A citation can point to a genuine webpage and still fail to support the sentence beside it. The page might discuss the same topic without establishing the claimed number, contain an older edition of the relevant rule, or support only half of a compound statement. Citation anchoring connects each material claim to the exact source passage, revision and context that justify it. The goal is not to make an AI answer look scholarly; it is to make its evidence checkable by the reader.
Did you know that the number of links in an answer is not a measure of how many claims were proved? A trustworthy agent identifies its meaningful assertions, finds the applicable source for each, checks the quotation and surrounding conditions, then distinguishes direct support, partial support, contradiction and unavailable evidence. For current education decisions, the source’s effective year and responsible issuer can matter as much as its title.
This guide builds an invented one-hundred-claim ledger, traces an answer containing six material assertions and transfers the method to an updated school notice. The numerical results are original teaching examples, not live eduKate citation quality statistics or evidence that artificial superintelligence exists. It extends How Super Intelligence Works and the separate chapters on Conflicting Evidence and Retrieval Coverage.
Choose a useful reading route
- Researching official school information? Start with exact claims, passages and source scope.
- Building a citation system? Study annotation selectors, provenance and link revisions.
- Learning Mathematics? Recalculate the hundred-claim and six-claim coverage examples.
- Publishing evidence-led articles? Follow link checks, meaningful qualifiers and saved-state review.
Contents
- 1. A source link and a supporting passage are different objects
- 2. Split complex sentences into material assertions
- 3. Source identity needs issuer, location and revision
- 4. The W3C annotation model helps describe precise targets
- 5. Quotation correctness does not prove claim correctness
- 6. Page-level relevance is weaker than claim-level support
- 7. Tables and numbers need exact row context
- 8. Current sources can change after a citation was created
- 9. Evidence provenance is a chain, not a decoration
- 10. An invented one-hundred-claim evidence laboratory
- 11. Six material claims can form one final answer
- 12. Evidence conflicts require source-scope reasoning
- 13. Citation locators must survive changing page layouts
- 14. Correct a citation without falsifying the original record
- 15. Practical citation review before publication
- Frequently asked questions
- Source and standards notes
- Further learning
1. A source link and a supporting passage are different objects
Imagine a paragraph claiming that a particular school changed its Mathematics subject entry rule. A link to the school’s home page is not enough evidence for the change, even if the home page is authentic and relevant to the institution. The supporting source should contain the specific statement, applicable cohort and effective date. A useful citation lets another person inspect the words that actually justify the claim without hunting through dozens of unrelated pages.
A model may retrieve a source title and generate a convincing explanation from memory. The resulting link looks credible, but the underlying page may not include the claim at all. The assistant should preserve a source ID and passage location during retrieval, then verify the particular sentence against that passage. A generated citation is a proposal for evidence, not proof that the page was read or that its contents support every nearby assertion.
Claims also have scopes. A document can correctly state that one programme accepts an applicant group while saying nothing about another programme. The source supports the narrower proposition, not a broad declaration about all routes to university. Citation anchoring requires the claim’s subject, action, condition and time scope to match the evidence rather than merely sharing key nouns.
Return to contents or continue to Split complex sentences into material assertions.
2. Split complex sentences into material assertions
Consider a sentence saying a subject is offered, requires a particular prerequisite and guarantees admission to a certain pathway. These are three distinct assertions. A source listing offered subjects supports only the first; a separate eligibility table may address the second; the alleged guarantee may be unsupported. Attaching one general link to the entire sentence can create a false impression that all three parts were checked.
Extract the claims a reader might act upon: amounts, deadlines, programme rules, cause-and-effect statements, comparisons and promises. Ordinary connecting words need not receive a new citation each, but material facts should have traceable evidence. A claim-to-passage matrix makes the coverage visible, including which conditions were established and which remain uncertain or contradicted.
Be careful with paraphrase. The original source might say an applicant may be considered under certain conditions, while the model writes every applicant will be admitted. Both discuss admissions, yet the stronger wording is not supported. The checker needs to compare modality, quantifier and exceptions, not just textual similarity. A useful final answer can retain the narrower accurate proposition.
Return to contents or continue to Source identity needs issuer, location and revision.
3. Source identity needs issuer, location and revision
Record the actual URL or stable document ID, the issuing organisation, a retrievable passage location, retrieval time and relevant source revision where available. A page title alone is not stable enough. Two annual notices can share the same title while governing different cohorts. A copied mirror can repeat the wording of an earlier version while the official current page has changed.
W3C’s PROV-O framework describes provenance relationships among entities, activities and responsible agents. Its vocabulary can help explain a claim’s derivation chain: a particular source document was used in an evidence-checking activity to support a generated response. This is a modelling framework, not a guarantee that the claim is true. Real verification still requires the actual source content and correct interpretation.
A passage anchor can be a heading, paragraph ID, page number, table row identifier or text selector. The chosen locator must be stable enough for the source format and reading environment. A PDF page number may shift after a new edition, while a dynamic webpage may re-render its elements. Keep the source revision and a short evidence excerpt so a reviewer can recover the intended text if the first link changes.
Return to contents or continue to The W3C annotation model helps describe precise targets.
4. The W3C annotation model helps describe precise targets
The W3C Web Annotation Data Model represents annotations that relate a body to a target resource, with selectors that can identify a segment rather than the entire page. It includes concepts for text quotes and positions. An AI citation system can use these ideas to store a precise link between one claim and one location in an external document. That precision improves auditability even when the final reader sees a simpler hyperlink.
Selectors have limitations. A text-position selector can break after edits insert paragraphs before the quoted segment. A text-quote selector can match the wrong occurrence when identical phrases appear several times. A heading anchor may stay stable while the text under it changes materially. Source identity, applicable revision and a stored short excerpt can supplement the locator rather than assuming one selector is infallible.
The evidence should remain legible to a person. A citation containing only a machine-generated fragment identifier may be difficult to inspect if the browser does not support the relevant link format. Provide a source title, accessible context and a fallback route to the document. Technical anchors exist to make verification easier, not to hide the evidence behind specialized tooling.
Return to contents or continue to Quotation correctness does not prove claim correctness.
5. Quotation correctness does not prove claim correctness
A model may copy an exact sentence from a source and still use it incorrectly. The sentence might describe a historical situation or a hypothetical example. The surrounding paragraph may narrow the subject or state an important exception. A quote whose bytes match is more reliable than an invented quotation, but matching bytes are only the first step in assessing support for the actual claim.
Consider a fictional school notice explaining that students who meet both A and B may apply. An AI answer could quote the fragment ‘may apply’ accurately but omit the two prerequisites. The resulting paraphrase would be misleading. Inspect the full relevant passage, conditions and context, then write the sentence no more strongly than the evidence warrants.
Respect source rights and reader attention. Citation verification usually requires only a short exact excerpt and surrounding source reference, not copying an entire article into the output. A long reproduced text can obscure which clause supports the claim and create unnecessary copyright or privacy concerns. Keep the cited quotation proportionate and add your own explanation of what it establishes.
Return to contents or continue to Page-level relevance is weaker than claim-level support.
6. Page-level relevance is weaker than claim-level support
A retriever can perform well at finding topical pages while the final answer remains unsupported. A high similarity score might put a Mathematics curriculum overview at rank one, but a particular school subject selection rule could appear only in a separate official notice. Retrieval quality measures which material enters the context; claim anchoring measures whether the provided evidence actually establishes what the answer says.
Separate failure labels. A missing source is a retrieval or source-inventory problem. A retrieved source that does not contain the claimed sentence is a citation or answer-generation problem. A source that contains similar wording but governs a different year is an applicability problem. A source that contradicts another current official notice is an unresolved evidence conflict. One generic hallucination label hides these different repairs.
An agent can be helpful while abstaining. If the current official source cannot be fetched, the answer might explain the general process and clearly say which precise current criterion remains unverified. Citing a historical guide as current simply because it is the only accessible page would be misleading. The reader benefits from an honest boundary more than a complete-looking answer fabricated from partial evidence.
Return to contents or continue to Tables and numbers need exact row context.
7. Tables and numbers need exact row context
A cited number can be correct for the wrong category. A table may list a percentage for one school cohort beside a different cohort, or show totals and rates in adjacent columns. An extraction system that misaligns a row can place a genuine number next to an unsupported label. Citation anchoring should identify the table, row, column heading and effective version needed to interpret the figure.
Imagine a fictional report showing 18 correct answers among 24 attempts. The numerical claim 75% can be verified by 18 ÷ 24 = 0.75. But a sentence saying this is the school’s overall Mathematics pass rate would be false if the table concerns one exercise for a single learner. The mathematics and the population definition are independent parts of the claim.
Units deserve attention. A source might state a duration in minutes while a model paraphrases it in hours incorrectly. A financial table can show thousands of dollars but the answer states the number as single dollars. A citation verifier should compare quantity, unit, denominator and scope, with independent arithmetic where appropriate. Accurate copying of one numeral does not certify the whole inference.
Return to contents or continue to Current sources can change after a citation was created.
8. Current sources can change after a citation was created
A URL can continue working while its content changes. An article may be updated with a new deadline or a public information page replaced for the next year. An older answer’s citation might now point to a different passage, confusing later readers about what the agent actually saw. Record retrieval time and a source revision or suitable content fingerprint when the task requires auditable history.
Cached snippets need a freshness policy. A page checked last week may still be adequate for a stable explanation of algebra, but not for tomorrow’s school event. If the underlying source is superseded, a previously anchored statement may become historical rather than current evidence. The application should re-evaluate source applicability before a consequential new publication or recommendation.
Do not treat a changed source as permission to rewrite the entire answer without review. The issuer might add an exception that affects only one cohort; a blanket search-and-replace can create new errors elsewhere. Maintain a claim-to-source map so editors can identify the affected sentences, compare versions and apply a narrow authorised correction. The final page should be checked after the update.
Return to contents or continue to Evidence provenance is a chain, not a decoration.
9. Evidence provenance is a chain, not a decoration
A generated claim can be linked to an extraction activity that consumed a particular source passage. The extracted passage itself can be linked to a document revision and issuer. A reviewer may then know not only which URL appears beside the sentence but also what text the agent used and when it retrieved it. This is closer to an auditable evidence trail than a list of unspecific links at the end.
The chain can break at several points. A source document may be authentic but parsed incorrectly; a passage may be accurate but relevant to another year; a correct passage may be overstated by the generated claim; or a final article’s hyperlink may point to the wrong URL after editing. Verification should inspect each material transition according to consequence, not assume one stored provenance field resolves them all.
Provenance must be protected. Private records should not be indiscriminately copied into shared citation ledgers, and access to a source may change later. An authorised reviewer can inspect relevant evidence through the appropriate service; a public reader should receive only the information they are entitled to see. Source precision is not a permission to expose another parent’s confidential report.
Return to contents or continue to An invented one-hundred-claim evidence laboratory.
10. An invented one-hundred-claim evidence laboratory
Build a fictional Cedar Evidence Desk audit with one hundred material claims. Sixty-one have direct support from an applicable current passage. Eighteen are only partially supported, often because a source establishes one condition but not the full sentence. Eleven cite the wrong effective revision or cohort. Six link to a genuine but mismatched passage, and four have no retrievable source at the cutoff. The exclusive categories total 61 + 18 + 11 + 6 + 4 = 100.
The directly grounded fraction is 61%. Partial support is 18%, wrong-version evidence 11%, passage mismatch 6% and unavailable source 4%. These values are invented educational arithmetic, not observed AI citation accuracy. A sentence with a partly relevant citation cannot automatically be counted as fully supported, particularly when the missing clause is a deadline or eligibility condition a family could act upon.
A single unsupported material guarantee can be more consequential than many well-supported background sentences. The purpose of the ledger is to identify which assertions need narrowing, new evidence or removal, not to produce an attractive total. A safe system can preserve a useful explanation while withholding the one unverified current rule until the responsible source is checked.
Return to contents or continue to Six material claims can form one final answer.
11. Six material claims can form one final answer
Imagine a fictional answer about subject selection containing six material claims: the school offers the subject, it runs at a given level, it has a stated prerequisite, applications close on a certain date, an exception applies to one group, and a particular university pathway is guaranteed. Suppose four are directly supported, one has partial evidence and one is unsupported. Direct claim coverage is 4 ÷ 6 ≈ 66.67%.
The numerical fraction does not mean the answer is safe to publish unchanged. The unsupported guarantee might be the most consequential sentence. The appropriate edit could remove the guarantee, qualify the partial prerequisite and retain the four supported facts. Once corrected, a new check can evaluate the revised claims rather than merely moving the old label from unsupported to supported.
Do not conflate a claim with a citation count. Two links repeating the same general overview might both support the first claim but provide nothing for the deadline. A useful matrix has one row per claim and records each source’s actual supporting passage, scope and result. The user can then see exactly what is known and what needs a current institutional check.
Return to contents or continue to Evidence conflicts require source-scope reasoning.
12. Evidence conflicts require source-scope reasoning
Two documents can appear to disagree when they actually describe different cohorts or effective periods. One school page may govern current Secondary 3 selection while another governs a historical programme. Citation anchoring can reveal the difference by linking each claim to its proper source segment. The agent should not resolve the conflict by counting which position appears in more retrieved links.
If two genuinely current authoritative notices still disagree, a model-generated tie-breaker is not evidence. The responsible issuer may need to clarify the rule. Preserve the conflicting passages, their revisions and what interpretation remains uncertain. A reader-facing answer can explain the open issue and avoid a high-consequence recommendation until the correct authority is established.
The companion Conflicting Evidence guide treats source precedence and safe abstention in depth. This chapter’s narrower contribution is making each competing assertion traceable to its exact paragraph or table row, so later review can assess whether the disagreement is real.
Return to contents or continue to Citation locators must survive changing page layouts.
13. Citation locators must survive changing page layouts
Web pages can insert banners, rearrange sections or render content through client-side components. A paragraph number based only on visual position may no longer identify the intended text. PDF editions can shift page numbers after additional front matter. Store document identity and version with a content excerpt or robust selector, and provide a usable fallback route to the original source.
Do not guess an HTML anchor because its words sound like a heading. Validate the actual link or selector in the current document. A citation pointing to a section that has moved is less useful to readers, even when the statement remains supported elsewhere. Broken or misdirected fragments should be recorded for repair rather than silently accepted as harmless formatting.
Some publishers restrict direct deep links or require authentication. A system should respect access rules and describe the source limitation rather than invent a permalink. If the evidence is accessible only to authorised users, its citation should not expose private text in a public article. A reviewer can use an appropriately protected source reference to verify the statement.
Return to contents or continue to Correct a citation without falsifying the original record.
14. Correct a citation without falsifying the original record
Suppose the school updates a notice and the article’s old link now opens a revised document. An editor discovers that the existing sentence was based on the previous version. Rather than quietly relabelling the original claim as verified against the new page, preserve the historical provenance and review whether the sentence needs correction. A new published version should cite the applicable current passage.
Versioned article edits should avoid lost updates. Another editor might already have fixed a related exception or link. Use the site’s supported content revision or optimistic-lock mechanism and check the saved result. A citation repair that overwrites a colleague’s current factual correction would create a second error while appearing to solve the first.
A correction note can be useful when readers may have acted on the previous statement. It can state which fact changed, the applicable effective date and where the current source can be inspected. The assistant should not invent a reason for the issuer’s change or claim that every old copy has been deleted. Factual repair and transparent history support trust.
Return to contents or continue to Practical citation review before publication.
15. Practical citation review before publication
Begin with the draft’s material assertions. For each, record the exact source and passage, responsible issuer, effective scope and whether the claim is directly supported, partially supported, contradicted or unverified. Use independent arithmetic for numeric inferences. Narrow overbroad wording rather than covering it with additional irrelevant citations. A concise well-supported answer is more useful than a long unsupported one.
Before making content public, check that every hyperlink points to the intended source and that no current-source claim relies solely on an obsolete cached excerpt. Confirm the published version actually saved the corrected text and source links. The model may prepare a readable explanation, but the final state belongs to the content-management system and the facts belong to the primary sources.
Students can practise by finding four relevant sentences in a fictional notice and writing one claim each can truly support. Then add an exception that changes one claim. The lesson is not to attach more links; it is to match the strength and scope of a conclusion to the evidence a second reader can independently inspect.
Return to contents.
17. A reviewer rubric should classify exact support
Use a clear scale that differentiates directly supported, partially supported, contradicted, wrong scope and no accessible evidence. A directly supported claim must match its source’s subject, action, numerical value, conditions and relevant time. A partially supported claim may have one established clause but still lack another. A contradictory source is different from an inaccessible source whose content has not been checked, so these categories need distinct next actions.
Give reviewers examples before measuring agreement. A school notice saying students ‘may apply if they satisfy conditions A and B’ does not support a statement that students will be admitted. Another example can contain an exact numerical amount but in the wrong column or reporting year. The rubric should explain why each claim fails, enabling consistent correction rather than a simplistic similarity judgment.
Some evidence remains genuinely ambiguous. A passage may contain undefined terms or cross-reference a missing appendix. The right outcome can be Unknown or qualified review, not a guess that the most likely interpretation must be correct. Preserve the unresolved source condition so a later official clarification can change the assessment without falsifying the earlier label.
Return to contents to revisit the cited claims.
18. Source fingerprints help distinguish editions
A short content hash or provider revision ID can help identify which edition of a document was retrieved. A page may keep the same public URL while changing the effective deadline. Recording the original evidence passage and version means the reviewer can later tell that the text was correct under one revision but no longer governs a current reader’s decision. The hash is an identity aid, not a proof of factual authority.
An exact hash match says the bytes matched a trusted reference at the checking point; it cannot establish the source issuer was qualified or the interpretation was sound. Similarly, a valid hyperlink can still lead to an obsolete version or a page whose publisher has changed. Keep issuer, effective date and source contract beside the fingerprint when those facts matter.
A citation system should avoid storing full copies of private documents in every article or log merely to preserve reproducibility. Where an authorised reviewer can inspect a protected source by stable ID, retain the minimum necessary excerpt or reference under suitable access controls. A public response need not expose a learner’s confidential report for a provenance chain to remain auditable to responsible staff.
Return to contents to revisit the cited claims.
19. Recover from broken fragments and reflowed PDFs
A text fragment or page anchor may cease to work after a website redesign, especially when headings are renamed, paragraphs moved or a dynamic renderer changes content. The saved citation record should identify the original document and a short evidence quote, enabling a reviewer to locate the claim even when the first deep link breaks. A model should not invent an anchor from a guessed heading.
PDF page numbers present their own problem. A new cover page can shift the displayed page labels, while the source passage remains substantively the same. A citation must state which edition was checked and which table, heading or paragraph the claim depends upon. When available, use the document’s own labels rather than assume zero-based screenshot indices correspond to reader-facing pages.
Repair a citation by checking the actual current source and preserving the meaning of the claim. Do not simply point a broken citation to the document’s home page and call the claim verified. If the current edition contradicts the earlier statement, the article itself may need correction under editorial approval and a fresh saved-state check. Link repair and factual revision can be different kinds of work.
Return to contents to revisit the cited claims.
20. Fix the hundred-claim laboratory without moving the denominator
The invented initial audit contains sixty-one directly supported claims, eighteen partly supported, eleven citing the wrong applicable revision, six linked to the wrong passage and four without accessible source evidence. Suppose later review corrects ten partial claims by narrowing their wording and four wrong-version claims by finding the applicable current notice. The direct-support count rises from sixty-one to seventy-five.
Remaining categories become eight partial, seven wrong-version, six mismatched and four unavailable. Check the sum: 75 + 8 + 7 + 6 + 4 = 100. The directly supported proportion is now 75%, compared with 61% at the original cutoff. This is an explicit revised evaluation of the same fictional hundred claims, not a random production benchmark or evidence that every corrected sentence is permanently current.
Do not inflate results by deleting the hardest claims from the denominator while keeping the original count in the headline. If an unsupported statement is removed from the final article, the revised article has a changed claim population and should be evaluated under its actual new scope. A historical correction report can still preserve the original hundred-case audit for process improvement. Both views are useful when properly labelled.
Return to contents to revisit the cited claims.
21. One unsupported guarantee can outweigh many background facts
A parent may read an answer containing several accurate descriptions of subject options and one unsupported promise about admission. The many supported statements cannot compensate for that promise when the family uses it to choose an education pathway. Citation evaluation needs non-compensatory gates for material deadlines, prerequisites and outcomes, rather than one averaged percentage that rewards adding harmless factual sentences.
Write final claims no stronger than the sources. If a notice says that eligible students may be considered, report consideration under the stated criteria rather than certainty of placement. If an official issuer has not released the future intake’s details, describe the current known pathway and clearly mark the pending update. A calm limitation is more helpful than a fabricated assurance.
Review factual scope independently from reading quality. A statement can be accurate but too technical for a student, or easy to read but wrong about effective dates. The ideal article meets both: it explains a supported claim in plain language and lets a reader inspect the relevant issuer’s current passage. The purpose is informed decisions rather than decorative scholarship.
Return to contents to revisit the cited claims.
22. Citation metrics must not reward link proliferation
Counting ten hyperlinks in a paragraph says little about whether its three material statements are justified. The links may all lead to one syndicated source or none may contain the claimed condition. A useful metric measures directly supported material claims within the accepted article scope, and tracks unsupported, partial and unavailable claims separately. It can also measure whether anchors are still resolvable under the source versions reviewed.
Source diversity should follow independent issuers and evidence roles, not sheer URL count. A school’s specific subject table and a ministry’s broader qualification framework may play different complementary roles. Five mirrors repeating one table do not create five independent confirmations. A provenance ledger can record original issuer and derivation so a reviewer sees the evidence’s actual breadth.
Over-citation can burden readers. For a short stable algebra definition, one appropriate authoritative source and a clear explanation may be enough. For a current admissions rule with several exceptions, multiple exact passages may be necessary. The target is sufficient justified evidence for the user’s decision, not the maximum number of blue links that fits in the page.
Return to contents to revisit the cited claims.
23. A changed-context school notice changes citation validity
Imagine a fictional school’s revision 4 notice states an application window for one intake, and revision 5 later replaces it for the following cohort. A historical question about revision 4 may still be answered accurately using that document, while a current application question requires revision 5. A citation’s factual words can be correct and still inapplicable to the user because its effective context is wrong.
Now suppose an agent copied a correct figure from a PDF table but the paragraph referred to the wrong subject row. The passage text and number may both be genuine, yet the association is false. The repair is to inspect row and column headings and the correct issuer, not to add another general link. A truthful answer may need a corrected number and a new table-specific citation.
Change the permission condition: the evaluator has access to a private staff note that a public reader cannot view. The public article must not quote it merely to improve evidence coverage. Find an authorised public source or state the limitation and provide an appropriate official inquiry route. Source precision does not override confidentiality or legitimate reader access.
Return to contents to revisit the cited claims.
24. A publication checklist tied to actual evidence
Before publishing, extract material claims and map each to one or more actual source passages. Check source identity, applicable version, issuer, date and conditions. Validate numerical transformations separately; compare paraphrases against the source’s modality and quantifiers. Mark unresolved exceptions explicitly rather than bury them under a vague citation. This is a disciplined editorial process a careful human can understand and review.
Inspect the final saved article, not only the draft. A content editor might normalize markup, alter a fragment link or strip unsupported metadata. Confirm the links open the intended sources, the key text remains accurate and the page has the intended status. A model’s statement ‘all references are checked’ is not a substitute for these observations.
For learners, try a six-claim fictional answer and ask which statements its two sources actually support. Remove one unsupported guarantee, clarify a partial prerequisite and then evaluate the revised answer under the new claim count. They will learn that trustworthy writing is not simply having references: it is making precisely the claims the evidence can carry.
Return to contents to revisit the cited claims.
- 17. A reviewer rubric should classify exact support
- 18. Source fingerprints help distinguish editions
- 19. Recover from broken fragments and reflowed PDFs
- 20. Fix the hundred-claim laboratory without moving the denominator
- 21. One unsupported guarantee can outweigh many background facts
- 22. Citation metrics must not reward link proliferation
- 23. A changed-context school notice changes citation validity
- 24. A publication checklist tied to actual evidence
Frequently asked questions
Does a real citation link prove a claim?
No. The linked source must contain an applicable passage that actually supports the exact assertion.
What is a passage anchor?
A locator such as a section, paragraph, page, table row or text selector identifying the relevant part of the cited source.
Can an exact quote still be misleading?
Yes. Removing conditions or effective-date context can change the meaning of a perfectly copied fragment.
Why record source revisions?
Because a document can be updated while its URL remains the same, changing which claims it supports.
Does a high retrieval score establish a grounded answer?
No. Ranking and coverage concern retrieved evidence; claim-to-passage checking concerns how that evidence was used.
How should the assistant handle an unsupported current rule?
Qualify or withhold that claim and provide a route to verify it with the responsible current authority.
Source and standards notes
Sources reviewed 24 October 2026. The Cedar Evidence Desk, its hundred claims and all school examples are invented for teaching, not real citation accuracy measurements.
- W3C PROV-O: Entities, activities and agents used to describe provenance chains.
- W3C Web Annotation Data Model: Target selectors including passage and text-quote references.
- MDN Text Fragments: Browser-supported text-targeting links and their constraints.
- Microsoft Azure Retrieval Evaluation: The distinction between relevant source retrieval and final answer quality.
- NIST AI RMF: Measure: Evidence, evaluation validity and limits of automated assessments.
Further learning
Continue through Super Intelligence curriculum; How X Works hub; Conflicting Evidence; Retrieval Coverage; Evidence Ledgers; Concurrent Edits; Gold Standard of Education. Useful citations connect specific justified claims to source passages that real readers can inspect.