing browser cache will remove the model and require it to be downloaded again.

Semantic analysis requires WebAssembly

The semantic layer runs through WebAssembly. On browsers or environments where WebAssembly is unavailable or disabled, the semantic toggle has no effect and the tool falls back to the standard calculator. This is silent from the user's perspective — the tool works, it just does not apply semantic adjustment.

Service-worker caching delays updates

The site uses a service worker to cache pages for offline use. When a new version of the site is deployed, returning visitors may see the previous version until the service worker updates in the background. To force an immediate update, clear the site's cache or open the site in a private window.

Client-side-only architecture limits diagnostics

Because the tool runs entirely in the browser, there is no server-side logging of unusual behaviour. Errors that occur during a check are visible only to the user in the browser console. If you encounter a problem, the fastest way to report it is through the contact form.

No A/B testing or remote configuration

The scoring rules, base values and weights are fixed in the client code. There is no mechanism to test alternative rules against live traffic, and no way to update scoring without deploying a new version of the site. This is a consequence of the client-side design, not a limitation that the tool intends to remove.

Known limitations

Supience is a self-contained text analyser. It has no external connection and no knowledge base. The following limitations follow directly from that design.

It cannot verify anything outside the submitted text

It has no knowledge base of its own

Its date awareness is limited to extraction

It does not resolve named entities

It does not detect stance or perspective

It does not handle quoted material separately

General limitations of any text-based indicator

Human judgment remains part of the process

Supience deliberately keeps the human decision-maker in the loop. The tool can organize observable signals, but the user still has to determine what matters, which sources are appropriate and how much verification is justified.

For important claims, users should move from the indicator to the underlying evidence. That may mean checking primary documents, official records, original research, qualified professionals or multiple independent sources.

Privacy-conscious design

The core Supience calculator is designed to perform its analysis in the browser rather than requiring every submitted passage to be sent to an external artificial-intelligence service. This reduces the need for a server-side analysis system for the basic tool.

The site's actual data practices may change if analytics, advertising, forms or other third-party services are introduced. The Privacy Policy should therefore be consulted for the site's current practices.

Methodology updates

The scoring model may evolve as the tool is tested and improved. Changes should be documented so that the indicator remains understandable rather than becoming an opaque score.

A methodology change can affect how the same text is evaluated at different points in time. Results should therefore be treated as outputs from the current version of the tool, not permanent facts about the submitted information.

The right way to use the indicator

Think of the Supience result as a signal for investigation. It can help answer:

It should not be used to conclude that a person, company, article, AI response or website is truthful or untruthful solely because of its score.

Bottom line: Supience measures information-quality signals that can be observed in submitted text. It does not measure truth itself.

Try the Supience Information Check

Methodology questions

What does the Supience score measure?

The score summarizes four observable text signals: evidence, specificity, uncertainty and verifiability. It is designed to organize a preliminary review, not to estimate the probability that a claim is true.

Why does Supience use four dimensions?

The four dimensions cover different parts of an information review. Evidence asks what support is visible, specificity looks for concrete detail, uncertainty considers qualifications, and verifiability asks whether a reader has something practical to investigate.

Why is the score not a truth percentage?

A text can contain strong evidence signals and still contain an incorrect claim. Conversely, a short statement can have few detectable signals without being false. A truth percentage would imply a level of factual verification that this browser-based method does not perform.