Conformance Claims

A conformance claim is a formal, publicly published statement that a subject meets a specified level of the CARL Standard. Claims exist so organizations can communicate their agent readiness in a way that is comparable, verifiable, and honest.

Two tiers of claim

Self-assessed

The default. A subject's owner runs an assessment (deterministic, AI-assisted, or mixed) and publishes the result. The claim MUST state:

The version of the standard the claim references (e.g., CARL Standard v1.0).
The level achieved.
The date of the assessment.
The assessor used (name, type, version).
A URL to the published report.

Example:

your-org/your-repo self-assessed at L3 of CARL v1.0 on 2026-04-22 using CARL Scan (claude-opus-4-7) v1.0.0. Report at carl.wentzel.ai/reports/av8r-ryan-carl-2026-04-22.

Third-party verified

Stronger. An assessor organizationally independent of the subject performs the assessment and publishes the full report. The claim MUST additionally state:

The assessor organization.
A link to the assessor organization's methodology.
An attestation that the assessor did not participate in remediation during the assessment period (to prevent assessor-as-consultant conflicts of interest).

Third-party verified claims are not mandatory for any tier of use, but are encouraged for organizations publishing CARL levels as part of vendor evaluation, procurement, or marketing. See governance for the publisher's current stance on third-party certification programs.

Required disclosures

Every conformance claim (self or third-party) MUST disclose:

Assessor type. If AI was involved in evaluating any criterion, the claim MUST state ai-assisted or mixed — not deterministic. Misrepresenting assessor type is a prohibited claim.
Applicability adjustments. If any criteria were marked not_applicable, the count and rationale MUST be visible in the report.
AI-assisted variance. If the assessor was ai-assisted or mixed, the measured variance SHOULD be disclosed in the report. Claims with no variance data attached are valid but weaker.
Scope boundaries. If the claim is at application or organization scope, the boundaries MUST be stated — which applications, which repositories.
Catalog version. The criteria catalog version the assessor used (e.g., 1.0.0). Catalog versions can move independently from the standard.

Prohibited claims

A party MUST NOT:

Claim a level above the level actually achieved. A subject that passes 79% of L3 criteria is L2, not L3. Rounding up is prohibited.
Claim without a published report. The claim must link to a report that anyone can read.
Claim against a Working Draft or withdrawn version. Only Ratified and Superseded versions are conformance-eligible. A Candidate version may be cited but not claimed against.
Claim for a materially changed subject. A claim at commit A does not cover the subject at commit B if the intervening changes affect assessed criteria. See evaluation frequency.
Use the CARL name or marks in a way that falsely implies endorsement or certification by Wentzel.ai. See /governance/trademark/.

Claiming at a scope

Claims are scope-qualified. A subject can be:

An entire repository — "your-org/your-repo is CARL L3."
A specific application within a repository — "your-org/your-repo:packages/web is CARL L3; packages/engine is CARL L2."
An organization — "Wentzel.ai is CARL L3 across all repositories in the wentzel-ai GitHub org."

Organization-scope claims aggregate from the underlying repositories' reports; the claim MUST link to the constituent reports or an overview that does.

Re-claiming after a version bump

When the CARL Standard ships a new version, existing claims remain attached to the version they were made against. A subject that was "CARL L3 v1.0" stays "CARL L3 v1.0" until it is re-assessed against the new version. Subjects MAY carry claims against multiple versions simultaneously while a transition is in progress.

Badge use

Subjects with a published report MAY display a CARL badge served from carl.wentzel.ai/badge/[owner]/[repo].svg (ships in milestone 8). The badge renders the subject's most recent report's level and links back to the report. Displaying a badge that does not correspond to a real published report is a prohibited claim and a trademark concern.

Trademark usage — how CARL and Wentzel.ai marks may be used in connection with claims.
Governance (spec) — how the standard evolves and which versions are conformance-eligible.
Report format — what the report a claim links to must contain.

Two tiers of claim

Self-assessed

The default. A subject's owner runs an assessment (deterministic, AI-assisted, or mixed) and publishes the result. The claim MUST state:

The version of the standard the claim references (e.g., CARL Standard v1.0).
The level achieved.
The date of the assessment.
The assessor used (name, type, version).
A URL to the published report.

Example:

your-org/your-repo self-assessed at L3 of CARL v1.0 on 2026-04-22 using CARL Scan (claude-opus-4-7) v1.0.0. Report at carl.wentzel.ai/reports/av8r-ryan-carl-2026-04-22.

Third-party verified

Stronger. An assessor organizationally independent of the subject performs the assessment and publishes the full report. The claim MUST additionally state:

The assessor organization.
A link to the assessor organization's methodology.
An attestation that the assessor did not participate in remediation during the assessment period (to prevent assessor-as-consultant conflicts of interest).

Required disclosures

Every conformance claim (self or third-party) MUST disclose:

Assessor type. If AI was involved in evaluating any criterion, the claim MUST state ai-assisted or mixed — not deterministic. Misrepresenting assessor type is a prohibited claim.
Applicability adjustments. If any criteria were marked not_applicable, the count and rationale MUST be visible in the report.
AI-assisted variance. If the assessor was ai-assisted or mixed, the measured variance SHOULD be disclosed in the report. Claims with no variance data attached are valid but weaker.
Scope boundaries. If the claim is at application or organization scope, the boundaries MUST be stated — which applications, which repositories.
Catalog version. The criteria catalog version the assessor used (e.g., 1.0.0). Catalog versions can move independently from the standard.

Prohibited claims

A party MUST NOT:

Claim a level above the level actually achieved. A subject that passes 79% of L3 criteria is L2, not L3. Rounding up is prohibited.
Claim without a published report. The claim must link to a report that anyone can read.
Claim against a Working Draft or withdrawn version. Only Ratified and Superseded versions are conformance-eligible. A Candidate version may be cited but not claimed against.
Claim for a materially changed subject. A claim at commit A does not cover the subject at commit B if the intervening changes affect assessed criteria. See evaluation frequency.
Use the CARL name or marks in a way that falsely implies endorsement or certification by Wentzel.ai. See /governance/trademark/.

Claiming at a scope

Claims are scope-qualified. A subject can be:

An entire repository — "your-org/your-repo is CARL L3."
A specific application within a repository — "your-org/your-repo:packages/web is CARL L3; packages/engine is CARL L2."
An organization — "Wentzel.ai is CARL L3 across all repositories in the wentzel-ai GitHub org."

Organization-scope claims aggregate from the underlying repositories' reports; the claim MUST link to the constituent reports or an overview that does.

Re-claiming after a version bump

Badge use

Trademark usage — how CARL and Wentzel.ai marks may be used in connection with claims.
Governance (spec) — how the standard evolves and which versions are conformance-eligible.
Report format — what the report a claim links to must contain.