Fact-Locked Claim Auditor

Keep AI-generated claims anchored to your approved facts.

Fact-Locked Claim Auditor is a deterministic quality-control Skill for checking declared factual claims against a buyer-provided Fact Sheet.

It helps catch common factual drift before AI-generated content moves further into your workflow — without relying on another AI model to guess whether the content is correct.

What It Does

Provide:

• An approved Fact Sheet containing the facts your content is allowed to rely on
• A claims file containing the factual claims you want to audit

The Skill checks the declared claims against the supplied facts and produces structured JSON and human-readable Markdown audit reports.

It can flag issues such as:

• Changed quantities or numeric values
• Singular scope being expanded into plural scope
• Required attribution being dropped
• Claims with no declared fact mapping
• References to unknown Fact IDs
• Missing declared factual claims

Example

Your approved Fact Sheet may state:

The official evaluation included 1 benchmark and reported an accuracy of 82%.

But generated content may claim:

The system achieved 92% accuracy across multiple benchmarks.

The auditor can flag the changed quantity and expanded scope instead of silently accepting the rewritten claim.

Designed for AI Content Workflows

This Skill is useful when you already have a set of approved facts and want an additional deterministic QC step before generated content is reviewed, published, or passed to another workflow.

Possible uses include:

• AI-assisted article production
• Research-based content workflows
• Script and report generation
• Internal content pipelines
• Fact-controlled generation workflows
• Pre-publication claim review

Deterministic and Local

The core audit does not require:

• AI or LLM APIs
• External web services
• Network access
• A database
• A GPU
• Browser automation

The same supported input is designed to produce stable, reproducible audit results.

Fail-Closed Validation

The Skill validates its inputs instead of silently guessing what malformed data was intended to mean.

It can reject problems such as:

• Unknown Fact IDs
• Duplicate Fact IDs
• Duplicate Claim IDs
• Invalid required fields
• Invalid confidence values
• Unsupported or malformed structures

This helps prevent malformed input from producing a misleading successful audit.

Reports

Each audit produces structured results with an overall status:

• PASS
• REJECT
• INVALID_INPUT

Detected violations are reported with explicit reason codes so the result can also be consumed by automated workflows.

JSON and Markdown reports are supported.

Important: What PASS Means

PASS means the declared claims passed the deterministic checks supported by this version of the Skill.

It does NOT mean:

• The Fact Sheet itself is objectively true
• The Skill verified the claims against the internet or external sources
• Every possible semantic change was detected
• Every factual claim in arbitrary prose was automatically discovered
• The content is guaranteed to be factually correct

Fact-Locked Claim Auditor checks fidelity to the facts you provide. It is not a general-purpose truth verification service.

Known Limitations

Version 1.0.0 intentionally uses bounded, deterministic checks rather than open-ended AI semantic judgment.

Complex cases involving number reuse, negation, unit changes, attribution movement, or other semantic rewrites may require human review.

The Skill should be used as a quality-control layer, not as a replacement for source verification or final editorial review.

Requirements

• Python 3.10+
• A Skill-capable Agent environment
• UTF-8 structured Fact Sheet and claims input

No external AI service or paid API is required.

Version

1.0.0

Built for deterministic, local, fact-locked claim auditing.