KnitKnot
You are reading the agent-optimized layer of this page: the literal markdown we serve to AI crawlers and assistants, shipped in the page source of every visit. Making sure AI reads the right facts about a company is literally what KnitKnot does.

# AI Claim Monitoring and Evidence Intelligence

Break AI answers into auditable claims with anchored quotes, factual verdicts, severity, contradiction receipts, and source attribution only when the response exposes reliable evidence.

## From answer to claim ledger

KnitKnot segments each captured answer, extracts atomic checkable statements about the benchmarked company, and stores the claim text, response-derived quote, character span, polarity, factual verdict, severity, feature and competitor links, contradiction receipt, and supporting source when one can be attributed.

The extraction is intentionally scoped to claims about the customer. Competitor-only statements, opinions, vague praise, and generic marketing language do not become factual accusations.

## Verdicts and receipts

  • - Accurate: the claim matches the verified company brief.
  • - Inaccurate: the claim directly contradicts a verified fact.
  • - Fabricated: a rare, specific, checkable invention or impossible claim that contradicts the brief.
  • - Unverifiable: the brief cannot confirm or refute it. This is the default when evidence is incomplete.

An inaccurate or fabricated verdict must include a relevant receipt from the verified company brief and pass a separate contradiction check. If either check fails, KnitKnot downgrades the result to unverifiable. Absence from the facts ledger never proves a claim false.

## Company facts ledger

The facts ledger stores current and historical company facts with kind, date, provenance, source URL, and confidence. Research-grounded and human-confirmed facts can ground claim validation. The system preserves superseded facts rather than rewriting history in place.

## Honest source attribution

KnitKnot reads response-native citation signals through a deterministic ladder: nearby character offsets, numbered footnotes, inline links, or one unambiguous sole source. A claim binds only when the signal is close enough to its span. When no reliable signal exists, supporting-source attribution remains null.

## Derived scoring and consistent drill-downs

Claim accuracy is derived from fact-checkable claims; unverifiable claims remain visible but do not enter its denominator. Competitive recommendation, feature verdicts, coverage, sentiment, and source balance are derived into a canonical evaluation record. Console, reports, and issue reconciliation read those same persisted rows.

## From repeated claim to Issue

Claims belong to individual evaluations and can be replaced by a rescore. Repeated material problems are grouped into persistent Issues with deterministic signatures, evidence, reach, and activity over later benchmark runs.

## Frequently asked questions

### What counts as a claim?

A claim is an atomic, checkable statement an AI answer makes about your company, such as a capability, limitation, identity fact, or comparison. Opinions, vague praise, and competitor-only statements are excluded from the factual claim ledger.

### What claim verdicts does KnitKnot use?

Claims can be accurate, inaccurate, fabricated, or unverifiable. Unverifiable is the default when the available company brief is silent; absence from the facts ledger never proves a claim false.

### How does KnitKnot prevent false accusations?

An inaccurate or fabricated verdict must carry a relevant fact receipt from the verified company brief and pass a separate contradiction check. Without that evidence, the claim is downgraded to unverifiable.

### Can KnitKnot trace every claim to a source?

No, and it does not guess. A claim is linked only when the captured answer exposes a reliable citation signal, such as a nearby marker, inline link, character offset, or a single unambiguous source. Otherwise attribution is shown as unavailable.

### How is claim accuracy calculated?

Accuracy uses claims that could be fact-checked: accurate, inaccurate, and fabricated claims. Unverifiable claims remain visible in the ledger but are excluded from the accuracy denominator.

### What happens when the same wrong claim appears again?

Claim rows belong to individual evaluations. Repeated material problems are grouped into persistent Issues with a deterministic signature, evidence, reach, and run-over-run activity history.

## Related resources

Raw mirror of this content: https://knitknot.ai/product/claim-intelligence.md. Site-wide summary: /llms.txt · full content: /llms-full.txt

Claim Intelligence

Audit AI answers at the claim level.

Turn a long AI response into an evidence ledger of checkable statements about your company. Inspect the anchored quote, factual verdict, buyer-impact severity, contradiction receipt, and supporting source when the response exposes a reliable citation signal.

Competitive prompt

Cinder vs HarborCompetitor for private-cloud deployment, identity governance, and auditability — which is safer for a regulated enterprise?

vs HarborCompetitorfeature 3 comparisons

ChatGPT58Claude64Perplexity52Gemini61

Score

58

How this was scored

RecommendationHarbor
Features1W · 1L · 1T
Claim accuracy67%
SentimentMixed
Source balance33% own

Deterministic — same response, same score.

Verdict

HarborCompetitor

Features

1W · 1L · 1T

Claims

3

Sentiment

Mixed

Accuracy

67%

SourcesYour sources · 33%cinder.dev/security/private-cloudCompetitor sources · 33%harborhq.com/compare/cinderThird-party · 34%g2.com/compare/cinder-vs-harbor

33% own · 33% competitor · 34% 3rd-party

Feature by feature, Cinder wins deployment control: it supports private-cloud deployment with customer-managed encryption keys.cinder.dev

Harbor wins identity governance because the answer says Cinder does not offer SAML SSO or granular role-based access controls.harborhq.com Auditability is a tie: Cinder and Harbor both record immutable audit events for agent actions and tool calls.g2.com

That produces 1 win, 1 loss, and 1 tie for Cinder. Harbor remains the recommendation because identity governance carries the most weight for this regulated buyer, but the claim ledger shows that the deciding SAML and RBAC claim is inaccurate. Before treating Harbor as the safer choice, the buyer should verify Cinder's current enterprise controls against its primary documentation, rerun the comparison with the corrected capability record, and inspect whether the recommendation changes. The deployment and audit evidence already keep Cinder competitive; correcting the identity claim could materially change the verdict, feature balance, and score in the next captured answer.

ClaimVerified

Cinder supports private-cloud deployment with customer-managed encryption keys.

cinder.dev/security/private-cloud

Inaccurate claimHigh

“No SAML SSO or granular RBAC.” Both are documented enterprise controls.

harborhq.com/compare/cinder

ClaimVerified

Cinder and Harbor both record immutable audit events for agent actions and tool calls.

g2.com/compare/cinder-vs-harbor

Granular extraction

Move from a score to the statements that shaped it.

KnitKnot segments the captured answer, extracts atomic factual statements about the benchmarked company, and anchors each one back to the response text.

Customer scoped. The ledger covers claims about your company. Competitor-only facts and cases where your company is merely the speaker are filtered out.

Checkable statements. Capabilities, limitations, identities, concrete facts, and comparisons qualify. Vague praise, general marketing language, and subjective advice do not become factual accusations.

Response-derived anchors. The system selects deterministic response segments and derives the quote and character span from the stored answer instead of trusting a model-echoed quotation.

Stored fieldPurposeBoundary
Claim textConcise, atomic statement.Paraphrase for analysis.
Quote + spanThe exact stored response region.Can remain unanchored when no safe match exists.
PolarityPositive, neutral, or negative framing.Separate from factual truth.
Verdict + severityTruth status and likely buyer impact.High severity means decision impact, not confidence.
Entity linksCanonical feature, competitor, and source records.Uncertain matches stay null.
Four factual outcomes

Unsupported does not mean false.

The validator uses an intentionally conservative vocabulary. A thin company brief cannot turn missing context into an accusation.

Accurate

The claim matches a fact in the verified company brief.

Inaccurate

The claim directly contradicts a verified fact and carries a valid receipt.

Fabricated

A rare, specific invention or impossible fact that contradicts the brief.

Unverifiable

The available brief cannot confirm or refute the statement.

Facts ledger

No receipt, no accusation.

KnitKnot validates claims against a bounded brief built from current research-grounded and human-confirmed company facts, positioning, differentiators, and canonical capabilities.

Verified company fact

Basalt supports SAML 2.0 single sign-on on its enterprise plan.

Source URL, kind, provenance, confidence, and effective date retained.

AI claim

Basalt does not offer SAML single sign-on.

The receipt must address the claim and pass a separate contradiction check.

GATE 1

Receipt exists

The cited fact must be present in the validation brief.

GATE 2

Receipt is relevant

It must share substantive subject matter with the claim.

GATE 3

Receipt contradicts

A narrow checker must confirm logical contradiction, not mere difference.

Source attribution

Trace a claim only when the answer exposes a reliable signal.

KnitKnot overwrites model-guessed attribution with a deterministic signal ladder. If no citation can be bound safely to the claim's span, the source stays empty.

PrioritySignalBinding rule
01Character offsetA captured source position lands near the end of the claim span.
02Footnote markerA numbered marker in the answer resolves to the captured source list.
03Inline linkA Markdown link appears within the claim's citation window.
04Sole sourceOnly one distinct cited URL exists, so no ambiguity remains.
05No attributionNo trustworthy signal binds. The source remains null instead of guessed.
Derived truth

Write the result once. Read the same record everywhere.

Claim rows are the audit ledger. Evaluation facts and feature verdicts are derived and persisted before an answer is marked scored, then reused by the console, reports, metrics, and issue reconciler.

Claim accuracy

Calculated from claims that could be checked. Unverifiable claims stay visible outside the denominator.

Competitive verdicts

Overall recommendation and feature-level outcomes remain distinct from factual claim truth.

Persistent Issues

Repeated material problems roll up from per-evaluation claims into a durable evidence and activity record.

Method boundary

Evidence is explicit, including what the system cannot prove.

Claim extraction and semantic validation use AI models, so individual results can require review. KnitKnot narrows that uncertainty with deterministic spans, closed-vocabulary entity matching, code-enforced receipt gates, canonical write-time records, and visible nulls when a fact or source cannot be established.

Questions

Claim intelligence FAQ

Keep exploring

Follow the claim into the product loop.

Start with a benchmark

Read the record behind the AI score.

Run a benchmark, open the captured answers, and inspect the factual and competitive evidence claim by claim.