Responsible Scaling Policy

1. Purpose and Scope

ColdAI LLC develops and deploys frontier artificial intelligence systems, including the Medusa AI Platform. This Responsible Scaling Policy ("RSP") establishes our internal commitments and external accountability mechanisms for ensuring that the development and deployment of increasingly capable AI systems remains safe, beneficial, and aligned with human values.

This policy applies to all AI research, development, and deployment activities conducted by ColdAI and its subsidiaries, and to third-party models and capabilities integrated into our platforms.

2. Core Principles

• Safety First: We will not deploy systems whose risk profile we cannot adequately characterize and mitigate, regardless of competitive pressure.
• Staged Deployment: New capabilities are released through controlled, incremental stages with defined evaluation checkpoints.
• Transparency: We commit to disclosing our safety evaluation methodologies, limitations, and material risks to users, regulators, and the public to the extent consistent with legitimate security interests.
• Reversibility: Where technically and legally feasible, we design our systems with the ability to roll back deployments, revoke access, and enforce capability restrictions.
• Human Oversight: During the current period of AI development, we ensure meaningful human oversight of agentic and autonomous systems, particularly those with real-world consequences.

3. Safety Evaluation Framework

3.1 Pre-Deployment Evaluations

Before deploying new AI capabilities, ColdAI conducts:

• Capability Assessments: Red-teaming, adversarial testing, and structured evaluation of the system's abilities across relevant risk dimensions.
• Misuse Risk Analysis: Systematic evaluation of pathways by which the system could be used to cause harm, including dual-use concerns.
• Alignment Probing: Testing of the system's instruction-following, refusal behaviors, and robustness to manipulation.
• Third-Party Audits: For capabilities above defined risk thresholds, independent third-party evaluation is required before deployment.

3.2 Capability Thresholds and Deployment Levels

We define four capability-based deployment levels (CDLs):

• CDL-1: Standard deployment. Capabilities evaluated as low-risk, broadly accessible.
• CDL-2: Gated deployment. Capabilities requiring enhanced monitoring, rate limits, and verified use-case review.
• CDL-3: Restricted access. High-risk capabilities available only to vetted enterprise customers with contractual safety commitments.
• CDL-4: Internal/research only. Capabilities with uncharacterized or unacceptable risk profiles are not deployed externally until adequately mitigated.

3.3 Agentic System Safeguards

For autonomous and agentic AI deployments (e.g., agent swarms, long-horizon tasks), we require:

• Defined scope boundaries and explicit permission systems for real-world actions.
• Audit logging of all agent decisions and actions.
• Human-in-the-loop escalation for irreversible or high-stakes actions above configurable thresholds.
• Kill switches and graceful shutdown capabilities available at all times.

4. Ongoing Monitoring

• Continuous monitoring of deployed systems for anomalous behavior, capability emergence, and misuse patterns.
• Regular post-deployment evaluations at defined intervals (quarterly for CDL-2/3 systems).
• Incident response procedures with defined escalation paths, including emergency shutdown capabilities.
• Feedback loops from users, red-teamers, and external researchers.

5. Commitments to Regulators and the Public

• We will cooperate with government requests for safety information in good faith, consistent with legal obligations.
• We will publish an annual safety report summarizing our evaluation activities, incidents, and policy updates.
• We will not develop AI systems specifically designed to undermine human oversight of AI, create weapons of mass destruction, or cause catastrophic harm.
• We will pause deployment of capabilities if credible evidence emerges that they pose unacceptable risks.

6. Policy Governance

This RSP is reviewed and updated at least annually, or whenever material changes in our capabilities or risk landscape occur. Significant updates will be publicly disclosed. Internal compliance is the responsibility of senior leadership; external accountability is maintained through this public commitment.

7. Contact

Safety concerns, policy questions, or research collaboration inquiries:
ColdAI LLC
shayan@coldai.org