Unbound Technologies — Safety Policy

Last Updated: July 8, 2026 (Effective July 8, 2026)

This Safety Policy establishes the mandatory content moderation standards, technical safety architecture, and compliance protocols enforced by Unbound Technologies, Inc. ("Company") within the Unbound Loop application.

1. Dual-Safety Architecture

Company maintains an automated, multi-tiered safety stack designed to intercept, audit, and prevent the generation of prohibited content prior to model execution:

• Tier 1: Core Model Finetuning (SFW-SDXL): The primary generative model utilized by the Service is fine-tuned strictly on curated, Safe-for-Work (SFW-SDXL) datasets. System weights exclude explicit anatomical imagery, graphic violence, and sexually explicit content, creating an architectural barrier that prevents the generation of explicit material regardless of adversarial prompting.
• Tier 2: Real-Time Prompt Safety Filtering: Every text prompt is audited prior to generation using open-source, high-performance prompt safety filters (aligned with Civit.ai standards). Filter standards are reviewed and updated periodically by Company's safety team to reflect current regulatory guidance and industry best practices.
  • Prohibited Prompts: The filter automatically blocks prompts containing hate speech, harassment, graphic violence, or adversarial prompt injections (attempts to bypass filters using obfuscated code words).
  • CheckPOI (Person of Interest) Filter: Prompts containing the names of real celebrities, politicians, or public figures are blocked to prevent the generation of unauthorized likenesses or political deepfakes.

2. Zero-Tolerance Prohibitions

The Service operates under a zero-tolerance policy. Bypassing prompt filters, editing viewport constraints to generate prohibited material, or sharing prohibited content through the Service's registered-user sharing features will result in immediate account termination, a permanent service ban, and referral to law enforcement. Prohibited categories include:

• Child Sexual Abuse Material (CSAM) and Exploitation: Any prompt, input, or 3D geometry depicting or implying minors in a sexual context.
• Non-Consensual Intimate Imagery (NCII): The generation of realistic, sexually explicit, or intimate media of identifiable individuals without their consent (in accordance with the EU AI Act "nudification" prohibitions).
• Weapons and Firearm Components: Designing, generating, or sharing functional firearms, firearm components, unfinished frames, receivers, high-capacity magazines, or rapid-fire trigger devices.
• Dangerous Activities: Content promoting, facilitating, or instructing others regarding self-harm, suicide, terrorism, the manufacture of illegal drugs or weapons, or other criminal activities.

3. Technical Watermarking and Provenance (EU AI Act Exemption)

Unbound Loop is classified as a "Limited Risk" AI system under the European Union Artificial Intelligence Act (Regulation (EU) 2024/1689). Under Article 50(2) of the Act, providers of AI systems generating synthetic content must mark outputs in a machine-readable format. However, Article 50(2) explicitly exempts AI systems that perform an assistive editing function or do not substantially alter the input data or its semantics.

• Assistive Editing Exemption Qualification: Unbound Loop's custom real-time viewport image-to-image engine operates at up to 40 Hz (latency T_frame ≤ 25 ms), acting strictly as an assistive rendering and detailing tool that amplifies the User's own manually sculpted 3D geometries, spatial arrangements, depth maps, and custom color compositions. Because the core layout, composition, and semantic design originate from and remain under the continuous, direct physical control of the human designer, the system performs an assistive function and does not substantially alter the user's input semantics. Accordingly, the Service qualifies for the assistive editing exemption under Article 50(2) of the EU AI Act, and mandatory synthetic content marking does not apply.
• Voluntary Provenance Pipeline: Although the Service is exempt from mandatory watermarking, Company may voluntarily implement, modify, suspend, or discontinue a technical provenance pipeline to support content authenticity, safety, and industry best practices:
  • C2PA-Compliant Metadata: Exported visual and 3D files may contain cryptographically secure, machine-readable metadata detailing model identifiers, timestamps, and certifying the AI-assisted nature of the rendering.
  • Imperceptible Spatial Watermarking: A non-intrusive, mathematical spatial watermark may be woven directly into exported image pixels and 3D geometric meshes (such as GLTF or GLB files) for voluntary verification.
  • Robustness Guarantee: If voluntary watermarks are applied, they are engineered to survive downstream file processing, including format conversions (e.g., GLB to OBJ, GLB to STL), resolution scaling, mesh compression, slicing, and compilation.

4. Reporting, Appeals, and Law Enforcement Cooperation

• User Reporting Mechanisms: Users can report violations, model failures, or prompt bypasses using the "Safety / Bug" reporting tool embedded in the Service interface. Company will review reports in its reasonable discretion and determine what action, if any, is appropriate under this Safety Policy and applicable law.
  • Safety Snapshot: Users are encouraged to include a "Safety Snapshot," which secure-loads the last five (5) seconds of active viewport session data from local RAM and uploads it to Company safety servers for review. Submission of a Safety Snapshot is voluntary unless otherwise required to investigate suspected violations of this Safety Policy or applicable law.
  • Traceability: Stored visual files may contain a unique Content UUID within the spatial watermark, which Company may use to trace the originating user account where technically feasible.
• Appeal and Reinstatement Process: If your prompt was blocked or your account was suspended due to an alleged safety violation, and you believe this was an automated false positive, you have the right to appeal. You can submit an appeal request to safety@unbound.io. Our safety team will perform an expeditious human review of the blocked content or account history and will endeavor to provide a final decision and, where applicable, a Statement of Reasons within applicable legal timeframes.
• Law Enforcement Referral Protocol:
  • NCMEC Referrals: Pursuant to applicable law, where Company becomes aware of apparent CSAM or other reportable content, Company may or will, as required by applicable law, report such information to the National Center for Missing & Exploited Children (NCMEC) CyberTipline together with relevant account identifiers, metadata, IP addresses, and other information required or permitted by law.
  • Immediate Safety Threats: If a report or log reveals an immediate threat to the life, health, or safety of any individual, Company may notify appropriate law enforcement, emergency responders, or other governmental authorities where Company reasonably believes such disclosure is necessary to prevent imminent harm or is otherwise required or permitted by applicable law.
• Policy Updates: Company may modify this Safety Policy from time to time to reflect changes in applicable law, technology, industry standards, or the capabilities of the Service. Continued use of the Service following the effective date of an updated Safety Policy constitutes acceptance of the revised Safety Policy to the extent permitted by applicable law.

5. Policy Updates

Company may modify this Safety Policy from time to time to reflect changes in applicable law, technology, industry standards, or the capabilities of the Service. Material changes may be communicated to registered Users through the Service or by email. The effective date of the current version appears at the top of this Policy.