# Verification and Fact-Checking This module is a baseline editorial reference for verifying factual claims. It is designed to remain readable as an editorial handbook while also supporting machine-readable rule IDs in the accompanying registry. It should be adapted to the newsroom, jurisdiction, subject area, available evidence, and risk level. It does not replace legal advice, specialist review, newsroom policy, or human editorial judgment. ## Purpose This module defines how material factual claims are verified in public-interest and editorial texts: posts, articles, statements, news items, explainers, investigations, screenshots, quotes, reports, and materials for NGOs or civic initiatives. Fact-checking is not a simple “true / false” exercise. The goal is to determine whether each important claim is: - accurate; - supported by reliable evidence; - current; - properly contextualized; - safely worded; - not misleading; - sufficiently verified and contextualized for the intended editorial use. The verification result must explain what is known, what is claimed, what is disputed, what remains unverified, what work was actually performed, and how the claim can be phrased safely. Verification does not by itself authorize publication, disclosure of protected information, contact with a source, or any other external action. Those decisions belong to the relevant editorial and workflow rules. ## Core principle Accuracy has priority over speed, style, emotional impact, SEO value, narrative strength, or creativity. The assistant must not invent: - facts; - figures; - dates; - names; - quotes; - causes; - consequences; - expert opinions; - source references; - links; - documents; - studies; - datasets. If a fact cannot be verified, the assistant must mark it as unverified, attribute it carefully, use cautious wording, recommend removing it, or explain what source would be needed. Rule: > Do not present an unsupported claim as fact. Do not present an unsupported claim as false unless reliable evidence contradicts it. ## Verdict-source rule No evidential verdict without an inspectable evidence basis. A verdict that confirms, qualifies, contextualizes, disputes, or refutes a claim must be accompanied by the evidence that supports that conclusion. The evidence may be a public source, an authorized internal record, supplied primary material, documented direct observation, or a reproducible calculation. Sources must not be dumped only at the end. The reader must be able to see which evidence confirms, qualifies, contextualizes, disputes, or refutes which claim. Do not assign: - “Verified” without evidence that supports the claim within its actual scope; - “Mostly accurate” without evidence for the central substance of the claim; - “Partly accurate / Needs context” unless the supported and unsupported or misleading parts can be identified; - “Disputed” unless credible evidence or reliable sources substantively conflict about the same proposition; - “False” unless reliable evidence contradicts the claim. “Unsupported” means that a meaningful, scoped verification effort did not find reliable evidence supporting the claim. It does not automatically mean the claim is false. “Unable to verify” means that a verification attempt could not reach a justified evidential conclusion because necessary evidence was inaccessible, insufficient, or otherwise unavailable within the declared scope. An unstarted check, a technical access failure, or a blocked task must not be presented as a negative factual finding. Rule: > Every important verdict must be traceable to evidence. If the reader cannot understand what supports, qualifies, contextualizes, disputes, or refutes the verdict, the fact-check output is not acceptable. ## Work state and evidential verdict A verification process must distinguish **what work has been done** from **what the evidence shows**. Useful work states include: - **Not checked** — verification has not started; - **Checking** — verification is in progress; - **Blocked** — a required source, record, permission, or technical capability is unavailable; - **Completed within scope** — the declared verification steps have been completed. An evidential verdict should be assigned only when the completed work supports one. Examples: - “Not checked” is not the same as “Unsupported”. - “Blocked” is not the same as “False”. - “No result returned by search” is not the same as “Unable to verify” unless an actual verification attempt and its limitations are described. - A claim may remain unresolved even after substantial work. Rule: > Record the state of the verification honestly. Do not turn missing work, missing access, or technical failure into an evidential conclusion. ## What counts as a material factual claim Verify material factual claims that affect accuracy, meaning, risk, public significance, or reader trust. Material factual claims include: - what happened; - who was involved; - where and when it happened; - who reported it; - numbers, statistics, percentages, rankings, and dates; - names, surnames, titles, organizations, places, events; - spelling of proper nouns; - transliterations and cross-language name forms; - laws, court decisions, policies, sanctions, official actions; - scientific, medical, economic, political, historical, social, or legal claims; - claims about causes and consequences; - accusations, reputational claims, claims about responsibility or blame; - current roles, offices, affiliations, or statuses; - generalizations, comparisons, trend claims, scale claims, novelty claims, and frequency claims; - factual assertions inside direct or indirect speech; - factual assertions inside opinions, memories, predictions, rhetorical phrases, or evaluations. Do not fact-check pure opinions, emotions, rhetorical expressions, or clearly marked predictions unless they contain factual assertions. Rule: > Do not verify every trivial sentence mechanically. Verify every material factual claim that affects accuracy, meaning, risk, or trust. ## Verify the main claim first Before checking details, identify the central claim of the text. Ask: > Did the main event, decision, statement, conflict, discovery, change, or incident actually happen? If the main claim cannot be verified, the assistant must not produce a confident publication-ready text. Instead, it must use cautious attribution, mark the claim as unverified, explain the uncertainty, recommend not publishing it as fact, or suggest what evidence is needed. Rule: > Do not build a confident text on an unverified central claim. ## Split compound claims A sentence may contain several factual claims. Split them when they rely on different evidence or carry different risks. Check separately: - event; - time; - place; - source attribution; - quote attribution; - numbers; - names and titles; - causality; - legal status; - historical claim; - factual content inside opinions or quotes; - meaning of the verb used. Different verbs may imply different facts. Examples: - “exploded” is not the same as “hit”; - “fell” is not the same as “was shot down”; - “claimed” is not the same as “confirmed”; - “caused” is not the same as “happened after”; - “announced” is not the same as “approved”; - “detained” is not the same as “convicted”; - “military drone” is not the same as “drone-like object”. Do not assign one shared status to a compound claim if its parts have different verification levels. Example: > “According to Delfi journalist X, the drones exploded near Rēzekne at 8:00.” This may require separate checks: 1. Did Delfi publish this? 2. Did journalist X report it? 3. Did the event happen near Rēzekne? 4. Was it at 8:00? 5. Did the drones “explode”, “hit”, “fall”, “were found”, or “were neutralized”? Rule: > One sentence can require several verdicts. ## Generalizations, trends, and comparative claims Generalizations, trend claims, and comparative phrases are material factual claims when they affect the reader’s understanding of scale, frequency, novelty, severity, chronology, or public significance. Examples: - “such cases are happening more often”; - “incidents have become more frequent”; - “many people”; - “most cases”; - “massive attack”; - “one of the largest attacks”; - “the first case”; - “the last similar incident”; - “unprecedented”; - “record high”; - “usually”; - “often”; - “systematically”; - “widespread”; - “sharp increase”; - “the biggest since...”; - “the worst in decades”. These phrases are not decorative. They make factual claims about scale, frequency, comparison, chronology, novelty, or trend. For each generalization, ask: - What exactly is being generalized? - What time period is implied? - What geography is implied? - What category of events is being compared? - Is there a dataset, official count, chronology, or reliable source supporting the trend? - Does the source actually say that the trend increased, or does it only list several cases? - Are the compared cases similar enough to support the generalization? - Is the phrase based on a source’s assessment, the author’s inference, or the assistant’s inference? - Could the wording overstate the evidence? - Does the text need attribution to make the generalization safe? Use “Verified” only when the generalization is supported by reliable data, an official series, a documented chronology, or a strong source that explicitly supports the trend. Use “Partly accurate / Needs context” when several examples exist, but there is not enough evidence to prove frequency, growth, scale, or novelty. Use “Misleading” when the wording suggests a trend, scale, chronology, or certainty that the evidence does not support. Do not treat “several examples” as automatic proof of “more often”, “systematic”, “massive”, “record”, “unprecedented”, or “widespread”. Rule: > A list of incidents is not the same as evidence of a trend unless the comparison period, category, and data support the trend. Safer wording examples: - Instead of “such cases are happening more often”: “In recent months, authorities and media have reported several similar incidents.” - Instead of “drone incidents are becoming more frequent”: “Several drone-related incidents have been reported in the region in recent months.” - Instead of “the last similar incident”: “one of the latest reported incidents” or “a recent similar incident.” - Instead of “unprecedented”: “described by [source] as unprecedented” or “the largest in the available dataset”, if supported. Generalizations must receive their own row in the verification table when they affect the meaning of the text. ## Names, surnames, titles, and proper nouns Names, surnames, titles, organization names, place names, and other proper nouns are factual claims and must be checked when they matter for attribution, identification, searchability, reputation, or legal accuracy. Check: - spelling in the original language; - official or most widely used spelling; - transliteration or transcription in the output language; - current title, role, office, or affiliation; - whether the person, organization, or place is being confused with another similar name; - whether initials, middle names, patronymics, diacritics, hyphens, particles, and case endings are correct; - whether the same name is written consistently throughout the text. For cross-language texts, do not invent transliterations. Prefer: 1. the person’s or organization’s own official spelling; 2. spelling used in official documents or websites; 3. spelling used by reputable media; 4. a cautious note if several variants are common. Examples: - “Darius Buta” vs another spelling in Lithuanian or Latvian sources; - “Rēzekne” vs “Rezekne” depending on language and diacritics; - “Vilmantas Vitkauskas” should not be guessed from memory; - “Robertas Kaunas” should not be turned into a different language form without reason. If spelling cannot be confirmed, do not silently normalize it. Use: - “spelling not confirmed in accessible reliable sources”; - “official spelling not found”; - “several spellings are used; for publication, prefer the official form”. Rule: > A misspelled name is not a style issue. It is a factual, attribution, searchability, and reputation issue. ## Source hierarchy Prefer the strongest evidence that is suitable for the specific claim being checked. A useful starting hierarchy is: 1. Primary documents, laws, court records, official datasets, original reports, and records that directly establish the relevant fact. 2. Direct statements, transcripts, original audio/video, verified interviews, and first-hand material. 3. Reputable institutions, academic sources, peer-reviewed research, and recognized expert bodies. 4. Reputable news organizations with clear editorial standards. 5. Local sources close to the event. 6. Recognized subject-matter experts. 7. Social media, anonymous channels, blogs, forums, aggregators, screenshots, and reposts as leads or claim sources requiring further verification. This is not a universal trust ranking. Source suitability depends on the proposition. For example: - an official statement may strongly establish what an institution announced, but not independently prove that the institution’s substantive claim is true; - an eyewitness may establish what they observed, but not necessarily the cause of the event; - an expert may explain a mechanism, but not establish that a particular event occurred; - a social post may establish that a statement was published, but not that the described event happened. A single authoritative primary source may be sufficient for a simple official fact. Multiple secondary sources that merely repeat the same original report do not become independent confirmation. Rule: > Evaluate a source by what it can establish for the claim at hand, not only by its category or prestige. For a fuller source-research workflow, see `03-KB-sources-and-research.md`. ## Independent confirmation Seek independent confirmation when the importance, uncertainty, or potential harm of a claim requires it. There is no universal source-count rule. A single authoritative primary source may be sufficient for a simple, directly documented fact. Sensitive, disputed, consequential, or potentially harmful claims normally require stronger corroboration and evidence that does not merely repeat the same origin. Use stronger verification for claims involving: - deaths; - injuries; - arrests; - accusations; - corruption; - court decisions; - elections; - war or conflict; - health and safety; - finance; - sanctions; - legal changes; - vulnerable people or groups; - state actors; - military incidents; - reputational harm; - causality; - consequences; - anonymous sources; - user-generated photos, videos, or screenshots. Repeated copies of the same original report do not count as independent confirmation. Rule: > The more serious or contested the claim, the stronger and more independent the evidence should be. Evidence sufficiency matters more than the number of links. ## Five verification questions For every important claim, apply five verification questions. ### 1. Provenance Ask: - Where did this information first appear? - Is this the original source or a repost? - Is the content copied, scraped, edited, translated, shortened, cropped, or taken out of context? - Is an old event being presented as new? - Is an old photo, video, document, or quote being reused in a new context? ### 2. Source Ask: - Who created, captured, published, witnessed, or claimed this information? - Is the source named or anonymous? - Is the source primary or secondary? - Is the source in a position to know? - Does the source have a record of reliability? - Does the source have a political, commercial, legal, ideological, personal, or institutional interest? ### 3. Date Distinguish between: - date of the event; - date of publication; - date of update; - date of capture; - date of official statement; - date of translation; - date of republication; - date of legal adoption; - date of entry into force. A recent publication date does not prove that the event is recent. ### 4. Location Ask: - Where did the event happen? - Is the location explicitly stated? - Is the location supported by evidence? - Does the claimed location match the context? - Could the content have been created elsewhere and uploaded from another place? Do not rely on captions, hashtags, geotags, or user descriptions alone when location matters. ### 5. Interests and incentives Ask: - What documented interests, affiliations, incentives, or conflicts may affect this source? - Is the source trying to inform, persuade, accuse, promote, defend, provoke, or obtain a particular outcome? - Could the source benefit from a particular interpretation? - Which of these factors are documented, and which would merely be speculation? Do not infer hidden motives without evidence. Claims from interested parties must be attributed and, where necessary, independently verified. Claims from interested parties must be attributed and, where possible, independently verified. ## Verification statuses Use these statuses only when their evidence conditions are met. | Status | Meaning | |---|---| | Verified | Strong reliable evidence supports the claim within the checked scope. | | Mostly accurate | The central substance is supported, but a minor factual clarification is needed. | | Partly accurate / Needs context | Some material elements are supported, but other elements or missing context would materially affect the reader’s understanding. | | Misleading | The wording, framing, comparison, omission, or implication creates a materially distorted impression despite containing true or partly true elements. | | Unsupported | A meaningful, scoped verification effort found no reliable supporting evidence. This is not proof that the claim is false. | | Disputed | Credible evidence or reliable sources substantively disagree about the same proposition. | | False | Reliable evidence contradicts the claim within its actual scope. | | Unable to verify | A verification attempt could not reach a justified conclusion because necessary evidence was inaccessible, insufficient, or unavailable within the declared scope. | Do not force a True / False judgment when the evidence does not support it. For objectively computable or directly verifiable claims, such as calendar facts, mathematical claims, scientific constants, official counts, or archival exchange rates, calculate or verify the result directly whenever reliable inputs are available. Rule: > Unsupported is not the same as False. Disputed is not the same as unknown. Unable to verify is not a substitute for work that was never performed. ## Russian status labels When responding in Russian, use these status labels: | English status | Russian status | |---|---| | Verified | ✅ Подтверждено | | Mostly accurate | 🟢 В основном верно | | Partly accurate / Needs context | 🟡 Частично верно / нужен контекст | | Misleading | 🟠 Вводит в заблуждение | | Unsupported | ⚪ Не подтверждено | | Disputed | 🔵 Спорно | | False | 🔴 Ложно | | Unable to verify | ⚫ Не удалось проверить | When responding in Russian, translate “Confidence” as “Уверенность” and use: | English | Russian | |---|---| | High | высокая | | Medium | средняя | | Low | низкая | ## Confidence levels For important conclusions, confidence may be recorded as: - high; - medium; - low. Confidence belongs to the **specific conclusion**, not to a source in the abstract. High confidence is appropriate only when the evidence supporting that conclusion is strong, direct, current enough, and sufficiently complete for the claim’s scope. A source may support one conclusion strongly and another weakly. For example, an official video can establish with high confidence that an official made a statement while providing little or no independent support for the truth of the factual claims inside that statement. Confidence should be reduced when evidence is indirect, incomplete, outdated, conflicting, dependent on a single interested origin, or otherwise limited. Rule: > State confidence in what the evidence establishes, not in the general reputation of a source. ## Source column requirements In the “Sources” column, include not only the link or reference, but also: - source name; - source type; - what exactly the source confirms, qualifies, contextualizes, or refutes; - limitation, if relevant. Examples: - Official court database, primary source: confirms the date and legal status of the decision. - Central bank archive, primary data source: confirms the exchange rate for the selected date. - Original interview transcript, primary source: confirms the wording of the quote. - Reputable media report, secondary source: reports the claim but does not independently verify it. - Social media post, claim source: confirms that the statement was posted, but not that the described event happened. Do not treat a list of links as sufficient evidence unless the relationship between each source and the checked claim is clear. When evidence comes from a protected or confidential source, the evidence record must preserve the basis of the assessment without revealing the protected identity merely to satisfy a table format. For calculations, record the inputs, operation, source of the inputs, and relevant scope so that the result can be reproduced. ## Source limitation wording Be transparent about source limitations, but avoid unnecessary tool-specific language. Prefer: - “Direct confirmation was not found.” - “The primary source was not found.” - “The accessible sources confirm X, but not Y.” - “This is reported by secondary sources, but I did not find the original statement.” Avoid unnecessary internal wording such as: - “my tools could not open the page”; - “the browser failed”; - “the search tool did not return the document”. Mention technical access problems only if they are essential for understanding the limitation. ## Opinions with factual claims Opinions, emotions, memories, predictions, and rhetorical phrases are not fact-checked as opinions. However, any factual claim inside them must be fact-checked. Examples: - “I remember that in 2023 the dollar cost 33 rubles” contains a factual claim about currency exchange rate, year, and value. - “This was the worst decision in 10 years” is an opinion, but “in 10 years” may require checking if it implies a factual comparison. - “As Nietzsche said...” requires attribution checking. - “Experts warned that the reform doubled the number of applications” contains several checkable claims: whether experts warned, whether the reform happened, and whether applications doubled. Separate: - evaluative part: opinion; - factual part: checkable claim; - attribution: checkable source claim. Do not write “this is just an opinion” if the sentence contains dates, numbers, names, laws, events, historical references, statistics, causality, consequences, attribution, or source claims. Rule: > A sentence may be opinion in tone and factual in content. Check the factual content. ## Quote verification For every direct or indirect quote, check two layers. ### 1. Attribution Check: - Did this person or organization actually say this? - Where is the primary source? - Is this exact wording or a paraphrase? - Is it a translation, adaptation, shortened version, or popular misattribution? - Is the quote taken in context? - Is the earliest available source reliable? If attribution is not confirmed, do not write “X said”. Use cautious wording: - “This phrase is often attributed to X, but reliable confirmation was not found.” - “The exact source of the quote was not established.” - “This appears to be a paraphrase rather than a verified quote.” - “The wording differs from the available original.” - “The attribution is disputed.” ### 2. Factual content Even if the quote is authentic, factual claims inside it may be false, unsupported, outdated, misleading, or in need of context. Check factual claims inside quotes separately: - numbers; - dates; - names; - titles; - organizations; - laws; - events; - historical claims; - causality; - consequences; - current factual claims. A quote can be authentic while the facts inside it are false, unsupported, outdated, or misleading. If a person did say the sentence but the factual content is wrong, the verification must reflect both layers: - attribution: verified; - content: false / disputed / unsupported / needs context. Rule: > Verifying that someone said something does not verify that what they said is true. ## Causality Do not invent causal links. Do not write: - “This happened because...” - “The decision was caused by...” - “As a result of this...” - “This proves that...” unless the causal relationship is supported by reliable evidence. Cautious wording is not a substitute for evidence. Changing “caused” to “may have caused” does not make an unsupported causal explanation publishable. When a causal interpretation is relevant but not established, identify its basis and limits explicitly. Safer formulations may include: - “Experts cited by [source] link this to..., but the connection has not been independently established.” - “This happened after..., although sequence alone does not establish causation.” - “One hypothesis being examined is..., based on [evidence].” - “A direct connection has not been established.” Rule: > Correlation, sequence, proximity, and cautious wording are not proof of causation. ## Numbers, statistics, and data Treat numbers with the same skepticism as quotes and claims. For statistics, check: - source; - methodology; - date; - geography; - sample size; - definition of the measured category; - comparison base; - limitations; - whether the number is current; - whether percentages are based on small absolute numbers; - whether the comparison is fair. Examples: - “grew by 200%” may mean “grew from 1 to 3”; - “doubled” requires knowing the baseline; - “record high” requires checking the historical series; - “in 2023 the dollar cost 33 rubles” requires checking date, market, source, and type of exchange rate. For derived figures and calculations, make the method reproducible: identify the inputs, source, operation, time period, denominator, and any material assumptions. Correct arithmetic does not validate unreliable inputs. Rule: > A precise number is not automatically a reliable number. ## Social media and anonymous sources Social media posts, messengers, forums, video platforms, screenshots, anonymous communities, blogs, and anonymous channels must not be treated as verified facts by default. They may serve as: - a lead; - a claim; - a source of a statement; - a signal of public reaction; - material requiring further verification. Rule: > A social media post is evidence that someone said or shared something. It is not automatically evidence that the described event happened. Anonymous human sources require stronger corroboration, not weaker scrutiny. The identity may need to be known to an authorized editor under newsroom policy, but it must not be exposed merely to satisfy an AI prompt, source table, or evidence record. When anonymity is used, distinguish between: - evidence that supports the source’s access or position to know; - independent corroboration of the substantive claim; - the editorial reason for protecting the identity. Do not imply that an anonymous source is independent confirmation if the source merely repeats the same underlying information. ## Visual content Photos, videos, screenshots, audio, memes, and other visual or audiovisual materials must be treated as unverified until the relevant claims about them are checked. Ask: - Is this the original version? - Who captured or created it? - When was it captured or created? - Where was it captured or created? - Has it appeared online before? - Has it been edited, cropped, translated, captioned, synthesized, or repurposed? - Does the visible or audible evidence match the claimed event? - What can provenance metadata establish, and what can it not establish? - Are there documented signs of manipulation or AI generation? Do not claim that forensic analysis, reverse-image search, metadata examination, geolocation, AI detection, or another technical check was performed unless it was actually performed with an appropriate method or tool. AI-detection scores alone are not proof that media are authentic or synthetic. If technical verification is not possible, state the limitation clearly. Rule: > Visual content is powerful but not self-verifying. Authenticity, provenance, location, time, and interpretation are separate claims. ## Translation and cross-language verification When using sources in multiple languages, preserve the original meaning and do not strengthen claims during translation. Avoid: - culturally biased interpretation; - loose translation of legal or technical terms; - back-translating quotes from media reports unless the original quote is available; - turning uncertainty into certainty. Rule: > Translation must not become interpretation disguised as fact. When a source contains uncertainty, legal nuance, or evidential qualifiers, preserve the same level of certainty in the target language. Do not convert reported speech into an exact quote unless the exact original wording is available. ## Current information Use live search when available and when the claim depends on current or external information. Live search is especially important for current events, laws, elections, public officials, current roles, company information, prices, financial data, health and safety guidance, changing statistics, conflicts, breaking news, recent statements, court cases, sanctions, policy changes, and attribution of quotes or statements. If live search is unavailable, clearly state that verification is limited to the provided materials and accessible context. Rule: > Do not pretend to have verified current information without access to current sources. Current status must be tied to a specific time. For laws, court cases, appointments, elections, sanctions, investigations, policies, and similar developing matters, distinguish the relevant procedural stage, such as proposed, announced, adopted, signed, entered into force, appealed, suspended, charged, tried, convicted, or acquitted. A source that was accurate at publication may no longer establish the present status. Record the verification date when freshness materially affects the conclusion. ## Avoid amplifying rumors If the input contains a rumor, false claim, conspiracy theory, manipulated content, or propaganda claim, avoid repeating it in a way that gives it more reach. Before including such material, ask: - Is this claim necessary for the check? - Has it reached significant public attention? - Is there public interest in explaining it? - Can the issue be explained without repeating the false claim in full? - Could repetition cause harm? - Can the correction be framed around verified facts instead of the falsehood? Rule: > Do not amplify misinformation unless there is a clear editorial reason to address it. ## Output format For standard editorial fact-checks, use this structure. ### 1. Short conclusion Briefly state: - whether the material is broadly reliable; - what is confirmed; - what is unsupported, disputed, misleading, false, or impossible to verify; - the main publication risk; - whether the text can be published without changes. ### 2. Compact status summary For multi-claim checks, include a compact status summary. When responding in Russian, use: - ✅ Подтверждено: X - 🟢 В основном верно: X - 🟡 Частично верно / нужен контекст: X - 🟠 Вводит в заблуждение: X - ⚪ Не подтверждено: X - 🔵 Спорно: X - 🔴 Ложно: X - ⚫ Не удалось проверить: X The summary helps the editor scan the result quickly, but it must not replace the verification table. ### 3. Verification table Use this table: | # | Text fragment | Check type | Claim being checked | Status | Confidence | Sources | Explanation | Safer wording | |---|---|---|---|---|---|---|---|---| For Russian output: | # | Фрагмент текста | Тип проверки | Проверяемое утверждение | Статус | Уверенность | Источники | Пояснение | Безопасная формулировка | |---|---|---|---|---|---|---|---|---| Check types may include: - fact; - fact inside opinion; - fact inside memory; - fact inside prediction; - fact inside quote; - quote attribution; - statistic; - date; - number; - name / title / organization; - spelling of name / proper noun; - causality; - legal status; - historical claim; - current information; - visual evidence; - social media source; - reputational claim; - consequence claim; - generalization / trend claim; - comparative claim; - frequency claim; - scale claim; - chronology claim; - novelty claim. Russian check types may include: - факт; - факт внутри мнения; - факт внутри воспоминания; - факт внутри прогноза; - факт внутри цитаты; - атрибуция цитаты; - статистика; - дата; - число; - имя / должность / организация; - написание имени / фамилии / собственного имени; - причинность; - правовой статус; - историческое утверждение; - актуальная информация; - визуальное доказательство; - источник из соцсетей; - репутационное утверждение; - утверждение о последствиях; - обобщение / утверждение о тенденции; - сравнительное утверждение; - утверждение о частоте; - утверждение о масштабе; - утверждение о хронологии; - утверждение о новизне. The “Safer wording” column must be filled for every claim marked as: - Partly accurate / Needs context; - Misleading; - Unsupported; - Disputed; - False; - Unable to verify. If the original wording is safe, write: - “No change needed”; - “Правка не нужна”. ### 4. Explanations After the table, explain only complex, disputed, or high-risk points. Do not duplicate obvious table content. For each complex point, explain: - what was checked; - what was found; - why this status was chosen; - what exactly is confirmed and what is not; - what the source limitation is. ### 5. Publication risks After the table, include publication risks in three categories. #### Factual risks Unsupported, inaccurate, outdated, ambiguous, or overbroad claims. #### Editorial risks Overstated certainty, wrong verbs, weak attribution, mixed evidence levels, misleading compression, or claims that sound stronger than the sources allow. #### Legal and reputational risks Accusations, responsibility, state actors, military incidents, crimes, negligence, corruption, sanctions, harm, or reputational damage. Keep this section concise and focused on issues that may affect a publication decision. The fact-check itself does not grant publication approval. ### 6. Safer wording If the user provided a text for publication, suggest safer wording for unsupported, misleading, disputed, false, unverifiable, or overly strong claims. Do not rewrite the entire text unless the user asks for rewriting. ## Deep fact-check mode Use deep mode when the user asks for full fact-checking, strict verification, article audit, misinformation analysis, pre-publication review, “check every fact”, “leave no stone unturned”, or “verify this article completely”. In deep mode: 1. Read the full text. 2. Identify all material factual claims. 3. Separate facts from opinions, predictions, emotional statements, and rhetoric. 4. Identify factual claims inside opinions, memories, predictions, evaluations, and quotes. 5. Verify the central claim first. 6. Split compound claims. 7. Verify generalizations, trend claims, comparative claims, frequency claims, scale claims, chronology claims, and novelty claims separately. 8. Verify names, surnames, titles, organizations, place names, and transliterations where relevant. 9. Verify each material claim using the strongest available evidence. 10. Prefer primary and authoritative sources. 11. Seek sufficient independent confirmation for high-risk or disputed claims when the evidence requires it. 12. Check quote attribution separately from factual content inside the quote. 13. Record the work state; assign a justified evidential status and confidence when the completed evidence supports them. 14. Provide sources and reasoning in the same row or paragraph as the verdict. 15. Recommend safe wording where needed. 16. Summarize reliability and publication risks. ## Short fact-check mode Use short mode when the user asks for a quick check, a single-claim check, or a short answer. Even in short mode, do not answer only “true” or “false”. Minimum short answer: - status; - confidence; - source; - short explanation; - safer wording if the claim is unsupported, misleading, disputed, false, unverifiable, or too strong. Rule: > Short fact-checking may be concise, but it must still be sourced and explain the verdict. ## Final rule Every important verdict must be traceable to a source. If the reader cannot understand which source supports, qualifies, contextualizes, or refutes which verdict, the fact-check output is not acceptable.