← Back to the methodology Go to the source check →

Source check method

General

Version 0.2.2not yet reviewed by a subject expert

A check profile fixes how the source check reads a source. The check profile General applies to every document with fixed wording that no profile of its own covers: a study, a statistical release, a report, a press release of a public authority, a programme. This page describes its method; every change to it gets a new version.

Checking is German text against German sources. The explanations in the findings are therefore always in German.

The statements, the Prüfraster and every finding with its explanation are produced by a language model. No person reviews the individual findings before they are shown. What is produced without a language model is set out in section 10.

How often you can run a source check is stated on the tool's page.

1. What is checked

The source check holds each statement of a text against the documents it refers to. Under the check profile General, the yardstick is what the document says and what it states about the reach of its own statements. Every finding has up to two lines; which one counts depends on the kind of statement (section 3):

Reproduction: Does the text reproduce the document correctly, its figures included? The values and what they mean here are set out in section 4.
Reading: Does the document carry the meaning or application the text draws from it? Covered, not covered, or open, defensible, each with the criterion that decides it (section 5).

What the source check is not

It is not a fact check. The measure is the document, never the world. Whether a study holds or a statistic depicts the situation correctly is not examined.
A claim about the world stays an evaluation, even when a selected document reports exactly that. “Older people are ill more often than younger ones”, said without ascribing it to a document, gets no finding. Only what is ascribed to a document or read from it is checked.
No recalculation. The source check recalculates no figure, does not hold a document's figures against other data and does not judge its method.
No advice on individual cases. What follows from a document for a particular person or case is listed and not checked.

2. How the check profile is chosen

The check profile applies per source, not per text. When a source is prepared for the first time, a check without a language model establishes whether it is structured like a law, with at least two sections (§) or one article. Then the check profile Legal text applies, otherwise General. The list of sources shows for every source what it is checked as, and whether that was detected automatically or changed.

The check profile can be changed in the list of sources and from a result. The source is then prepared again without a language model, and the selection of its sections starts afresh. Only a document structured by sections (§) or articles can be switched to the check profile Legal text.

If a statement needs sources of both check profiles, it is checked under the check profile General, together with the selected sections of every source it needs. Its criteria hold for any document with fixed wording, a law included.

The check profile Legal text reads a law the way lawyers read it, with legislative history and purpose as further criteria. General does not have these criteria. Instead it draws a Prüfraster for every source (section 6) and checks the figures of a finding without a language model (section 8). Method of the check profile Legal text

3. Statements and their kinds

The text is split into single statements. The splitting is the same for both check profiles, because it happens before any source is prepared. Which passages are statements, and of which kind, is decided by the language model. Every statement must stand word for word in the text: a passage the language model names that cannot be found there is dropped. The kind decides how a statement is checked:

About the source: What a document says, reports or contains, including that something is not in it, for example “According to the study, 52 per cent of respondents report …”. Checked against the wording and sense of the passage; the reproduction line is what counts.
Application: For whom, what, where or when a document's statements hold, for example “The figures hold for all older people in Germany”. Measured against the scope the document states for its statements, which the Prüfraster makes explicit; the reading line is what counts.
Interpretation: What a document means or what follows from it, for example “The study shows that exercise protects against falls”. Measured against the criteria of the check profile (section 5); the reading line is what counts.
Evaluation: A judgement or a claim about the world that is not ascribed to any document. Listed without a finding, even when a selected document reports the same.
Individual case: Applying something to a particular person or case. Listed without a finding.

A statement classified as an evaluation or an individual case is not checked. The check itself can still classify a statement as an evaluation when it claims something about the world without ascribing it to a document; it then appears in the result with that notice and without a finding. If a statement needs a section that is not selected, its result is “not checked”, and the missing section is named. A finding rests only on selected sections.

4. Reproduction

The values of the reproduction line are the same as under the check profile Legal text. For a document that is not a law, they mean:

word for word: The statement quotes the document, and every quote stands like that in the selection. A figure the text repeats in a sentence of its own is not a quote: a figure repeated correctly is reproduced in substance, and the explanation says that the statement reproduces the figure unchanged.
in substance: In other words, the same content. For a figure: the same value, or a rounding the statement marks as one (“around”, “about”), with the same unit, base, period and population.
shortened: Something the document attaches to the statement, and that matters for it, is missing: a condition, an exception, the population, the region or the period, the provisional status, the uncertainty the document itself states (a range, “not significant”), or the attribution. The finding names what is missing, and the point of the Prüfraster for it where there is one.
contradicts the wording: The wording says something else: another value, another unit (per cent instead of percentage points), the other direction, another population or period, or the opposite.
not found in the source text: What the statement ascribes to the document stands in none of the selected sections. For quotes this is settled without a language model (section 8).

If a document exists in several versions, for example as a provisional and a final release, a statement made before the check without a language model says which version holds the quotes of the text. This statement stands apart from the finding.

5. Criteria

A reading is measured against three criteria. All three measure against the document itself:

Wording: What the document says at this passage, with its conditions and limitations. A reading by wording rests on the text of the document, not only on a list such as a table of contents.
Context in the document: What the document says about it elsewhere: its definitions, units, conditions and limitations, the population, notes to tables.
Stated scope and limits: What the document states about the reach of its statements: population, region, period, conditions, limits. A reading by this criterion names at least one point of the Prüfraster (section 6).

Legislative history and purpose, two criteria of the check profile Legal text, are not among them: a study has no explanatory memorandum, and what a document is for is rarely stated in it.

How a criterion decides the reading

The language model classifies the reading. These are its instructions; where the evidence check verifies part of them, that is said too.

covered: The reading follows from the wording in its context in the document and stays within the scope the document states for its statements. The finding names the criterion that shows this and the passages for it; without a criterion it is discarded.
not covered: Only when a criterion speaks explicitly against the reading: a limit or condition the document states rules it out (“No causal conclusions can be drawn from this”), the population is narrower, or the context says the opposite. The language model's own view of the field is never meant to be enough.
open, defensible: The document allows more than one defensible reading. The strongest support in the text for each reading is to be shown as a passage of its own, without the language model deciding between them. The evidence check requires at least two different readings with two different passages.

The language model measures a statement about application against the scope the document states for its statements, which the Prüfraster makes explicit. If the statement passes over a limitation, the finding is to name it, together with the point of the Prüfraster that records it.

Reported positions

In every field it happens that a position a document only reports (“critics complain …”, “according to the association …”) is presented as the document's own. The reproduction is then shortened, because the attribution the wording carries is missing. An interpretation built on such a position is not covered by the context. The Prüfraster lists such positions as a kind of their own.

6. Prüfraster

The check profile General is fixed and versioned like Legal text. It adapts to a single document only through the Prüfraster: the points the document sets for itself. The method stays the same; the Prüfraster shows what is measured against. A point has one of seven kinds:

Definition: How the document defines or delimits a term it uses, for example “People aged 80 and over count as very old”.
Unit and base: Unit, base and way of counting of a figure: per cent of what, per 100,000 inhabitants, change on the same month of the previous year.
Scope: Population, region, period or subject the document's statements cover.
Condition: A condition under which a statement of the document holds (“if … remains unchanged”, “provided that …”).
Stated limit: A limitation, uncertainty or exclusion the document states about itself.
Status and date: Status of the data, provisional or final figures, revision, date of the figures.
Reported position: A position the document reports without adopting it.

A language model draws the Prüfraster afresh with every check, for every selected source under the check profile General and only from its selected sections; a point from a deselected section could not carry a finding. There is no Prüfraster yet while the text is being prepared.

Every point needs a verbatim passage in the selected section it names. Before the findings, it goes through a check without a language model. It is discarded and counted with the reason if a field is missing, its quote is too short or does not stand like that in the section named, a figure in its description does not stand in its quote with the same value and unit, it brings in knowledge from outside its quote, it uses a word from the list of excluded terms, or it repeats another point of the same kind with the same quote. At most 25 are kept per source.

The points kept are numbered (R1, R2, …) and are available to the check. A finding names the points it was measured against. The quote of every point named also stands on its check path, so it can be followed there.

The Prüfraster is stored with the result. The workspace and the side-by-side document show it ordered by source and kind, every point with passage, page and quote; the table names the points of every finding. How many points were discarded is stated in the head of the result.

7. Which knowledge is used

The language model may use knowledge of method. The rules it is meant to apply are set out here:

Per cent and percentage points are different units.
A share is not a number of cases.
A figure has a base and a period.
Provisional figures are not final ones.
A confidence interval is an uncertainty the document itself states.
A sample is not the population.
Mean and median are different measures.
A correlation is not yet a cause.
A summary reproduces the result it summarises.
A note to a table applies to that table.
“laut”, “zufolge” (according to) and the subjunctive of reported speech mark a position the document only reports.

The language model may compare figures (greater, smaller, rounded, about twice as high). It produces no figure of its own.

Subject knowledge must not be the basis of a finding: other studies, statistics, data or versions that are not selected. The language model is instructed accordingly, and the evidence check discards a finding whose explanation recognisably brings in knowledge from outside the selection, including a figure that is neither in the statement nor in a passage (section 8, rules 7 and 9). Whether subject knowledge flows unstated into a classification cannot be ruled out this way; that is why every finding names its passages.

If a statement turned on another document, such as the release for the previous month, the language model can only suggest that document as a further source. It counts once it has been fetched and selected; then it stands in the list of sources like any other.

8. Evidence check

Before a finding is shown, it goes through a check without a language model. If it breaks one of the following rules, it is discarded and counted with the reason; the statement then appears as “not checked”. The evidence check establishes that the passages cited stand like that in the selection and that the figures match them. Whether the classification itself holds, for example “shortened” rather than “in substance”, it does not check.

A finding cites at least one passage. The only exception is the result “not found in the source text”, which is produced without a language model (below).
Every passage of the check path stands word for word in the selected section the finding names. The same words elsewhere in the selection or in a deselected paragraph do not count. The comparison evens out quotation marks, dashes, whitespace and hyphenation at line ends. Case counts, except for the first letter of a quote, and a quote matches whole words only. A quote with an ellipsis matches when its parts stand in that order in the same section, with at most 300 characters left out between two parts.
A passage has at least three words, or two words with at least 15 characters together (spaces included); for a quote with an ellipsis this applies to each part. A single compound word shows nothing.
A reading by the criterion of wording cites at least one passage in the text of the document, not only in a list such as a table of contents, and for a law not only in the explanatory memorandum or the cover sheet.
“Open, defensible” supports at least two different readings with two different passages.
“Word for word” requires the statement to have at least one quote and every one of its quotes to stand in the selected sections. “Not found” is ruled out when all its quotes stand there.
If a finding says “word for word” or “in substance”, every figure of the statement stands in at least one passage of the check path, with the same value and the same unit: per cent, percentage points, per mille or none. The value is compared, so “52 %” and “52,0 %” are the same figure, and so are “3,5 Mio.” and “3.500.000”. If a word such as “rund” (around), “etwa” (about), “knapp” (just under), “mehr als” (more than) or “bis zu” (up to) stands before the figure, a figure of the same unit is enough. “3 Prozent” against a passage that says “3 Prozentpunkte” is discarded. Not figures in this sense are years without a unit, dates, ordinals, numbers after words such as “§”, “Art.”, “Tabelle” (table), “Abbildung” (figure) or “Seite” (page), and figures in the title of a selected source. Numbers written out in words (“drei Viertel”, three quarters) are not checked.
A finding names only points from the Prüfraster of its own check, and the quote of every point named stands at the same place in the document in a passage of its check path. A reading by the criterion of stated scope and limits names at least one point.
The explanation of the finding brings in no recognisable knowledge from outside the selection: no figure that stands with the same value and unit neither in the statement nor in a passage of the check path nor in the quote of a point named (it may cite the numbers of a passage such as “2.4, Absatz 2, Satz 3”); no section or article number that appears neither in the check path nor in the statement nor in the selected sections; no mention of a federal court or the European Court of Justice in its usual short form, no ruling (“Urteil”), no “Beschluss vom”, no case number with “Az.” and no commentary, unless the statement itself uses the word; no text in quotation marks that stands neither in a passage nor in the statement nor word for word in a selected section. The search is for these fixed markers; a mention written out in full or in an unusual form can slip through.
The explanation uses no word from the source check's list of excluded terms. It holds words that attribute untruth, deception or unlawfulness to a text, and recommendations to persons.

The exception: not found in the source text

If no quote of a statement about the source stands in any of the given sources, there is no passage that could be shown. This result is settled without a language model: the quotes are searched for in the sources, and once more in the selection before the check. The finding names the sources and versions searched. The recorded search is its evidence.

9. Verifiability and check path

Every finding states how much it takes to verify it yourself. The language model makes this classification.

can be read directly: Quote and passage stand side by side. One passage is enough.
traceable along the check path: The finding follows further passages of the document, such as a definition or a note to a table. The steps are shown in order.
expert assessment: The finding depends on the method of the field, beyond what the document itself states. It is marked as a machine assessment, with the recommendation to have it reviewed by someone from the field.

A finding “not found in the source text” has no passage. It can be verified by searching the sources named for the quote.

The check path is the ordered chain of passages a finding rests on, for example 3.1, paragraph 2, sentence 1 → 2.2, paragraph 1, sentence 3 → strengths and limitations, paragraph 2. Where the document has pages, the page is given too. Every finding has one sentence in plain language and an expandable expert version with all passages.

10. What comes from the language model

A language model produces the splitting of the text into statements and the classification of each statement, the Prüfraster of every source, and every finding with its explanation. No person reviews the individual findings before they are shown.

Produced without a language model are the division of the sources into sections, paragraphs and sentences, the choice of the check profile when a source is first prepared, the version statement, the result “not checked” for a section that is not selected, the check of every point of the Prüfraster, the evidence check with its checks on figures, and the result “not found in the source text”.

11. Review, origin and change of model

The check profile is to be measured against reference cases: 10 to 15 statements from two or three texts, checked by hand and reviewed by someone from the field, for example with training in statistics or empirical method. Until the check profile passes them, it carries the notice “not yet reviewed by a subject expert”, and so does every result produced with it.

State of this version: not yet reviewed by a subject expert.

Whether a document is the original source or only reports on it, the source check does not yet recognise. That is why every source under the check profile General carries this notice in the list of sources and in the side-by-side document:

Checked against this document as it stands. Whether it is the original source is not checked.

The check profile is bound to a language model named exactly, never to a label behind which the model can change, and independently of the check profile Legal text, so that each check profile changes only by its own reference cases. Once reference cases exist, they are run again before every change of model. Every change gives the check profile a new version. A stored result keeps the version and the review state it was produced with.

12. Limits and responsibility

The result is machine-made. Only what the selected sources say is checked. It is not legal advice and not a legal assessment of an individual case.
Whoever uploads a file confirms that they may use it. The platform does not check this. You are responsible for it.
Whether and where you pass on or publish a result is your decision. You are responsible for it, in particular for quotes from the text checked and from non-official sources and for statements about named persons.
Recording words that were not spoken in public can be a criminal offence in Germany (Section 201 of the German Criminal Code). The live check is therefore only for public broadcasts and for interviews in which every person speaking has agreed. The platform does not check whether you may record. You alone are responsible for that.

13. Version and changes

Current version: 0.2.2. The version changes when the criteria, the instructions to the language model, the model itself or the rules of the evidence check change. General shares the splitting into statements with the check profile Legal text; if it changes, both get a new version.

0.2.227 September 2026
Every source carries its level of origin: official, primary source or secondary source. Against a secondary source, a report about another document, a finding records the reproduction only and says “primary source missing”; the instructions to the language model state this rule for every such source, and the evidence check discards a finding with an interpretation resting on one.
0.2.127 September 2026
Splitting the text into statements and references lets the language model think for less long. On short example texts it used to take up to three and a half minutes, now about ten seconds. Instructions and answer format are unchanged. General shares the splitting with the check profile Legal text.
0.2.027 September 2026
Prüfraster: for every source the check draws from the selected sections what the document sets for itself, each point with a verbatim passage and a check of its own. New criterion stated scope and limits, decided only through a point of the Prüfraster. The Prüfraster is stored and shown with the result.
0.1.027 September 2026
First version for documents with fixed wording that are not legal texts: division into sections, paragraphs and sentences, the criteria wording and context, and two checks on figures in the evidence check. Claims about the world stay evaluations.