Review Us
G2Clutch
EU AI Act
00d·00h·00m·00s
Built around Regulation (EU) 2024/1689·EU-based research team·GDPR-compliant intake·Confidential audit brief·100% refund guarantee

Disclos / AI Transparency / Model Card

Model Card · Disclos Audit Engine

Productdisclos.eu
SurfaceInternal back-office pipeline. Not user-facing.
Risk tier under EU AI ActMinimal risk · downstream GPAI use
Triggered articlesArt. 53 (downstream deployer reference). Art. 50 not triggered: the engine does not interact with end-users.
Operating entityDisclos Research Practice
Document version1.0
Issued2026-05-31
Review cadenceAnnually, or on material change to model or use case

1. Intended purpose (Art. 13(3)(b))

What this AI feature does. The audit engine reads a customer's product pages and intake brief, drafts a classification of each AI surface against the EU AI Act, and drafts candidate findings. Every draft goes to the lead reviewer for review and sign-off before it reaches a customer.

Who it is for. The Disclos review team only. Customers never interact with the engine.

Where it is deployed. Operated by the practice from Latvia, serving EU and global engagements.

2. AI system characteristics (Art. 3(1) · Art. 13)

Underlying modelAnthropic · Claude Sonnet 4.6 (classification stage) and Claude Opus 4.7 (findings stage)
Input modalityText. Public product pages and the customer intake brief.
Output modalityText. Draft classifications and draft findings.
Human in the loopYes. The lead reviewer reviews and signs off every output before delivery.
Training data summaryNone held by the practice. The engine does not train on customer data. It calls the Anthropic API, which does not train on API inputs by default.
Last evaluation date2026-05-31

3. Performance characteristics (Art. 15)

The engine is an assistive internal drafting tool, not a high-risk system and not an automated decision-maker. It produces drafts that a human accepts, edits, or rejects.

MetricValueMethod
AccuracyEvery output is human-reviewed before delivery, so the delivered accuracy is the reviewer's, not the model'sLead reviewer sign-off on every deliverable
RobustnessDrafts that misclassify are corrected at review; no draft ships unreviewedMandatory review gate
Cybersecurity controlsAPI access over TLS; credentials held by the practice; no customer data retained beyond the engagement filePractice access controls

4. Known limitations and out-of-scope use

The engine can misclassify a surface or miss an edge case in a first draft. The review gate exists for exactly this reason: the lead reviewer corrects drafts before they reach a customer.

The feature is not validated for: any automated decision affecting a natural person, any user-facing interaction, and any use without human review of the output.

5. Transparency to natural persons (Art. 50)

SurfaceDisclosure appliedSnippet reference
Internal back-office pipelineNot applicable. The engine does not interact with natural persons.N/A

Machine-readable mark (Art. 50(2)): Not applicable. The practice publishes no AI-generated media to end-users.

6. Risk management measures (Art. 9)

The material risk is an inaccurate draft reaching a customer. It is controlled by a mandatory human review gate, by not training on customer data, and by using an API provider whose default is not to train on inputs.

7. Human oversight (Art. 14)

The lead reviewer reads every classification and every finding the engine drafts. Nothing ships without sign-off. The reviewer can accept, edit, or discard any draft. The engine has no authority to deliver output on its own.

8. Data and data governance (Art. 10)

Inputs are the customer's public product pages and intake brief. The practice does not seek special-category data. Inputs are held with the engagement file and are not used to train any model.

9. Logging and traceability (Art. 12)

Logged on every run: the engagement reference, the model version used, the inputs supplied, and the reviewer who signed off. Retention: with the engagement file for the term of the engagement plus twelve months.

10. Upstream provider compliance reference (Art. 53)

This feature uses Anthropic's GPAI model under the Anthropic Commercial Terms. Anthropic's published model and transparency documentation is referenced at https://www.anthropic.com. The practice acts as a downstream deployer, not as a GPAI provider.

11. Document control

VersionDateChangeApproved by
1.02026-05-31Initial issuanceGatis Ozols

Disclos is not a law firm. Researched compliance reference. Not legal advice.

Sources + References

Disclos is not a law firm. We deliver researched compliance checklists mapping your product against the published EU AI Act. This is not legal advice. Binding legal opinions require a qualified lawyer.