48. Fact vs opinion: the difference is not “objective words” versus “personal words”

School exercises often teach a useful first approximation: a fact can be checked, while an opinion expresses a judgement or preference. That distinction helps beginners, but serious reasoning needs one more layer. Many opinions contain factual premises, and many factual statements are expressed through words that require definitions.

“The bridge opened in 2010” is a factual claim because evidence can bear on whether that event occurred in that year. “The bridge is beautiful” is primarily evaluative because beauty depends partly on aesthetic judgement. “The bridge is the longest pedestrian bridge in the district” is factual only after we define pedestrian bridge, district and length. “The bridge is badly designed” could be an opinion, or it could compress several testable engineering claims about capacity, accessibility, maintenance or safety.

The important question is therefore not whether a sentence sounds factual. Ask what part of it is truth-evaluable, what part depends on criteria, and whether those criteria have been made explicit.

A mature fact-versus-opinion exercise teaches students to unpack mixed sentences. “This was a disastrous policy because waiting times rose 40%” contains at least two layers: a quantitative claim about waiting times and an evaluative conclusion about whether that consequence makes the policy disastrous. The first can be verified statistically. The second requires a normative standard and possibly other outcomes.

49. Fact vs belief: sincerity is evidence about the believer, not automatic evidence about the world

A person can sincerely believe something that is false. They can also sincerely disbelieve something that is true. Belief tells us about a cognitive state: what a person accepts, expects or takes to be the case.

This distinction matters in testimony. If a witness says, “I am certain the car was blue,” the certainty is relevant evidence that the witness holds a strong memory. It does not independently establish the car’s colour. Lighting, distance, memory contamination and later discussion can all affect the relationship between sincere belief and external reality.

Conversely, a learner may say, “I think I guessed,” even when their explanation shows real understanding. Self-belief and demonstrated competence are different variables.

Fact checking should therefore resist two shortcuts: “They believe it, so it must be true” and “They are uncertain, so it must be false.” The correct move is to treat belief as one kind of evidence whose reliability depends on the question and context.

50. Fact vs knowledge: truth and justification answer different questions

A proposition can be true even when nobody currently knows it. An event can occur without being observed. A lost archive can contain the correct date long before anyone discovers it. This is one reason facts should not be defined merely as what people currently accept.

Knowledge introduces an epistemic question: what grounds are available for accepting the proposition? Philosophers disagree about the exact analysis of knowledge, just as they disagree about the exact ontology of facts. This article does not attempt to settle those debates. For practical verification, the distinction is enough: truth concerns how the claim relates to what is the case; justification concerns how responsibly a knower is entitled to accept it.

That means “it happened” and “we know it happened” are different statements. The first concerns the world. The second adds a claim about the strength of the evidence available to us.

This distinction becomes especially useful in developing stories, historical research and science. A result may be true before replication, but the community’s justified confidence can strengthen after independent replication. The world does not become truer; the knowledge position becomes stronger.

51. Fact vs inference: derived conclusions can be factual, but the derivation must stay visible

Some facts are directly observed in a bounded sense: the instrument displayed 17.2, the document contains a paragraph, the door was visibly open in a verified image. Other conclusions are inferred from several observations.

Suppose a machine is known to be operating at 10:00, a failure alarm appears at 10:07 and inspection at 10:10 finds it stopped. If no direct observation captures the instant of failure, we can still infer that the transition occurred within a bounded interval. The inference can be strong enough to use operationally without pretending that 10:07:00 was directly observed as the physical failure instant.

Inference is not a weaker word for guess. A valid inference can be extremely strong. The discipline is to preserve the premises and rule that connect them to the conclusion.

For AI systems, this separation is especially important. Retrieval can supply observations and source claims; the model then synthesises an inference. If the final answer hides that step, readers may mistake a derived conclusion for a directly stated source fact. Good factual writing can be concise while still signalling “records indicate”, “the evidence supports”, “this implies” or “the exact time remains bounded”.

52. Testimony: a witness is a source with a viewpoint, not a camera with perfect coverage

Human testimony can be indispensable. Many events are known partly because people experienced them, recorded them or later described them. The mistake is to treat testimony as either infallible or worthless.

A witness has a position in space and time. They may see one side of an event and not another. Attention may be divided. Memory can change. Incentives can matter. Expertise can improve interpretation of some observations while introducing assumptions in others.

A strong testimony analysis asks: What could this person directly observe? What are they reporting from memory? Which parts are interpretation? How soon after the event was the account recorded? Did later information contaminate the account? Are there independent observations with different failure modes?

Two witnesses who disagree may both be sincere and partially correct because they observed different parts of the event. Conversely, many witnesses can share the same mistaken assumption. Testimony becomes more informative when its observational boundary is explicit.

53. Primary source vs secondary source: closer to the event is not automatically closer to the truth

A primary source is typically close to the event, phenomenon or original research process being investigated: an original dataset, official record, experiment report, interview, photograph, diary, contract or direct observation. A secondary source analyses, interprets, synthesises or contextualises primary material.

Primary sources are often essential because they reduce some layers of transmission. They are not automatically accurate, complete, unbiased or sufficient. A primary witness can be mistaken. An original spreadsheet can contain a coding error. An official announcement can accurately record an intention that never becomes an outcome.

A high-quality secondary source can sometimes be the better starting point because it compares multiple primary sources, explains method, identifies known corrections and places a claim inside a larger body of evidence. The International Fact-Checking Network’s commitments emphasise source transparency and the use of suitable primary sources where available, while also requiring a methodology that allows readers to understand how conclusions were reached.

The practical rule is not “always trust primary” or “always trust expert synthesis”. Use the source whose role best fits the claim, then inspect method, provenance, independence, currency and corrections.

54. Authenticity vs accuracy: a genuine document can contain a false statement

Authenticity asks whether an artifact is what it purports to be. Accuracy asks whether its content correctly describes the relevant reality.

An authentic memo from 1987 can contain a mistaken figure. A genuine newspaper page can repeat a rumour that was later disproved. A real photograph can be captioned with the wrong place. A legitimate database entry can preserve a value that was subsequently corrected.

These distinctions matter because verification has layers. First establish whether the artifact is authentic enough for the task. Then establish what it actually says or depicts. Then ask whether that content supports the external-world claim.

When fact checking fails at this boundary, two opposite mistakes appear. A forged artifact is accepted because its content sounds plausible, or an authentic artifact is treated as infallible because its provenance is genuine. Authentication answers a necessary question. It does not answer every question.

55. Authority vs independence: the best source for status may be the worst source for self-evaluation

Authority and independence contribute different kinds of evidential value.

If we ask whether a university officially awarded a degree, the university’s authorised record may be the controlling institutional source. If we ask whether the university’s programme produces better outcomes than alternatives, independent evaluation may be more informative than the university’s marketing material.

A company is authoritative about what it announced and how it labels its own product. It is not automatically independent evidence of product superiority. A government agency can be authoritative about the text and effective date of a rule while independent researchers may be better placed to estimate the rule’s social effects.

Good fact checking therefore builds a source portfolio, not a prestige ranking. One source establishes status, another measures outcome, another supplies historical context, and another challenges the interpretation. Reliability comes from matching source role to claim role.

56. Fact status is a lifecycle, not a permanent sticker

Facts about a changing world and conclusions drawn from incomplete evidence need status over time. A practical lifecycle can include candidate, supported, established, disputed, corrected, superseded, retracted or invalidated. These are communication states, not metaphysical grades of truth.

A claim may begin as a candidate because one source reports it. Independent evidence can move it to strongly supported. A later audit may reveal an error and correct it. A policy fact can be established for one period and later become superseded because the rule changes. A publication can remain part of the historical record while its evidential standing is withdrawn.

The key design principle is that status transitions should leave a trail. We should be able to answer: What was the earlier claim? Why did its status change? Which evidence triggered the change? What is the current owner of the answer? Which past decisions depended on the earlier state?

This is where fact reasoning connects directly to What Changed?. A knowledge system needs both the history of the world and the history of its own conclusions about the world.

57. Provisional facts: useful conclusions do not always have to wait for perfect finality

Some decisions cannot wait until every uncertainty is resolved. Emergency operations, breaking news, preliminary science, project control and classroom diagnosis often require action under incomplete evidence.

The solution is not to call guesses facts. It is to state provisional conclusions with their status and revision conditions. “Preliminary counts indicate approximately 120 participants; final reconciliation is pending” can be useful and honest. “There were exactly 120 participants” would claim more than the current process supports.

A provisional fact record should identify what is stable enough to use, what remains open, and what later evidence will trigger review. This makes action possible without erasing uncertainty.

The distinction is especially important for AI-generated summaries. An assistant may need to answer now, but it can preserve “reported”, “preliminary”, “as of”, “not yet independently confirmed” and similar status language instead of flattening every current best estimate into final truth.

58. Disputed facts: disagreement is a property of the evidence state, not automatic proof that truth is unknowable

When credible sources disagree, “disputed” can be the correct present status. That status should not be confused with “both sides are equally supported” or “there is no fact of the matter”.

Disagreement can arise from different definitions, time windows, measurement methods, incentives, data access or genuine uncertainty. One source may later prove better supported. Sometimes the disagreement persists because the underlying evidence is irrecoverably incomplete.

A responsible disputed-fact note explains what is agreed, what conflicts, which evidence supports each claim and what additional observation could discriminate between them. It avoids performing false balance by treating a large body of convergent evidence as equivalent to a weak contrary assertion simply because both exist.

Dispute status is therefore a routing instruction: preserve the conflict, investigate the discriminating evidence, and avoid premature certainty. It is not a permanent destination unless the evidence boundary truly prevents resolution.

59. Unsupported, unknown and contradicted are three different outcomes

Fact checking becomes more accurate when it refuses to collapse every failed verification into “false”.

Unsupported means the evidence currently supplied does not adequately justify the claim. Unknown or unresolved means the available evidence does not settle the question. Contradicted means stronger relevant evidence supports an incompatible conclusion.

Suppose someone says a meeting occurred at 4 pm. No calendar entry is found. That alone may leave the claim unsupported or unresolved. Now suppose access logs, room records and multiple independent participants establish that the meeting occurred at 2 pm and ended before 3 pm. The 4 pm claim is no longer merely unsupported; it is contradicted by a stronger reconstructed timeline.

This three-way distinction prevents a common reasoning error: treating “I cannot verify it” as equivalent to “I verified the opposite”. It also prevents false agnosticism when strong contrary evidence really exists.

60. Correction, retraction and supersession change status in different ways

A correction changes some content while leaving the larger work in force. A retraction or invalidation withdraws standing from a work or claim under the relevant authority. Supersession replaces a current version with a newer one while often preserving the earlier version as historically valid for its period.

These transitions matter because a citation can remain technically real while no longer representing the current evidential state. An article can still exist at the same URL after a correction. A retracted paper can remain accessible for transparency. An old policy can remain archived after a new policy takes effect.

Crossref’s Crossmark service is built around this problem: readers need a way to check the current status of scholarly content, including updates, corrections and retractions. The general lesson reaches beyond journals. A verifier should ask not only “Can I find the source?” but also “What is its current status?”

The correct historical record may therefore preserve both the original and the correction while making the current authoritative state unambiguous.

61. Published does not mean current

Search engines are very good at finding pages that exist. Existence is not the same as current validity.

An old help page can remain indexed after software behaviour changes. An archived government PDF can remain downloadable after a regulation is amended. A product specification can describe a discontinued revision. A medical guideline can be superseded while mirrors of the earlier document remain online.

Current-state verification therefore needs a status check: publication or issue date, version, effective interval, update notice, correction notice, retraction notice, successor document and the authority responsible for the state.

Historical questions reverse the priority. The superseded document may be exactly the correct evidence if the task is to reconstruct what rule applied then. “Old” and “wrong” are not synonyms; “new” and “relevant” are not synonyms either.

62. Reproducibility helps verify a result, but reproducibility is not identical to truth

If another analyst can take the same data, code and procedure and obtain the same result, an important part of the result’s provenance has been verified. We have learned that the reported transformation can be reproduced.

That is valuable, but it does not prove that the inputs describe the world correctly or that the research design answers the intended question. A reproducible calculation can reproducibly process biased data. A reproducible simulation can implement the wrong assumptions faithfully.

Replication adds another layer by testing whether a result can be obtained with new observations or an independently implemented method. Even replication should be interpreted in context: populations can differ, instruments can differ, and a genuine effect can vary across settings.

The verification ladder is therefore layered: can we reconstruct what was done, can we reproduce the calculation, can independent evidence support the result, and does the result actually bear on the target claim?

63. Evidence sufficiency: verification needs a stopping condition

It is possible to search forever. Every source can lead to another source, every measurement to another calibration, every archive to another box. Practical verification needs a principled reason to stop.

The stopping condition should depend on consequence, reversibility, uncertainty and the cost of being wrong. A low-stakes classroom question may need one authoritative textbook and a quick cross-check. A public safety claim, legal determination or high-consequence medical decision may require formal domain-specific standards, multiple evidence streams and expert review.

We stop not because absolute certainty has been reached, but because the evidence has crossed an appropriate decision threshold and additional searching is unlikely to change the conclusion enough to justify its cost.

This makes evidence sufficiency a decision problem as well as an epistemic one. The same proposition can deserve different verification effort depending on what will be done with it.

64. Stopping rules prevent both premature certainty and infinite research

A useful verification plan decides in advance what would count as enough evidence, what evidence would reopen the question, and which unresolved uncertainties are acceptable.

For example: verify a current office-holder using the official institutional page plus one independent current source; stop if both agree and no transition announcement is found. Reopen if a dated succession announcement appears. For a numerical research claim, reproduce the calculation and inspect the underlying method; stop only after the denominator, exclusions and uncertainty are understood.

Predeclared stopping rules reduce two biases. The first is stopping as soon as the preferred answer appears. The second is continuing to search until some source can be found to support a desired alternative.

Good research remains open to surprise while refusing to turn search duration into a proxy for truth.

65. Positive and negative claims often need different search completeness

Finding one verified example can establish an existential claim: “At least one red book is in the box.” Failing to find a red book does not establish “No red books are in the box” unless the search process was capable of examining the complete relevant set.

This asymmetry appears in archives, databases and web search. One authentic record can establish that an event occurred at least once. Establishing that no record exists may require a complete inventory, a well-defined archive boundary or a search protocol with known coverage.

Negative claims should therefore carry a completeness statement. “No matching entry appears in the indexed records searched from 2020–2025 under these identifiers” is more informative than “There is no record.”

The search boundary belongs inside the fact because it defines what absence can legitimately mean.

66. Reference classes: a number needs the right comparison group

Facts about frequency, risk, performance and unusualness depend on the reference class used for comparison.

A student score of 75 may be high relative to one assessment cohort and ordinary relative to another. A machine failure rate can be alarming for a new model and normal for an ageing fleet. A disease incidence can appear high when compared with the national average and lower when compared with a demographically similar population.

The factual claim “this is unusual” is incomplete until unusual relative to what has been specified. Reference classes should be chosen because they match the question, not because they create the most dramatic contrast.

This is also a safeguard against misleading benchmark comparisons. The benchmark, denominator and eligibility rule are part of the statement, not background decoration.

67. Base rates: background frequency changes how surprising evidence should be

An unusual test result can look decisive when the underlying condition is very rare. A common event can look suspicious when attention focuses only on a vivid example.

Base rates provide background information about how frequently the target state occurs before the new evidence is considered. They do not replace the new evidence; they help interpret it.

In education, a single difficult paper should be interpreted against the learner’s normal variation. In fraud detection, an alert should be understood against the prevalence of genuine fraud and the system’s false-positive rate. In source analysis, an extraordinary transcription error may be less plausible than a routine version mismatch—until direct evidence shows otherwise.

Fact checking becomes stronger when it asks both “What does this evidence show?” and “How common would this observation be under the main alternatives?”

68. Stronger claims need evidence capable of discriminating stronger alternatives

A claim can be surprising because it conflicts with well-supported background knowledge, because it asserts a rare event, or because it attributes a strong causal mechanism. In such cases, ordinary low-resolution evidence may be insufficient.

This does not mean unusual claims should be rejected automatically. It means the evidence should be capable of separating the unusual claim from more ordinary explanations such as error, misidentification, coincidence or measurement failure.

If a photograph appears to show snow in Singapore, a verifier should not stop at “the pixels look white”. Check location, date, weather records, image provenance and whether the scene might instead contain foam, hail, artificial snow or another location. The stronger conclusion demands evidence targeted at the plausible alternatives.

Proportional verification is therefore about discriminatory power, not prejudice against novelty.

69. Fact checking across languages: translation can change the strength of a claim

Translation introduces another representation layer. A source can be authentic and accurately quoted in its original language while an English rendering changes certainty, tense, scope or institutional terminology.

A modal verb meaning “may” can become “will”. A term meaning “proposal” can be rendered as “plan” or “policy”. An honorific or legal category can be mapped to an English word with a different institutional meaning. Numerals, dates and decimal conventions can also shift.

For consequential claims, preserve the original wording, identify the translator or translation method, and check disputed terms against domain usage. Machine translation can be excellent for discovery while still requiring human or specialist review for legal, technical or culturally specific distinctions.

The factual object is not merely the translated sentence. It includes the relationship between the translated sentence and the source-language claim.

70. Fact checking numbers: units, denominators and percentage points are part of the claim

Numbers feel concrete, which makes numerical errors especially persuasive.

Check units before arithmetic. A value in thousands can be mistaken for a raw count. Kilograms can be confused with pounds. Annual rates can be compared with monthly totals. Nominal currency can be compared across years without accounting for inflation when the intended question is purchasing power.

Check denominators before percentages. Moving from 40% to 50% is a 10-percentage-point increase and a 25% relative increase. “Risk doubled” can describe movement from one in a million to two in a million; the relative change is large while the absolute change remains tiny.

Check rounding and aggregation too. A displayed total may not equal the visible rounded components even when the underlying full-precision values do. Numerical fact checking is not only about recalculating; it is about preserving what the number measures.

71. Images and video: authentic media can still be attached to the wrong event

Visual evidence can answer questions that prose cannot. It can also mislead without a single pixel being fabricated.

An authentic flood photograph from 2022 can be reposted as evidence of a flood today. A real video can be cropped so that the preceding action is hidden. A genuine photograph can be taken from an angle that makes a crowd appear larger or smaller. A verified image can be accurately located but wrongly dated.

Visual verification therefore asks several independent questions: Is the media authentic enough for the task? Where was it captured? When? What appears inside and outside the frame? Is it the event being claimed? Has the caption preserved the original context?

Deepfake detection is only one part of visual fact checking. Context failure remains possible even when the media itself is completely real.

72. Metadata is useful evidence, not a magic truth field

Metadata can record creation time, modification time, device information, location, author, file version and processing history. It can be extremely valuable for reconstructing provenance.

Metadata can also be absent, stripped, copied, modified, generated by software or refer to a different event from the one a reader assumes. A file creation timestamp may record when a copy was created on a device rather than when the original scene was photographed. A webpage modified date may change because of a template update while the substantive data remain old.

Treat metadata as one evidence layer and cross-check it against content, provenance and independent records where consequence requires it.

The correct question is not “Does the file have metadata?” but “What event does this metadata field actually represent, and how reliable is the process that generated it?”

73. AI-generated text is evidence of an output, not evidence of the world it describes

If an AI system answers, “The building opened in 1998,” one fact is immediately available: under the relevant prompt, context and configuration, the system produced that sentence. The sentence’s truth about the building still requires external evidence.

This distinction seems obvious, yet generated text often acquires accidental authority because it is fluent, specific and well formatted. A model can also cite a genuine source while synthesising the wrong entity, date or scope.

AI output can become useful evidence in other ways. A model may classify thousands of records consistently, propose candidate matches or extract claims for human checking. In those cases, the model output is a derived record with its own provenance, model version, input set and validation state.

The public rule is simple: synthetic output is not an observation merely because it is coherent. Use it to navigate evidence, transform evidence or propose inferences, then keep verification attached to the underlying world-facing sources.

74. Retrieval freshness: correct evidence can become wrong for a current-state question

A source can be accurate, authoritative and irrelevant because it is too old for the question.

Suppose a university page from January correctly names Professor A as dean. In September, Professor B has taken office. The January page remains historically accurate. It is now insufficient evidence for “Who is the dean today?”

Freshness should be matched to volatility. Geological location facts change slowly. live service status, transport disruptions, office-holders, prices, policies and software versions can change quickly. A source that was adequate yesterday may need reverification today.

Good retrieval therefore includes a temporal intent: current, latest known, historical as-of, or valid during a stated interval. Without that intent, search systems can rank a perfectly good historical page above the source that answers the actual temporal question.

75. The evidence-conflict matrix: disagreement becomes diagnosable when its dimensions are separated

When two credible sources disagree, compare them across dimensions before deciding that one must simply be wrong.

If all these dimensions match and the values still conflict, the disagreement is genuine. Preserve it and seek discriminating evidence. The matrix does not force resolution; it prevents avoidable confusion from masquerading as deep contradiction.

76. Recheck triggers: a verified fact should know what could make it stale

Verification is stronger when it records not only why a claim is accepted now but what future event would require review.

• a new version or amendment is published;
• a cited source issues a correction or retraction;
• a role changes hands;
• a measurement instrument is recalibrated;
• a classification rule or denominator changes;
• a previously missing dataset becomes available;
• new evidence contradicts the current conclusion;
• the source exceeds the freshness horizon appropriate to the property;
• the same claim is reused in a higher-consequence decision.

This converts fact maintenance from periodic memory into event-driven responsibility. A stable historical date may never need another check. A live service status may need a new observation minutes later.

The recheck trigger belongs to the claim’s operational context. It says, in effect: “This conclusion remains good while these assumptions remain good.”

77. Write a human-readable fact note

A knowledge system may store dozens of fields, but readers often need one transparent paragraph explaining why the claim currently deserves its status.

A useful fact note can follow this pattern: claim → status → evidence → limits → time/version → revision trigger.

Example: “As of 16 September, the organisation lists Mei Tan as Director, effective 1 September. This is supported by the current official leadership page and an appointment notice dated 20 August. An older page naming the previous director is retained as historical evidence. Recheck if a later appointment or correction notice appears.”

This is much more informative than “Verified: Mei Tan is Director.” It shows the validity interval, source roles and reason the conclusion may later change.

78. A compact public fact-status schema

A public, domain-neutral representation can preserve the following fields without exposing proprietary runtime machinery:

• claim_id — stable identifier for the claim record;
• claim — exact proposition being evaluated;
• subject_id — stable referent where appropriate;
• claim_type — descriptive, quantitative, causal, predictive, normative, institutional;
• scope — population, geography, jurisdiction, method or other boundary;
• valid_from / valid_until — when the claim applies;
• verified_at — when the current status was assessed;
• status — established, supported, provisional, disputed, unresolved, contradicted, superseded, retracted;
• evidence_refs — supporting and contradicting records;
• provenance — derivation relationships among those records;
• method — how observations or calculations were produced;
• uncertainty — quantitative or qualitative limits;
• supersedes — earlier fact record, where relevant;
• recheck_trigger — event or age condition requiring review.

The schema is intentionally small enough to understand. Real systems should not collect personal data merely because a field could theoretically improve provenance. Data minimisation, permissions and retention obligations remain separate release gates.

79. Current fact table vs historical fact table

One table of “facts” often becomes confused because it tries to serve two questions at once: what is current, and what was true before?

A current-fact view should identify the best-supported state now. A historical-fact view should preserve earlier validity intervals, corrections and superseded states. A third view may sometimes be needed: what the system believed at an earlier date before later corrections arrived.

Suppose an address was recorded as A until March, changed to B in April, and corrected in June to show that the move actually happened in February. The current historical reconstruction can say B was valid from February. An audit of what staff saw on 1 March must still show A.

This is the factual version of bitemporal reasoning: history of the world and history of knowledge should not overwrite each other.

80. Worked case: the paper was corrected after you cited it

A researcher cites a paper reporting 18.2%. Months later, the publisher issues a correction: the correct value is 12.8%. The DOI and article title remain the same.

The original citation is authentic. The quoted 18.2% was genuinely present in the version read. The current scientific use should now account for the correction.

A status-aware workflow checks the article’s current state—through the publisher, Crossmark where available, or other authoritative update mechanisms—before reusing the number. The fact note should preserve both states where history matters: “The original publication reported 18.2%; a later correction changed the value to 12.8%.”

This prevents two bad outcomes: continuing to repeat an obsolete value, or silently rewriting history so that nobody can understand earlier analyses built from the original publication.

81. Worked case: the primary source is wrong and the secondary synthesis is better

An original expedition diary records that a party reached a site on 14 May. Later archival work compares the diary with ship logs, correspondence and astronomical records and demonstrates that the diary date was written one day out of sequence. A scholarly secondary history uses 15 May.

The diary remains a primary source and an authentic artifact. That does not make every factual statement inside it correct. The secondary work can be better evidence for the event date because it has access to multiple independent records and explicitly resolves the discrepancy.

A fact checker should cite according to the job: the diary for what the diarist recorded, the scholarly reconstruction for the best-supported event date, and both if the historical discrepancy itself is relevant.

Source hierarchy is therefore not a ladder in which “primary” always sits above “secondary”. It is a map of evidential roles.

82. Worked case: the number is true and the denominator makes the story misleading

A headline says complaints doubled. The count rose from two complaints to four. The statement is arithmetically correct: the count increased by 100%.

Now add context. The service handled 100 cases in the first period and 10,000 in the second. Complaint rates moved from 2% to 0.04%. The count doubled while the complaint rate fell dramatically because the denominator changed.

Neither number should be hidden. The honest account states both: “Complaints increased from two to four while handled cases increased from 100 to 10,000, so the complaint rate fell from 2% to 0.04%.”

This is why factuality is not only the correctness of each isolated sentence. Context can determine whether a set of true statements gives a materially accurate picture.

83. Worked case: the image is real and the event is wrong

A dramatic photograph circulates during a storm. Reverse searches and archival pages show that the image is genuine but was first published three years earlier in another country.

The image has passed one authenticity check and failed the event-identity check. “This photograph is real” and “This photograph depicts today’s storm” are separate claims.

The repair is to preserve the distinction in the fact check: identify the original event, show the earliest reliable context available, and explain that reuse—not necessarily image fabrication—created the false claim.

This pattern is common enough to deserve its own reflex: verify media-to-event binding, not merely media authenticity.

84. Worked case: the successor is announced but the incumbent still holds office

An organisation announces on 1 June that a new director will take over on 1 August. A profile article published on 15 June calls the incoming person “the director”.

The announcement establishes the intended future appointment. Unless the effective date changed, the incumbent remains the current office-holder on 15 June.

A status-aware fact record can hold both claims without contradiction: “Incoming director: B, announced 1 June, effective 1 August” and “Current director through 31 July: A.”

This is a recurring failure in biographies, leadership pages and news summaries. Announcement time, effective time and current-role state must remain separate.

85. Students and teachers: fact, opinion, inference and evaluation form a ladder

Students are often asked to separate fact and opinion as if every sentence belongs neatly to one box. Better teaching shows a ladder of reasoning.

1. Observation or record: “The graph shows a value of 40 in March.”
2. Factual description: “The March value is lower than the April value.”
3. Inference: “The change may indicate rising demand.”
4. Explanation: “Demand rose because the campaign began in April.”
5. Evaluation: “The campaign was successful.”

Each step adds something. Comparison adds a relationship. Inference adds interpretation. Explanation adds cause. Evaluation adds criteria and judgement.

Teachers can mark more intelligently by asking which step the question requires and whether the student supplied evidence appropriate to that step. A correct observation should not receive full credit for an explanation question; an unsupported causal explanation should not be rewarded merely because it sounds sophisticated.

This ladder also improves writing. Students learn to signal the difference between “the source states”, “the data show”, “this suggests”, “this may be because” and “therefore I judge”. Clear factual language is partly the art of keeping those verbs honest.

86. Researchers: report observed, estimated and modelled quantities separately

Research becomes easier to audit when tables and prose distinguish direct observations, derived variables, statistical estimates and model predictions.

An observed survey response is one kind of record. A weighted population estimate derived from those responses is another. A regression-adjusted effect estimate adds model assumptions. A forecast for next year adds assumptions about future process stability.

All can be legitimate scientific outputs. Problems arise when later summaries erase the distinctions and write as though modelled quantities were directly counted.

Researchers should therefore preserve the transformation chain: raw or source observations → exclusions and coding → derived variables → model → estimate → uncertainty → interpretation. When a later correction changes one link, the downstream claims that depend on it can be identified and re-evaluated.

This practice strengthens reproducibility and protects against fact inflation: the tendency for a carefully qualified research result to become a simpler and stronger public claim as it travels through abstracts, press releases and media summaries.

87. Editors and fact-checkers: freeze the claim before searching

Search is dangerous when the target claim keeps changing. A writer begins with “The city opened the line in 2019,” finds evidence that construction began in 2019 and quietly treats the search as successful. The evidence matched a neighbouring claim, not the original one.

Before researching, write the claim in checkable form: subject, predicate, date, scope and important definitions. Then record what evidence would confirm, contradict or leave it unresolved.

If research reveals that the original wording was wrong but a nearby statement is true, revise the sentence explicitly. Do not let search results retroactively redefine what was supposedly verified.

This “freeze then test” discipline also makes collaboration easier. Another editor can reproduce the verification because the target did not move invisibly while sources were being gathered.

88. AI fact maintenance: retrieval, synthesis and memory need separate status

An AI system can fail at three different stages. Retrieval can find the wrong source or an outdated version. Synthesis can misread a correct source. Memory can preserve a once-correct conclusion after the world changes.

Those failure modes need different repairs. Retrieval needs better entity, date and source selection. Synthesis needs claim-evidence binding and limits on inference strength. Memory needs validity intervals, supersession and recheck triggers.

A public-safe fact-maintenance loop can therefore be described simply: identify the claim → retrieve temporally and semantically fit evidence → verify source status → synthesise only what the evidence supports → assign a fact status → store the conclusion with provenance and validity → recheck when a relevant change occurs.

That pattern is intentionally general. It does not reveal hidden prompts, proprietary orchestration, private user memory stores or internal eduKateAI controls. The important educational principle is that an intelligent system should be able to update a conclusion without pretending that its previous information never existed.

89. High-stakes domains: ordinary fact-checking rules are a floor, not a substitute for domain standards

The general framework in this article is useful for medicine, law, finance, engineering and safety because identity, time, evidence, uncertainty and provenance matter there too. It is not a replacement for professional standards in those domains.

A clinical claim may require validated diagnostic evidence and qualified interpretation. A legal claim may depend on jurisdiction, authority, procedural posture and current legislation. An engineering claim may require calibrated measurement, test protocols and safety factors. A financial statement may be governed by accounting and audit standards.

The correct handoff is explicit: use this foundation to structure the question, then route to the specialist evidence regime that owns the consequence.

Foundational literacy should make expert work easier to understand, not create the illusion that a generic checklist can replace expert judgement where the cost of error is high.

90. True statements can still mislead through omission

A report can contain no literally false sentence and still produce a materially misleading picture by selecting only favourable facts.

“Complaints fell in August” can be true while complaints rose sharply across the full quarter. “Three students achieved full marks” can be true while most of the class deteriorated. “The project completed two milestones early” can be true while its critical final milestone is six months late.

This creates a distinction between sentence-level factuality and representation-level adequacy. Fact checking individual sentences is necessary. It may not be sufficient when the communication claims to summarise an overall condition.

To test omission, ask what reasonable comparison, denominator, time window or contrary evidence a reader would need to interpret the stated fact. The answer is not to include every possible detail. It is to include the details whose absence would materially change the implied conclusion.

91. Evidence debt: convenient claims become expensive when provenance is lost

Teams often copy a useful number into a slide, then into a report, then into a webpage. Months later nobody remembers which dataset or version produced it. The number may still be correct, but verifying it again becomes expensive.

This is evidence debt: future verification work created by failing to preserve source, method, version and scope when the claim was first used.

Evidence debt compounds. A derived chart copies the undocumented number; an AI indexes the chart; another report cites the AI-generated summary. Each layer increases apparent authority while making the original evidence harder to recover.

The repair is inexpensive at creation time: attach a stable source reference, retrieval or validity date, definition, unit and method note to claims likely to be reused. Provenance is not administrative decoration. It is deferred-cost control for truth maintenance.

92. Fact handoff: what the next reader or system needs

When a verified conclusion moves from researcher to editor, teacher to parent, database to dashboard, or tool to AI, the handoff should preserve enough structure that the next user does not have to reconstruct the entire investigation.

• the exact claim;
• current status and date verified;
• scope and validity interval;
• key supporting evidence;
• important contradicting evidence;
• source status and version;
• material uncertainty;
• known correction or supersession history;
• what would trigger recheck;
• the specialist owner if the claim leaves this article’s scope.

A good handoff is concise enough to use and rich enough to audit. It does not expose private information merely for completeness. It preserves the facts needed for the next legitimate job.

93. The boundary map: what this article owns

This article owns the foundational question: what is a fact, what makes a factual claim supportable, and how should fact status be maintained as evidence and the world change?

It should hand off rather than swallow neighbouring specialist jobs:

• For detailed claim construction, use What Is a Claim?
• For evidence types, corroboration and support, use What Is Evidence?
• For the operational checking process, use What Is Verification?
• For missing records and negative inference, use What Does a Blank Actually Mean?
• For identity, copies and versions, use When Are Two Things the Same Thing?
• For temporal state and version history, use What Changed?
• For specialist editorial procedures, use the dedicated editorial fact-checking and source-verification guides rather than duplicating their workflow here.

Canonical ownership is part of quality. A 20,000+ word article should be deep because its job genuinely requires depth, not because it absorbs every adjacent topic.

94. The 30-second answer

If you need the entire article compressed into one practical rule:

A fact is about what is the case; a factual claim is our statement about it. Verify the claim by matching it to evidence that fits the right entity, time, scope, definition and method, then keep its status current when the world or the evidence changes.

That rule leaves room for strong conclusions without pretending that citations, confidence, authority or repetition can manufacture truth.

95. The durable principle: facts become useful when they can survive a change of reader, time and source

A fragile fact is one that works only while the original author remembers where it came from. A durable fact can be handed to another person, revisited months later, compared with a new version and still explain why it deserves its present status.

Durability comes from boundaries: exact claim, stable referent, valid time, scope, evidence, provenance, method, uncertainty, status and revision history. Those boundaries do not weaken knowledge. They make knowledge portable.

This is why a serious learning library needs more than correct sentences. It needs a way to preserve how correct sentences were earned, when they apply, what supersedes them and what evidence would make them change.

The purpose of fact checking is not to make language timid. It is to make confidence transferable.