Preview candidate · independent learning, accuracy, brand and accessibility review has not passed. Do not treat this page as an approved or published LAiDIES book.

The LIBRAiRY · Reference Shelf · candidate under review

How to Check AI’s Work

You do not need to redo every word. You need to find the parts that could change a decision, then check those properly.

An AI turns meeting notes into a polished client update: “The client approved the revised launch date, expects a 20% lift and wants paid social added to phase one.” The notes actually say the client will review the date on Friday, somebody guessed “maybe 20%,” and paid social was only a question.

Which part deserves attention first?

The Receipt Loop

  1. Frame
  2. Split
  3. Find
  4. Inspect
  5. Cross-check
  6. Decide
  7. Record

The label is a mnemonic. The mechanism is the claim-to-evidence decision process taught below.

Chapter 1 of 7

Sort the output

Fluency is presentation. Your first job is to identify what kind of thing each line is and what could change a decision.

Four useful buckets

A draft is proposed wording, structure or creative material. A factual claim says something checkable about the world. An assumption or inference connects facts without being directly stated. A recommendation proposes action and may depend on both facts and values.

A sentence can contain more than one bucket. Split compound statements before you search so one true clause cannot lend credibility to a false neighbour.

  • Draft: “Here is a warmer opening for the email.”
  • Claim: “The contract renews on 1 September.”
  • Inference: “The delay probably means the client is unhappy.”
  • Recommendation: “Move the campaign to September.”

Turn the stakes dial

Ask what happens if the statement is wrong, how reversible the action is, how uncertain the evidence is and how quickly the answer can change. A playlist caption and a medication decision do not deserve the same checking plan.

Check decision-changing claims first: money, deadlines, permissions, safety, public promises, people’s rights or anything hard to undo.

Classify all eight lines before checking your reasoning.

Chapter 2 of 7

Know what the tool actually did

Search, grounding, citations and provenance can help. None of them is the same thing as verification.

Six different jobs

Generation
Creates the answer. Smooth language is not evidence.
Retrieval or search
Finds material. Finding a page does not make it the right page.
Grounding
Connects or constrains an answer to supplied or retrieved material. The material can still be weak, stale, out of scope or misread.
Citation
Points to a source. The source must exist, open and support the nearby claim.
Provenance
Records an asset’s origin or change history. It can help authenticate a trail without proving that the depicted event or assertion is true.
Verification
Connects a specific claim to adequate evidence under the right scope, date and stakes, then makes a qualified decision.

The chain

A source may be retrieved, summarized by a model and represented by a citation. The human check still asks whether the cited source is the right source and whether its exact content supports the exact claim.

The same model asking itself “are you sure?” may expose inconsistency, but it is not independent evidence. Independence comes from the underlying authoritative record, a qualified owner or a genuinely separate evidence path.

Match five artifacts to what they establish.

Chapter 3 of 7

Frame the claim before searching

A vague question retrieves vague evidence. Make the statement small enough to test.

Write the claim you actually need

Name the product or person, plan or version, place or jurisdiction, date, conditions, denominator and intended decision. Preserve qualifiers such as “may,” “in this sample” and “for eligible accounts.”

“She doesn’t even go here” is a useful scope reminder: evidence about a different plan, country or version may be real and still not apply. The film reference is only a mnemonic; it does not test the evidence for you.

  • Vague: “It is 50% faster.”
  • Checkable: “In the named study, people completing which task, compared with what baseline, finished 50% faster under which conditions?”
  • Vague: “The plan is free.”
  • Checkable: “Is this feature included in the provider’s free plan for Canadian individual accounts on 25 July 2026?”
Choose the repair that makes each claim checkable.

Chapter 4 of 7

Choose and inspect evidence

There is no universal source ladder. Match evidence to the claim, then inspect what it actually says.

Claim-matched evidence

Claim typeBest starting evidenceNever sufficient alone
Current feature, price, plan or policyCurrent official documentation, account UI or contract; record region, plan and access dateAI memory, search snippet or undated screenshot
Law, regulation or public ruleCurrent legislation, regulator, court or official authority; qualified professional for consequential applicationGeneric blog or a non-jurisdictional answer
Scientific or empirical resultOriginal paper, dataset and method; correction/retraction status; suitable expert synthesisPress release or percentage without method/denominator
Quote, statement or eventOriginal transcript, recording, filing, document or direct statementQuote graphic, repost or citation without accessible source
Internal commitmentApproved record, contract or named decision ownerModel-written recap or unattributed notes
Media originOriginal asset/context plus provenance where presentMetadata treated as a truth verdict

The seven source questions

  • Identity: who created or published it, and what authority or interest do they have?
  • Relevance: does it address this exact claim rather than a nearby topic?
  • Scope: correct person, product, plan, place, jurisdiction and conditions?
  • Date/version: current enough for this decision?
  • Support: does it actually entail the claim, numbers and qualifiers?
  • Method: how was the evidence produced; what is missing or uncertain?
  • Independence: does corroboration trace to a genuinely separate origin?
You need today’s seat limit for a paid software plan in Canada. Compare all three, choose the most diagnostic source, then explain why and what you would inspect.

Chapter 5 of 7

Run the Receipt Loop

Frame → Split → Find → Inspect → Cross-check → Decide → Record.

A worked decision

Return to the client update. Frame and split three claims: approval of the date, expected 20% lift and paid-social scope. Find the meeting record and named decision owner. Inspect the exact lines. The notes support none of the three as written.

The useful draft survives, but the claims change: “The client will review the proposed launch date Friday. The team discussed a possible lift but has not agreed a forecast. Paid social is an open scope question.”

  • SUPPORTED: adequate evidence supports the framed claim.
  • CONTRADICTED: adequate evidence conflicts with it.
  • UNRESOLVED: evidence is missing, mixed or too weak.
  • NOT APPLICABLE: the evidence addresses a different scope, version or question.

Record the decision receipt

Keep the exact claim, source, location, access date, scope, verdict, uncertainty and next review trigger. That turns a private moment of checking into a correction path.

Do not force a true/false answer when the honest result is unresolved. Abstaining from invented certainty is a useful outcome.

Complete the decision receipt before revealing the model answer.
ClaimSourceYour verdict
Ali approved the launch.Meeting notes: “Ali will decide Friday.”
The campaign will lift results 20%.Meeting notes: “Maybe 20%?” No forecast source.
Paid social is in phase one.Meeting notes list paid social as an open question.

Chapter 6 of 7

Currentness, media and independent checks

A real source can be stale. A real provenance trail can accompany a false claim. A second answer can repeat the first answer’s blind spot.

Stale but real

A page may accurately describe an old model, plan, price or law and still be unusable for today’s decision. Record publication or version date, access date, region and change trigger.

Current provider examples belong in dated, replaceable cards—not inside the durable definition of verification.

Trace the origin

Five articles may repeat one press release. Corroboration requires separate provenance or evidence, not merely separate URLs.

For images, audio and video, provenance can help establish who made or changed an asset. Investigate the event, context and claim separately.

Resolve both cases and explain the evidence gap.

Chapter 7 of 7

Use, qualify, escalate or stop

The goal is a better decision—not infinite research or performative certainty.

Proportionate checking

  • Low and reversible: label it draft; one quick source inspection may be enough.
  • Material but reversible: use a claim table, authoritative source, scope/date check and independent corroboration where uncertainty remains.
  • High-stakes or hard to reverse: use authoritative records and the qualified human owner or professional. Stop or escalate when evidence is inadequate.

Write the honest answer

State what is supported, what remains uncertain, the scope and date, and what evidence would change the conclusion. A disclaimer does not repair bad advice.

This Rulebook is not legal, medical, financial, HR or safety advice. In those domains, it can help you organize questions and sources, but it cannot replace the responsible professional or decision owner.

Choose the proportionate next action for all three stakes levels.

The source-bound prompt

Use this to reduce risk and expose missing evidence. It is not a truth machine.

Now open the source yourself. This prompt cannot certify the source, the quote or the conclusion.

The unseen transfer challenge

A board memo says: “Tool A is the cheapest option for us, and a new study proves teams finish analysis 50% faster.” The linked price is for a different region and plan. The memo links a news story that mentions 50% but does not identify the task, baseline, sample or original study.

  1. Split the material claims.
  2. Identify the decision-changing claim.
  3. Choose the most diagnostic evidence action.
  4. Name the scope and freshness checks.
  5. Give a qualified verdict and state what would change it.

Review, freshness and corrections

Published
Not published — candidate under independent review
Last reviewed
2026-07-25 · maker review only
Next review
Before any release, then annually for the durable method and quarterly for any named provider example.
What can change
Provider features, plans, policies, interfaces, laws, statistics, research corrections and external links. The core checking method is reviewed annually or when learner evidence shows confusion.

Report a correction about a chapter, claim or source. Include the relevant ID and what you found; do not email private source material.

Canonical source: /content/library-books/verification-rulebook.json
Source SHA-256: ac3ec15f27898ad69c9ac142a86e51948b5d024437b89dcd11cf336ad81fdacd

Source drawer

Every material claim below is generated from the same canonical record as the reader wording. A working link does not, by itself, prove semantic support or currentness.

Open the claim ledger

  1. VR-C001 · SUPPORTED

    Generative AI can produce false content and fabricated citations, so important sources and citations still need review and verification.

    Sources: SRC-NIST-GAI-600-1

    Support: NIST.AI.600-1, pp. 9 and 34: confabulation includes confidently false content and fabricated citations; suggested actions require review and verification of sources and citations.

    Version/access: NIST AI 600-1, July 2024 · 2026-07-25

    Scope: Generative-AI risk management; not a performance rate or guarantee for a named model.

    Qualification: Sourced fact, scoped to the NIST risk profile.

    Next review: 2027-07-25 or on NIST revision

    Report a correction for VR-C001

  2. VR-C002 · QUALIFIED

    Current AI products can search the web and present citations; those are present capabilities, not hypothetical future abilities.

    Sources: SRC-OAI-WEB-SEARCH

    Support: Web search guide, Overview and Citations sections: web search can access up-to-date information and responses include URL citation annotations.

    Version/access: Current web documentation · 2026-07-25

    Scope: OpenAI API web-search tool; availability and implementation depend on current provider documentation.

    Qualification: Provider documentation establishes the provider-described capability, not comparative accuracy.

    Next review: Before release and 2026-10-25

    Report a correction for VR-C002

  3. VR-C003 · QUALIFIED

    Search grounding can connect generated answers to current web material and return citation annotations.

    Sources: SRC-GOOGLE-GROUNDING

    Support: Grounding with Google Search, How it works and Grounding metadata sections: search, processing, grounded response and citation metadata.

    Version/access: Current web documentation · 2026-07-25

    Scope: Gemini API grounding feature; does not establish source quality, entailment or universal truth.

    Qualification: Provider documentation establishes the provider-described operation only.

    Next review: Before release and 2026-10-25

    Report a correction for VR-C003

  4. VR-C004 · SUPPORTED

    Grounding, direct quotations, citations and permission to express uncertainty can reduce hallucination risk, but they do not eliminate it; critical information still needs validation.

    Sources: SRC-ANTHROPIC-HALLUCINATION

    Support: Reduce hallucinations, Allow Claude to say “I don’t know,” Use direct quotes, Verify with citations, and Limitations sections.

    Version/access: Current web documentation · 2026-07-25

    Scope: Risk-reduction guidance, not comparative model evidence or a guarantee.

    Qualification: Provider guidance plus its explicit limitation.

    Next review: Before release and 2026-10-25

    Report a correction for VR-C004

  5. VR-C005 · QUALIFIED

    A citation is a pointer to a source; verification additionally checks the source’s identity, relevance, scope, date and support for the specific claim.

    Sources: SRC-NIST-GAI-600-1 · SRC-COR-ABOUT

    Support: NIST pp. 9 and 34 require source/citation review; COR About names source identity, evidence and other-source questions. The five-part formulation is LAiDIES’ synthesis.

    Version/access: NIST AI 600-1; current COR About page · 2026-07-25

    Scope: Instructional verification model, not a universal legal standard.

    Qualification: Qualified LAiDIES synthesis; not attributed verbatim to either source.

    Next review: 2027-07-25 or on learner-confusion evidence

    Report a correction for VR-C005

  6. VR-C006 · SUPPORTED

    Lateral reading means leaving an unfamiliar page to investigate who is behind it and what trusted sources say about it.

    Sources: SRC-COR-LATERAL

    Support: Intro to Lateral Reading, opening lesson description: defines the strategy as leaving the webpage and opening a new tab to see what trusted websites say about the unknown source.

    Version/access: Current COR lesson page · 2026-07-25

    Scope: Evaluating unfamiliar online sources; it is one verification action, not the entire Receipt Loop.

    Qualification: Directly supported by the named Digital Inquiry Group lesson.

    Next review: 2027-07-25 or on lesson revision

    Report a correction for VR-C006

  7. VR-C007 · SUPPORTED

    C2PA provenance can bind claims about an asset’s source and history so tampering can be detected; it does not decide whether the content or provenance is good, bad or true.

    Sources: SRC-C2PA-2-4

    Support: C2PA 2.4 Content Credentials specification §§1.2 Scope, 1.3 Technical Overview and 2.3 Core Aspects; the trust model provides tamper-evident provenance and does not make value judgments.

    Version/access: C2PA Specifications 2.4 · 2026-07-25

    Scope: C2PA-compatible Content Credentials; not every metadata or authenticity system.

    Qualification: Provenance and tamper evidence are not a truth score.

    Next review: On specification revision or 2027-07-25

    Report a correction for VR-C007

  8. VR-C008 · SUPPORTED

    At 320 CSS pixels, the rendered candidate must reflow without loss of information or functionality and without two-dimensional scrolling except where an exception applies.

    Sources: SRC-WCAG-22

    Support: WCAG 2.2 Success Criterion 1.4.10 Reflow.

    Version/access: WCAG 2.2 · 2026-07-25

    Scope: Rendered web reader acceptance; conformance requires more than this single criterion.

    Qualification: Acceptance requirement, not a claim that this candidate already conforms.

    Next review: On W3C errata or Recommendation change

    Report a correction for VR-C008

  9. VR-C009 · SUPPORTED

    The Receipt Loop uses four evidence verdicts: supported, contradicted, unresolved and not applicable.

    Sources: INTERNAL-ECO01-SPEC

    Support: ECO-01 build packet, correct mental model and chapter 5.

    Version/access: 2026-07-25 · 2026-07-25

    Scope: LAiDIES teaching taxonomy, not a claim that every external framework uses these labels.

    Qualification: Internal product rule.

    Next review: After newcomer study or 2027-07-25

    Report a correction for VR-C009

  10. VR-C010 · QUALIFIED

    Checking effort should be proportional to consequence, reversibility, uncertainty and freshness.

    Sources: INTERNAL-ECO01-SPEC

    Support: ECO-01 build packet, Stakes dial and proportionality contract.

    Version/access: 2026-07-25 · 2026-07-25

    Scope: General verification triage; does not replace professional standards or regulated duties.

    Qualification: Internal instructional recommendation pending newcomer evidence.

    Next review: After independent safety/instructional review

    Report a correction for VR-C010

  11. VR-C011 · QUALIFIED

    Separate URLs are not independent corroboration when they all trace to the same underlying source.

    Sources: SRC-COR-ABOUT · SRC-COR-LATERAL

    Support: COR asks what other sources say and teaches leaving a page to investigate an unknown source; the origin-tracing formulation is LAiDIES’ synthesis.

    Version/access: Current COR pages · 2026-07-25

    Scope: Evidence independence; distinct reporting may still add independent analysis.

    Qualification: Qualified synthesis, not a direct quotation or universal rule.

    Next review: 2027-07-25

    Report a correction for VR-C011

  12. VR-C012 · SUPPORTED

    The source-bound prompt is a risk-reduction aid, not a truth machine, and critical claims still require external validation.

    Sources: SRC-ANTHROPIC-HALLUCINATION

    Support: Reduce hallucinations, Limitations: techniques reduce but do not eliminate hallucinations; critical information needs validation.

    Version/access: Current web documentation · 2026-07-25

    Scope: Prompt use with generative AI; no prompt can certify arbitrary sources or conclusions.

    Qualification: Safety limit retained next to the prompt.

    Next review: Before release and 2026-10-25

    Report a correction for VR-C012

  13. VR-C013 · SUPPORTED

    The Rulebook is not legal, medical, financial, HR or safety advice and cannot replace the responsible professional or decision owner.

    Sources: INTERNAL-ECO01-SPEC

    Support: ECO-01 build packet, stakes dial and safety acceptance test.

    Version/access: 2026-07-25 · 2026-07-25

    Scope: LAiDIES product boundary.

    Qualification: Internal safety rule.

    Next review: Every release and on safety incident

    Report a correction for VR-C013

  14. VR-C014 · SUPPORTED

    A real source can accurately describe an old state and still be inapplicable to a current decision.

    Sources: INTERNAL-ECO01-SPEC

    Support: ECO-01 build packet, misconception correction and freshness contract.

    Version/access: 2026-07-25 · 2026-07-25

    Scope: Distinguishes historical accuracy from present applicability.

    Qualification: Internal instructional principle.

    Next review: 2027-07-25

    Report a correction for VR-C014

Open all source records
  1. NIST AI RMF: Generative Artificial Intelligence Profile · NIST Confabulation, fabricated citations, source verification and ongoing monitoring.
  2. Reduce hallucinations · Anthropic Grounding, quotation and citation techniques reduce but do not eliminate hallucinations.
  3. Grounding with Google Search · Google Current example distinguishing search grounding and citation annotations.
  4. Civic Online Reasoning — About · Digital Inquiry Group The source-identity, evidence and other-source questions.
  5. Intro to Lateral Reading · Digital Inquiry Group Defines lateral reading as leaving an unfamiliar page and consulting trusted sources in another tab.
  6. C2PA Content Credentials specification 2.4 · C2PA Full current specification containing §§1.2, 1.3 and 2.3 for provenance and tamper-evident source/history limits.
  7. Web Content Accessibility Guidelines (WCAG) 2.2 · W3C Accessibility requirements for the rendered reader candidate.

Where to go next

  • Episode 3 · The Burn Book Problem See the verification risk made memorable through story.
  • Dream Phone · evidence practice · HOLD Practice is still under source review; treat its links as review leads, not final authority.
  • NewsStand See timely reporting under the NewsStand’s separate sourcing and correction rules.