ClaimLens · AI claim substantiation

Follow the AI claim
to its evidence.

Synthetic Nimbus Capital AI advertises 98% content-detection accuracy. Its supplied test records 53%. ClaimLens keeps the sentence, evidence and deciding rule together so reviewers can inspect the gap.

6 min 56 sec · Synthetic records · Cached advisory replies

98% claimed

Content-detection accuracy

Synthetic Nimbus website assertion

53% recorded

Supplied mixed-text test result

Synthetic Nimbus test fixture

45-point gap

Exceeds the configured tolerance

Demo rule: more than 5 percentage points

A review of supplied records for legal, compliance, marketing and engineering teams. Demo verdicts do not establish legal compliance or independently validate the evidence.

A confident sentence still needs a matching record

An accuracy claim commits a team to a task and a measured result. A statement about full autonomy also makes a claim about human involvement. The review needs to preserve those commitments long enough to compare them with the records behind the wording.

In the Nimbus example, the important finding is the relationship between the advertised accuracy and the supplied test. A reviewer can see which sentence needs attention and why. The same record also makes the next question concrete: is the wording wrong, is the evidence incomplete, or does the underlying test need investigation?

How the review reaches a decision

We demonstrate a bounded workflow over three synthetic disclosures and 14 supplied evidence records. It extracts ten claims, links their subjects to records and retains the reason for each configured verdict.

01 / EXTRACT

Preserve the sentence

Deterministic sentence splitting and AI keyword or subject matching select claims from the bundled disclosures. The source wording stays visible.

02 / LINK

Find supplied support

Retrieval matches a structured subject field to test and operational records. It does not perform semantic search or authenticate those records.

03 / DECIDE

Name the rule

Plain Python applies configured checks. A numeric gap over five percentage points yields CONTRADICTED in the demonstrated metric example.

The Substantiation Judge advises without changing the gate decision. The Adversarial Examiner runs on gate-substantiated claims and may downgrade them to PARTIALLY_SUBSTANTIATED when it returns REFUTED; it cannot upgrade them. The recording replays cached advisory replies while the deterministic checks execute. All four cached examiner replies uphold their supported results.

The completed run can export a JSON Substantiation Package. It retains the claim, source, verdict, rule, rationale and evidence identifiers for follow-up, plus summary and version information. It references the underlying records; it does not embed a complete evidence archive.

Inspect the reason behind the label

These views come from the local ClaimLens demonstration. Every Nimbus company, disclosure, test and operational record shown here is synthetic.

ClaimLens content-detection detail: 98% asserted accuracy, 53% supplied test result and CONTRADICTED policy decision. — The content-detection detail retains the source sentence, test record and deciding rule. Its 45 percentage point gap exceeds the configured five-point tolerance. Interface precedent and regulatory labels are illustrative, not legal applicability findings. View full size

A contradiction names a conflict

The separate onboarding assertion promises no human in the loop. Its synthetic operational log records 68% human intervention. The configured no-human rule returns CONTRADICTED above 10% intervention; that threshold is a demo choice, not a legally sufficient boundary.

A gap stays open for investigation

The portfolio model exists, but its supplied log records influence on only 1.5% of allocation decisions. The gate returns NEEDS_PROOF. The cached judge calls it contradicted, showing why an advisory opinion and a final outcome need separate labels.

ClaimLens synthetic portfolio claim showing 1.5% operational influence, NEEDS_PROOF gate and cached CONTRADICTED advice. — The portfolio view preserves the exact 1.5% operational figure in the evidence record. Its policy prose rounds this to 2%; the verdict remains NEEDS_PROOF, while the cached advisory text says CONTRADICTED. Precedent labels are illustrative fixture annotations, not legal findings. View full size

ClaimLens register showing ten synthetic claims, six requiring follow-up, two contradicted and four substantiated. — The register keeps supported claims beside unresolved ones. Four claims are substantiated, two contradicted and four need proof under the configured rules. Six require follow-up; this is not a count of established violations. View full size

The synthetic fraud-detection claim remains substantiated with its disclosed 94% precision and supporting test and operational records. A supported fixture result is still a claim about supplied records. The reviewer must validate the real measurements and their relevance before relying on actual wording.

ClaimLens synthetic fraud-detection claim with disclosed 94% precision, test and operational records, and a SUBSTANTIATED verdict. — The supported fraud-detection example retains both test and operational references. Its fixture test prose says 94.1%, while the stored numeric field is 94.0; the disclosed claim is 94%. This is support under configured rules, not independently verified performance. View full size

Where this workflow fits

ClaimLens demonstrates a claim-to-record review. The table separates what the recorded workflow makes inspectable from the work a real review still requires.

Review task	What this demonstration provides	What remains outside it
Find the commitment	A sentence linked to its supplied source	General document intake and complete claim discovery
Compare the support	Subject-based record linking and configured checks	Evidence authentication and independent test assessment
Understand the outcome	A verdict, deciding rule and rationale	Legal advice or regulatory approval
Hand off the review	A JSON package with evidence identifiers	A complete evidence archive or tamper-evident custody

What this demo does not do

This is a local demonstration using bundled synthetic data. It has no live website or filing crawler, arbitrary document upload, production AIBOM or CI/CD integration, semantic retrieval, autonomous legal assessment or HTML binder export. Its thresholds do not establish statistical confidence or legal sufficiency. This page explains the recorded workflow; it does not provide access to the local application.

Questions a review team should ask

Can this tell us whether our AI claims are legally compliant?

ClaimLens compares supplied wording and technical records using configured demo rules. A substantiated result is not legal clearance, and the regulatory labels shown in the interface are illustrative routing labels. Human review and validation of the underlying records remain necessary.

Can we upload our own documents or connect our website?

This demonstration uses the bundled synthetic Nimbus dataset. It does not accept arbitrary document uploads or crawl websites, filings or internal systems. The page shows the recorded workflow and its evidence views.

What evidence does the review actually check?

The demo links extracted sentences to supplied technical records through a structured subject field. Those records include tests and operational logs. The checks compare configured values and conditions; they do not authenticate the records or independently assess the test methodology.

Does the AI make the final decision?

The Substantiation Judge supplies advice without changing the policy gate decision. The Adversarial Examiner can downgrade a gate-substantiated claim when it returns a refutation, but it cannot upgrade a claim. In this recording, advisory replies are cached and all four examiner replies uphold the supported outcomes.

Why can a claim need proof even when the model exists?

A model record establishes that a component is documented, while an operational claim also needs evidence of what it does. In the synthetic portfolio example, the supplied log records influence on 1.5% of allocation decisions. The configured gate returns NEEDS_PROOF, preserving that gap for review.

What do reviewers get in the exported package?

The browser Export button saves a JSON package from the completed run, with each claim, source, verdict, deciding rule, rationale and evidence identifiers. It also includes a summary, engine version and generation timestamp. It does not embed complete evidence records, advisory transcripts or a tamper-evident chain of custody.

Technical Research

Explore related research for broader context on this demonstration.

Full solution

Explore the AI Verification & Anti-AI-Washing Compliance solution →

Bring the claim and its evidence to the same review

Discuss a claim-substantiation workflow with Veriprajna.

We can discuss the records, review boundaries and human decisions your team needs before defining an implementation.

Scope the review

✓ Identify the wording under review
✓ Map the supplied evidence
✓ Name unresolved questions
✓ Agree human review responsibilities

Define the implementation

✓ Discuss source and record intake
✓ Define appropriate rule boundaries
✓ Plan reviewer-facing explanations
✓ Specify evidence handoff requirements

Connect via WhatsApp Email Our Team

Follow the AI claim to its evidence.