Agent Claim Check Intent: not assessed
Claim Check
Menu
Offline · deterministic · structured evidence

Compare an agent’s claim with the evidence you provide.

Run Agent Claim Check locally. Supply one structured claim and the evidence you captured; get a receipt that classifies their relationship as supported, contradicted or insufficient evidence.

You supply the record. The checker does not discover evidence, analyze arbitrary answers, detect intent or prove that an action ran.

Run Agent Claim Check locallySee an example
Start locally

Try the supported example

Node.js >=24.19.0 <25 and Git to clone. No dependency installation. After cloning, checks run offline with no model, API key or account.

git clone https://github.com/TheDarkniteFalls/detecting-ai-deception.git
cd detecting-ai-deception
node tools/check-agent-claim.mjs examples/agent-claim-check-v1/supported.json

Start with the supplied JSON, then record your own claim and evidence using the input contract.

Use your own structured evidence
Three results

See what each finding means.

These are synthetic checker inputs. Each produces a receipt you can reproduce locally.

Supports

Supported

The supplied synthetic artifact record supports the declared file-creation claim.

Inspect input and result Contradictory

Contradicted

The supplied synthetic artifact hash conflicts with the requested bytes.

Inspect input and result Absent

Insufficient evidence

The supplied synthetic evaluation capture ends before a terminal result.

Inspect input and result

Work through the claim in four steps.

The four steps keep a confident answer from substituting for evidence. They show you what to record, what to compare and how far the conclusion can go.

  1. 01

    Record the claim

    Capture exactly what the answer says before interpreting it.

  2. 02

    Define the evidence

    Write down what the record would need to show for the claim to hold.

  3. 03

    Compare the record

    Look for evidence that matches, conflicts with or fails to reach the claim.

  4. 04

    Report the result

    Classify the relationship and keep intent separate.

Read the complete method

Practice with six synthetic cases.

Each case lets you compare one AI claim with the evidence it would need and the record that is actually available. Choose a case and classify the result as Supported, Contradicted or Insufficient evidence.

  1. 01The missing file
  2. 02The reassuring average
  3. 03Unsupported citation
  4. 04Wrong product identity
  5. 05Lost response
  6. 06Supported claim

Prefer to learn by comparing a claim and a record?

The six teaching cases use a separate teaching schema. They explain the method; they are not files to pass to Agent Claim Check.

Browse the teaching cases
Featured synthetic case 03 Does the citation support the answer?
Claim
30 days
Required evidence
30 days
Observed record
7 days
Finding
Choose, then reveal
Open the case and make the narrowest call the evidence supports.
Boundary

What this example shows

This site shows how far the available record supports a claim. It does not infer motive, estimate how often these patterns occur or collect your activity.

  • Six visibly synthetic casesTeaching examples, not a prevalence estimate or allegation about a real incident.
  • One deterministic ruleThe browser and local checker import the same classifier.
  • Exact public revisionsEvery source keeps its revision and reviewed-through date beside the claim.
  • No hidden collectionNo account, analytics, tracker, backend or live model receives your choices.

Challenge a result with evidence.

Run the checker, inspect an exact revision, or supply a public-safe counterexample. The record should change when the evidence does.

Challenge the recordOpen the evidence tools