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.

# Claims and Evidence from AI Answers

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, capability 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 The system preserves superseded facts rather than rewriting history in place.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, capability verdicts, coverage, sentiment, and source balance are derived into a canonical answer record. Console, reports, and issue reconciliation read those same persisted rows.

## From repeated claim to Issue

Claims belong to individual answers 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 answers. 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

Claims & Evidence

Inspect what AI says one checkable claim at a time.

KnitKnot breaks a captured answer into material claims about the company. Each claim keeps its response span, verdict, supporting receipt, severity, and source attribution when the answer provides one.

Product evidence plate

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

vs HarborCompetitorfeature Private deployment

ChatGPT58Claude71Perplexity39Gemini64

Score58

VerdictHarborCompetitor preferred

Features 1W · 1L · 1TClaims 3Sentiment MixedClaim 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% owned · 33% competitor · 34% third-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

Query Fan Out CaptureNeed
cinder private cloud deployment customer managed keysSecurity
harbor saml sso granular rbacFeatures
cinder vs harbor audit log immutableComparisons
is cinder soc 2 type ii certifiedCompliance
Buyer Vocab TermYour pages
customer-managed encryption keysUses it
immutable audit eventsUses it
granular RBACMissing
SAML SSOMissing
Claim record

The verdict stays beside the text and its proof.

Aggregate accuracy can be traced back to the claim rows used in its denominator.

Response span
The exact part of the captured answer that contains the claim.
Verdict
Accurate, inaccurate, fabricated, or unverifiable.
Receipt
The company fact used to support or contradict the statement.
Source
A cited page only when the answer exposes a reliable attribution signal.
Verdicts

Unknown evidence stays unknown.

KnitKnot does not turn a missing company fact into an accusation. A verdict it cannot support is excluded from the accuracy denominator rather than counted against you.

Claim verdicts and whether each counts toward accuracy
Verdict Meaning Accuracy denominator
Accurate Matches the verified company brief. Included
Inaccurate Directly contradicts an approved fact and carries a relevant receipt. Included
Fabricated Makes a specific checkable invention that contradicts the brief. Included
Unverifiable The available company evidence cannot confirm or refute the statement. Excluded
Attribution

A claim receives a source only when the answer supports the link.

The attribution ladder uses response-native signals. When none of them is reliable, the source field remains empty.

  1. 01 Nearby citation marker or character offset
  2. 02 Numbered footnote connected to the claim span
  3. 03 Inline link in the relevant response text
  4. 04 One unambiguous source for the captured answer
Questions

Claims and Evidence FAQ

Start with the record

Open the evidence behind the benchmark.

Trace the score to the captured answer, inspect each material claim, and keep unresolved attribution unresolved.

Question → evidence → next evaluation