Chapter 19- Claims — Mo Farahani

Chapter Goal

In Chapter 18, we established the relationship:

Corpus
↓
Retrieval
↓
Retrieved Passage
↓
Relevant Context
↓
Evidence

We also introduced a simple definition of a Claim:

A Claim is a statement in the AI's answer that can potentially be supported or contradicted by evidence.

Chapter 19 now examines that concept directly.

The central question is:

What exactly is the AI saying, and does the available Evidence really support it?

Consider:

Evidence:
Enterprise customers may request a refund
within 14 days of the original purchase.

The AI generates:

Enterprise customers can request a refund
within 14 days.

The generated statement is a:

Claim

The relationship is:

Evidence
↓
supports
↓
Claim

But this relationship is not always correct.

For example:

Evidence:
Enterprise customers may request a refund
within 14 days.

AI Claim:
Enterprise customers can receive a full refund
within 30 days.

The Claim contains information that the Evidence does not support.

This chapter explains how to reason about that relationship.

By the end of this chapter, you should understand:

what a Claim is;
who is making the Claim in the pipeline studied in this course;
how one answer can contain multiple Claims;
what Claim Extraction means;
what Claim Generation means;
what the Claim–Source Relationship is;
what Source Support means;
what Evidence Support means;
why source presence does not automatically mean claim support;
what Natural Language Inference is conceptually;
what Entailment means;
what Contradiction means;
what Neutral means;
what an Unsupported Claim is;
and why a strong Claim should be entailed by its Evidence.

The central model of this chapter is:

Evidence
↓
Claim

More precisely:

Evidence
↓
Does the Evidence support the Claim?
↓
Entailment / Contradiction / Neutral

19.1 Claim Fundamentals

1. What Is a Claim?

A Claim is a statement that asserts something about the world.

For example:

The Enterprise refund period is 14 days.

This statement asserts something that can potentially be:

true
false
supported
unsupported
contradicted

That makes it a Claim.

Other examples include:

Company X was founded in 2018.

Restaurant A is located in Amsterdam.

The Standard plan costs less than the Enterprise plan.

Each statement asserts something that can be evaluated.

2. Who Makes the Claim?

The word Claim can be used broadly.

A source document can contain claims. A journalist can make claims. A user can make claims.

But in the pipeline studied in this part of the course, our main focus is:

Claim
=
statement produced in the AI's answer

For example:

User:
How long is the refund period?

Evidence:
Refund requests may be submitted within 14 days.

AI Answer:
The refund period is 14 days.

The statement:

The refund period is 14 days.

is the Claim we want to evaluate.

Source
↓
Passage
↓
Evidence
↓
LLM Generation
↓
Claim

3. Claim Is Not the Same as the Whole Answer

An AI answer can contain several Claims.

Consider:

Enterprise customers can cancel from Billing Settings,
and refunds are available within 14 days.

This contains at least two factual assertions:

Claim 1:
Enterprise customers can cancel from Billing Settings.

Claim 2:
Refunds are available within 14 days.

Answer
↓
may contain
↓
Multiple Claims

Different Claims may require different Evidence.

4. One Sentence Can Contain Multiple Claims

Consider:

Company X was founded in 2018,
is headquartered in Amsterdam,
and operates in 12 countries.

This sentence contains three Claims:

Claim 1:
Company X was founded in 2018.

Claim 2:
Company X is headquartered in Amsterdam.

Claim 3:
Company X operates in 12 countries.

One source may support Claim 1.

Another may support Claim 2.

There may be no Evidence for Claim 3.

Evaluating the sentence as one indivisible unit can therefore hide unsupported information.

5. Claims Can Have Different Levels of Specificity

Compare:

Claim A:
Refunds are available.

with:

Claim B:
Enterprise customers can receive a full refund
within 14 calendar days regardless of account usage.

Claim B asserts much more:

Enterprise customers
full refund
14 calendar days
regardless of usage

It therefore requires more specific Evidence.

More Specific Claim
↓
More Evidence Requirements

6. Small Wording Changes Can Change a Claim

Suppose the Evidence says:

Customers may request a refund within 14 days.

Compare:

Claim A:
Customers may request a refund within 14 days.

with:

Claim B:
Customers will receive a refund within 14 days.

These are not equivalent.

Evidence:
may request

Claim B:
will receive

The certainty and meaning changed.

Another example:

Evidence:
The service is available in some European markets.

Claim:
The service is available throughout Europe.

The Claim is broader than the Evidence.

Claim evaluation therefore depends on exact meaning, not merely similar wording.

7. What Is Claim Extraction?

Claim Extraction is the process of identifying factual Claims contained in a piece of text.

Suppose the AI answer says:

Company X launched in 2018,
is based in Amsterdam,
and currently serves customers in Germany.

Claim Extraction can identify:

Claim 1:
Company X launched in 2018.

Claim 2:
Company X is based in Amsterdam.

Claim 3:
Company X serves customers in Germany.

AI Answer
↓
Claim Extraction
↓
Claim 1
Claim 2
Claim 3

8. Why Claim Extraction Matters

Suppose the answer is:

Enterprise costs €500 per year,
includes priority support,
and provides a 30-day refund period.

Available Evidence supports:

€500 per year
priority support

but actually says:

refund period = 14 days

Claim Extraction reveals:

Claim 1:
Enterprise costs €500 per year.
→ supported

Claim 2:
Enterprise includes priority support.
→ supported

Claim 3:
Enterprise provides a 30-day refund period.
→ contradicted / unsupported

This is more precise than evaluating the whole answer as a single block.

9. Claim Extraction Does Not Create New Facts

Claim Extraction should identify what the answer already says.

It should not invent additional information.

Answer:

The refund period is 14 days.

Correct extraction:

Claim:
The refund period is 14 days.

Incorrect extraction:

Claim:
All customers receive a full refund within 14 days.

Claim Extraction
→ identify assertions

not

Claim Extraction
→ invent new assertions

10. What Is Claim Generation?

Claim Generation is the process by which the model produces factual statements as part of its answer.

For example:

Evidence:
Annual Enterprise subscriptions may be refunded
within 14 days.

The model generates:

Enterprise customers have a 14-day refund period.

The generated statement is a Claim.

Context / Evidence
+
User Question
↓
LLM Generation
↓
Claims

11. Generation Can Transform the Evidence

The model does not usually copy the source word for word.

Evidence:

Annual Enterprise subscriptions may be refunded
within fourteen calendar days of the original purchase.

Generated Claim:

Enterprise annual plans have a 14-day refund window.

The wording changed.

The underlying meaning remained similar.

Evidence Text
≠
Claim Text

This does not automatically mean the Claim is unsupported.

The important question is:

Does the meaning of the Evidence support the meaning of the Claim?

12. Claim Generation Can Introduce Unsupported Information

Evidence:

Annual subscriptions may be refunded within 14 days.

Generated Claim:

Annual subscriptions receive a guaranteed full refund
within 14 days.

The model introduced:

guaranteed
full refund

Neither is established by the Evidence.

This produces an Unsupported Claim, or at least unsupported components within the Claim.

Having correct Evidence in context therefore does not automatically guarantee supported Claims.

13. Generated Claims Can Combine Multiple Evidence Pieces

Suppose:

Evidence A:
Standard costs €300 per year.

Evidence B:
Enterprise costs €500 per year.

The model generates:

Standard is €200 cheaper per year than Enterprise.

The Claim depends on both Evidence pieces.

Evidence A
+
Evidence B
↓
Generated Comparison
↓
Claim

One Claim therefore does not always map to one Passage.

19.2 Claim–Source Relationship

14. What Is the Claim–Source Relationship?

The Claim–Source Relationship describes how a Claim in the generated answer relates back to the source information from which its support originates.

Consider:

Claim:
The Enterprise refund period is 14 days.

Its support may come from:

Source:
Company Help Center

Document:
Enterprise Refund Policy

Passage:
Enterprise subscriptions may be refunded
within 14 days.

Source
↓
Source Document
↓
Passage
↓
Evidence
↓
supports
↓
Claim

The Claim–Source Relationship allows the generated statement to be traced back to supporting information.

15. Source Presence Does Not Mean Claim Support

Suppose the AI uses:

Enterprise Billing Policy

The document is relevant.

But the actual Passage says:

Enterprise subscriptions renew annually.

The Claim says:

Enterprise refunds are available within 14 days.

The Source is related to Enterprise billing, but that Passage does not support the refund Claim.

Related Source
≠
Supporting Source

More precisely:

Source Presence
≠
Claim Support

16. What Is Source Support?

Source Support refers to whether information from a Source supports a Claim.

Claim:

The refund period is 14 days.

If a Source Document contains:

Refund requests may be submitted within
14 days of purchase.

then the Source provides support for the Claim.

Source
↓
contains supporting information
↓
Claim is source-supported

17. Source Support Is Usually Passage-Level

Suppose a manual is 100 pages long.

Only one sentence states:

Refund requests must be submitted within 14 days.

It would be imprecise to treat the entire manual as the direct supporting evidence.

A more precise model is:

Source
↓
Document
↓
Supporting Passage
↓
Evidence
↓
Claim

Passage-level support therefore matters greatly when evaluating Claims.

18. What Is Evidence Support?

Evidence Support describes whether specific Evidence actually supports the meaning of a Claim.

Evidence:

Enterprise customers may request a refund
within 14 days.

Claim:

Enterprise customers may request a refund
within 14 days.

The support is strong.

Now change the Claim:

Enterprise customers may request a refund
within 30 days.

The same Evidence no longer supports it.

Evidence
↔
Claim

Evidence Support concerns the relationship between the meanings of these two pieces of information.

19. Source Support vs. Evidence Support

Source Support

Does the Source contain information that supports this Claim?

Evidence Support

Does this specific Evidence support this Claim?

Source
↓
Supporting Passage
↓
Evidence
↓
Claim

Evidence Support is the more precise relationship.

20. A Source Can Support One Claim but Not Another

Suppose a Source says:

The Enterprise plan costs €500 per year
and includes priority support.

It supports:

Claim 1:
Enterprise costs €500 per year.

Claim 2:
Enterprise includes priority support.

But it does not support:

Claim 3:
Enterprise has a 30-day refund period.

Same Source
↓
Claim 1 → supported
Claim 2 → supported
Claim 3 → unsupported

Support must therefore be evaluated Claim by Claim.

21. One Claim Can Require Multiple Sources or Passages

Suppose the Claim is:

Standard is cheaper than Enterprise
and offers a longer refund period.

Evidence may require:

Passage A:
Standard price = €300

Passage B:
Enterprise price = €500

Passage C:
Standard refund period = 30 days

Passage D:
Enterprise refund period = 14 days

Multiple Evidence Pieces
↓
One Composite Claim

Claim–Source mapping is therefore not always one-to-one.

22. One Evidence Passage Can Support Multiple Claims

Passage:

Enterprise costs €500 per year
and includes priority support.

can support:

Claim 1:
Enterprise costs €500 per year.

Claim 2:
Enterprise includes priority support.

One Evidence Passage
↓
Multiple Claims

The Claim–Evidence relationship can therefore be many-to-many.

23. Claim Support Must Match the Claim's Specificity

Evidence:

The product is available in Amsterdam.

Claim:

The product is available in Amsterdam.

This is strong support.

But:

Claim:
The product is available throughout the Netherlands.

is broader than the Evidence.

Amsterdam
≠
the entire Netherlands

Support must align with the Claim at the correct level of specificity.

24. Claim Support Can Fail Through Overstatement

Evidence:

Some customers reported faster response times.

Claim:

The service always provides faster response times.

The Claim changes:

some customers

into:

always

This is an overstatement.

Another example:

Evidence:
The treatment may reduce symptoms.

Claim:
The treatment eliminates symptoms.

A supported Claim should not go beyond what the Evidence establishes.

19.3 Natural Language Inference

25. What Is Natural Language Inference?

Natural Language Inference, or NLI, is a framework for reasoning about the semantic relationship between two pieces of text.

In this chapter, we can think of them as:

Evidence
+
Claim

The central question is:

Given this Evidence, what can we conclude about this Claim?

A simplified NLI classification is:

Evidence + Claim
↓
Entailment
or
Contradiction
or
Neutral

26. What Is Entailment?

Entailment means the Evidence supports the Claim strongly enough that the Claim follows from the Evidence.

Example:

Evidence:
Enterprise customers may request refunds
within 14 days of purchase.

Claim:
Enterprise customers have a 14-day period
in which they may request a refund.

The wording differs, but the meaning is supported.

Evidence
↓
entails
↓
Claim

This is the relationship we usually want for grounded factual Claims.

27. Entailment Does Not Require Exact Wording

Evidence and Claim do not need to use identical words.

Evidence:
The subscription can be cancelled before renewal.

Claim:
Customers can stop the subscription from renewing.

The wording is different, but the meanings may be equivalent in context.

Entailment
≠
Exact Word Match

NLI concerns semantic support, not simple lexical overlap.

28. Entailment Requires the Whole Claim to Be Supported

Evidence:

Enterprise refunds are available within 14 days.

Claim:

Enterprise customers receive a full refund
within 14 days.

The Evidence supports:

refund
14 days

but does not necessarily support:

full refund

Therefore:

Partial Support
≠
Full Entailment

29. What Is Contradiction?

Contradiction means the Evidence supports information incompatible with the Claim.

Example:

Evidence:
Enterprise refunds are available within 14 days.

Claim:
Enterprise refunds are available within 30 days.

Assuming both refer to the same policy and conditions:

14 days
≠
30 days

The Evidence contradicts the Claim.

Evidence
↓
incompatible with
↓
Claim

30. Contradiction Requires Matching Conditions

Consider:

Evidence:
Standard refund period = 30 days.

Claim:
Enterprise refund period = 14 days.

These are not necessarily contradictory.

They refer to different plans.

Before labeling something Contradiction, we should check whether both statements concern:

same entity
same product
same location
same time period
same policy
same relevant conditions

31. What Is Neutral?

Neutral means the Evidence neither supports nor contradicts the Claim.

Example:

Evidence:
Enterprise subscriptions renew annually.

Claim:
Enterprise refunds are available within 14 days.

The Evidence does not support the refund Claim.

But it also does not establish that the Claim is false.

Evidence
↓
does not establish
↓
Claim

and

Evidence
↓
does not contradict
↓
Claim

32. Neutral Is Extremely Important

A common mistake is:

Not Contradicted
=
Supported

This is incorrect.

Suppose:

Evidence:
Company X is headquartered in Amsterdam.

Claim:
Company X has 5,000 employees.

The Evidence does not contradict the employee Claim.

But it does not support it either.

The relationship is Neutral.

No Contradiction
≠
Entailment

33. What Is an Unsupported Claim?

An Unsupported Claim is a Claim for which the available Evidence does not provide adequate support.

For example:

Evidence:
Enterprise subscriptions renew annually.

Claim:
Enterprise subscriptions include a 30-day refund.

The Evidence does not establish the refund statement.

The relationship is Neutral with respect to that Claim.

Claim
↓
No Adequate Supporting Evidence
↓
Unsupported Claim

34. Unsupported Does Not Automatically Mean False

Suppose:

Evidence:
Company X was founded in 2018.

Claim:
Company X has an office in Berlin.

The Claim might actually be true.

But this Evidence does not establish it.

Unsupported
≠
Necessarily False

Unsupported means:

The available Evidence
does not support this Claim.

Whether the Claim is factually true is a separate question.

35. Contradicted vs. Unsupported

Unsupported

Evidence does not establish the Claim.

Example:

Evidence:
Company X is based in Amsterdam.

Claim:
Company X employs 5,000 people.

Contradicted

Evidence supports an incompatible statement.

Example:

Evidence:
Refund period = 14 days.

Claim:
Refund period = 30 days.

Unsupported
→ no adequate support

Contradicted
→ Evidence points against the Claim

36. Entailment, Contradiction, and Neutral Together

Consider:

Evidence:
Enterprise customers may request a refund
within 14 days of purchase.

Claim A

Enterprise customers may request a refund
within 14 days.

Relationship:

Entailment

Claim B

Enterprise customers may request a refund
within 30 days.

Relationship:

Contradiction

Claim C

Enterprise customers receive priority phone support.

Relationship:

Neutral

37. The Core Model: Evidence → Claim

The central model of Chapter 19 is:

Evidence
↓
Claim

But the arrow really means:

Evidence
↓
supports / does not support / contradicts
↓
Claim

A more useful representation is:

Evidence
+
Claim
↓
Relationship Analysis
↓
Entailment
Contradiction
Neutral

A strong Claim should ideally have:

Evidence
↓
Entailment
↓
Claim

38. Why Strong Claims Should Be Entailed by Evidence

Suppose the AI says:

The refund period is 14 days.

If the Evidence says:

Refund requests must be submitted within 14 days.

the Claim is well supported.

But if the Evidence only says:

The company offers refunds.

we cannot infer:

14 days

So Relevant Evidence alone is not enough.

Evidence Meaning
↓
must support
↓
Claim Meaning

That is the central idea behind Entailment.

39. Claim Support Is Not Based on Similarity Alone

Consider:

Evidence:
Refund requests are accepted within 14 days.

Claim:
Refund requests are accepted within 30 days.

The texts share almost every word.

But the important number differs.

High Text Similarity
≠
Entailment

This gives us another important distinction:

Similarity / Relevance
≠
Claim Support

Semantic meaning matters.

40. Claim Support Can Depend on Qualifications

Evidence:

Customers may request refunds within 14 days
if less than 80% of service credits have been used.

Claim:

Customers may request refunds within 14 days.

The Claim omitted the condition:

if less than 80% of service credits have been used

The Evidence describes a conditional rule.

The Claim sounds unconditional.

Conditional Evidence
↓
Unconditional Claim

may therefore fail full Entailment.

Conditions, exceptions, and qualifiers matter.

41. Claim Support Can Depend on Time

Evidence A:

Updated January 2025:
Refund period = 30 days.

Evidence B:

Updated September 2026:
Refund period = 14 days.

Claim:

The current refund period is 14 days.

The newer Evidence may support the Claim.

The older Evidence may no longer describe the current policy.

Claim Support
may depend on
↓
time
version
freshness

42. Claim Support Can Depend on Entity Identity

Evidence:

Standard Plan refund period = 30 days.

Claim:

Enterprise Plan refund period = 30 days.

The number matches.

The entity does not.

Correct Attribute
+
Wrong Entity
=
No Valid Support

Entity precision therefore matters in Claim–Evidence relationships.

43. A Complete Claim Analysis Example

Suppose the user asks:

Can Enterprise customers cancel their annual plan
and get a refund?

The system retrieves:

Evidence A:
Enterprise customers can disable future renewal
from Billing Settings.

Evidence B:
Enterprise annual subscriptions may be refunded
within 14 days of the original purchase.

Evidence C:
After the 14-day period, cancellation prevents
future renewal but does not automatically create
refund eligibility.

The AI generates:

Enterprise customers can cancel renewal at any time.
They can receive a full refund within 14 days,
and cancellation after 14 days still guarantees
a partial refund.

Step 1 — Claim Extraction

Claim 1:
Enterprise customers can cancel renewal.

Claim 2:
Enterprise customers can receive a full refund
within 14 days.

Claim 3:
Cancellation after 14 days guarantees
a partial refund.

Step 2 — Claim 1 Analysis

Evidence A:

Enterprise customers can disable future renewal
from Billing Settings.

Claim:

Enterprise customers can cancel renewal.

Relationship:

Entailment

Step 3 — Claim 2 Analysis

Evidence B:

Enterprise annual subscriptions may be refunded
within 14 days of the original purchase.

Claim:

Enterprise customers can receive a full refund
within 14 days.

The Evidence supports:

refund
14 days

but does not explicitly establish:

full refund

The Claim is therefore not fully entailed by the available Evidence.

Step 4 — Claim 3 Analysis

Evidence C:

After the 14-day period, cancellation prevents
future renewal but does not automatically create
refund eligibility.

Claim:

Cancellation after 14 days guarantees
a partial refund.

The meanings are incompatible.

Relationship:

Contradiction

Step 5 — Final Claim Map

Claim 1
→ Entailed

Claim 2
→ Not fully supported

Claim 3
→ Contradicted

This is more precise than asking only whether the overall answer is correct.

44. Claim-Level Evaluation

Answer-level evaluation can hide differences between individual statements.

A more precise approach is:

AI Answer
↓
Claim Extraction
↓
Claim 1 → Evidence Relationship
Claim 2 → Evidence Relationship
Claim 3 → Evidence Relationship

This allows us to identify exactly where support succeeds or fails.

45. The Full Source-to-Claim Pipeline

We can now connect Chapters 18 and 19.

Corpus
↓
Retrieval
↓
Source Document / Passage
↓
Context Selection
↓
Evidence
↓
LLM Generation
↓
Claims

Then:

Evidence
+
Claim
↓
Relationship Analysis
↓
Entailment
Contradiction
Neutral

The complete support chain becomes:

Source
↓
Source Document
↓
Passage
↓
Evidence
↓
Claim

This becomes the foundation for Grounding, Faithfulness, and Citations.

46. Why This Matters for RAG

A RAG system can retrieve excellent information and still generate unsupported Claims.

Correct Evidence
↓
LLM
↓
Incorrect Claim

Retrieval quality alone therefore does not guarantee answer quality.

A more complete model is:

Good Retrieval
↓
Good Evidence
↓
Correct Evidence Use
↓
Supported Claims

The transition from Evidence to Claim is one of the most important points in a RAG system.

47. Why This Matters for AI Visibility

From an AI Visibility perspective, a brand or Source does not influence an answer merely by being retrieved.

The information must move through several stages:

Source
↓
Retrieved Passage
↓
Evidence
↓
Claim Support
↓
Final Answer

Suppose a brand page says:

Company X provides Persian-speaking legal services
in Toronto.

That Passage could support:

Company X provides Persian-speaking legal services
in Toronto.

But if the AI generates:

Company X is the largest Persian-speaking law firm
in Toronto.

the original Evidence does not necessarily support:

largest

Even when a brand becomes Evidence, its representation in the final answer depends on how Claims are generated from that Evidence.

48. Common Misunderstandings

Misunderstanding 1: A Claim is something only the Source says

Incorrect. In this part of the course, we primarily analyze Claims produced in the AI's answer.

Misunderstanding 2: The entire AI answer is one Claim

Not necessarily. One answer or sentence can contain multiple Claims.

Misunderstanding 3: Claim Extraction means generating new information

Incorrect. Claim Extraction identifies assertions already present in the text.

Misunderstanding 4: Claim Generation and Claim Extraction are the same thing

Incorrect. Claim Generation produces statements during answer generation. Claim Extraction identifies those statements for analysis.

Misunderstanding 5: If a Source is relevant, every Claim about that topic is supported

Incorrect. Support must be evaluated Claim by Claim.

Misunderstanding 6: If a Source is cited, the Claim must be supported

Incorrect. A cited Source can be related but fail to support the specific Claim.

Misunderstanding 7: Similar wording means Entailment

Incorrect. Numbers, entities, conditions, qualifiers, and certainty can change the meaning.

Misunderstanding 8: If Evidence does not contradict a Claim, the Claim is supported

Incorrect. The relationship may be Neutral.

Misunderstanding 9: Neutral means the Claim is false

Incorrect. Neutral means the Evidence neither supports nor contradicts the Claim.

Misunderstanding 10: Unsupported means factually false

Incorrect. A Claim may be true but unsupported by the available Evidence.

Misunderstanding 11: Relevant Evidence automatically entails a Claim

Incorrect. Relevant Evidence may still be incomplete or insufficient.

Misunderstanding 12: Partial support is the same as full Entailment

Incorrect. Important components of the Claim must all be supported.

Misunderstanding 13: One Claim always requires one Source

Incorrect. A Claim can depend on multiple Passages or Sources.

Misunderstanding 14: One Passage can support only one Claim

Incorrect. A Passage can support several Claims.

Misunderstanding 15: NLI is simply keyword matching

Incorrect. Entailment, Contradiction, and Neutral concern semantic relationships between Evidence and Claims.

49. How the Concepts Connect

Chapter 18 established:

Corpus
↓
Retrieval
↓
Passage
↓
Evidence

Chapter 19 adds:

Evidence
↓
Claim

But the arrow represents a relationship that must be examined:

Evidence
+
Claim
↓
Natural Language Inference
↓
Entailment
Contradiction
Neutral

The larger architecture is:

SOURCE LAYER

Source
↓
Document
↓
Passage


RETRIEVAL LAYER

↓
Retrieved Passage


CONTEXT LAYER

↓
Relevant Context


EVIDENCE LAYER

↓
Evidence


GENERATION LAYER

↓
LLM
↓
Claims


SUPPORT ANALYSIS

Evidence
↕
Claim
↓
Entailment / Contradiction / Neutral

We now understand:

Chapter 18
→ What information can function as Evidence?

Chapter 19
→ What does the AI Claim,
  and does the Evidence support it?

The next question is:

If Claims are supported by the available Evidence, what does it mean for the overall answer to be grounded and faithful to that Evidence?

That takes us into:

Chapter 20 — Grounding and Faithfulness

Mental Model

Imagine a journalist has gathered a Source Document.

One Passage says:

Company X's Professional plan increased
from €100 to €120 per month.

This Passage functions as:

Evidence

The journalist writes:

The Professional plan increased
from €100 to €120.

That statement is a:

Claim

Compare:

Evidence:
€100 → €120

Claim:
€100 → €120

The Evidence supports the Claim.

Entailment

Now suppose the journalist writes:

The Professional plan increased
from €100 to €150.

Compare:

Evidence:
€100 → €120

Claim:
€100 → €150

These are incompatible.

Contradiction

Now suppose the journalist writes:

Company X has 2,000 employees.

The pricing Evidence says nothing about employee count.

Neutral

If no other Evidence supports the employee Claim:

Claim
→ Unsupported by available Evidence

The simplest mental model is:

Evidence
+
Claim
↓
Ask:
Does the Evidence support this Claim?
↓
Yes → Entailment
Opposite / incompatible → Contradiction
No relevant information → Neutral

The complete reasoning model is:

AI Answer
↓
Extract Claims
↓
For Each Claim
↓
Find Relevant Evidence
↓
Compare Evidence and Claim
↓
Entailment
Contradiction
Neutral

And the simplest reliable definition is:

A Claim is a factual statement produced in an AI answer. Evidence supports a Claim only when the meaning of the Evidence is strong enough to justify what the Claim says. Natural Language Inference gives us a useful framework for describing this relationship: Evidence can entail a Claim, contradict it, or remain neutral. A strong grounded Claim should be entailed by its Evidence rather than merely being topically related to it.

Chapter 19 Recap

Claim

A statement that asserts something about the world.

The refund period is 14 days.

In this part of the course:

Claim
=
factual statement in the AI's answer

Answer vs. Claim

An Answer can contain multiple Claims.

Answer
↓
Claim 1
Claim 2
Claim 3

Claim Extraction

Identifying factual assertions contained in an Answer.

AI Answer
↓
Claim Extraction
↓
Individual Claims

Claim Generation

The process by which the LLM produces factual statements during generation.

Question
+
Context / Evidence
↓
LLM
↓
Generated Claims

Claim–Source Relationship

The connection between a Claim and the source information that supports it.

Source
↓
Document
↓
Passage
↓
Evidence
↓
Claim

Source Support

Whether a Source contains information capable of supporting the Claim.

Source
↓
Supporting Passage
↓
Claim

Evidence Support

Whether specific Evidence supports the meaning of the Claim.

Evidence
↔
Claim

Source Support vs. Evidence Support

Source Support
→ Does the Source contain support?

Evidence Support
→ Does this specific Evidence support the Claim?

Natural Language Inference

A framework for analyzing the semantic relationship between Evidence and a Claim.

Evidence
+
Claim
↓
Entailment
Contradiction
Neutral

Entailment

The Evidence supports the Claim.

Evidence:
Refund period = 14 days

Claim:
Refund period = 14 days

→ Entailment

Exact wording is not required. Semantic support is what matters.

Contradiction

The Evidence supports information incompatible with the Claim.

Evidence:
Refund period = 14 days

Claim:
Refund period = 30 days

→ Contradiction

Neutral

The Evidence neither supports nor contradicts the Claim.

Evidence:
Enterprise renews annually.

Claim:
Enterprise has priority phone support.

→ Neutral

Unsupported Claim

A Claim without adequate supporting Evidence.

Claim
↓
No Supporting Evidence
↓
Unsupported Claim

Unsupported does not automatically mean false.

Unsupported vs. Contradicted

Unsupported
→ Evidence does not establish the Claim

Contradicted
→ Evidence supports an incompatible statement

Partial Support

Evidence may support only part of a Claim.

Claim:
Full refund within 14 days

Evidence:
Refund within 14 days

Supported:
refund
14 days

Not established:
full refund

Partial Support
≠
Full Entailment

Evidence → Claim

The core model of this chapter is:

Evidence
↓
Claim

But the relationship should be analyzed as:

Evidence
+
Claim
↓
Entailment
Contradiction
Neutral

Strong Claim

A strong factual Claim should ideally be supported and entailed by Evidence.

Evidence
↓
Entailment
↓
Claim

The complete Chapter 19 pipeline is:

Corpus
↓
Retrieval
↓
Passage
↓
Evidence
↓
LLM Generation
↓
Answer
↓
Claim Extraction
↓
Claims
↓
Claim–Source / Claim–Evidence Analysis
↓
Natural Language Inference
├── Entailment
├── Contradiction
└── Neutral
↓
Supported or Unsupported Claims

The critical section of the course pipeline has now become:

Source
↓
Passage
↓
Evidence
↓
Claim

The next chapter is:

Chapter 20 — Grounding and Faithfulness

If an answer contains many Claims, how do we judge whether the answer as a whole remains supported by and faithful to the available Evidence?

The Single Most Important Idea from Chapter 19

A Claim is a factual statement made in the AI's answer. The existence of a relevant Source or Passage does not automatically support that Claim. The specific Evidence must semantically justify what the Claim says. Natural Language Inference gives us the core framework: the Evidence may entail the Claim, contradict it, or be neutral. A strong Claim should be entailed by its Evidence and should not introduce details, conditions, certainty, or specificity that the Evidence does not support.