Safety
How we think about scaling safely
Our Responsible Scaling Policy defines the conditions under which we will, and will not, continue to scale our models. It is a binding internal commitment, not a public relations document.
Current status: Capability Level 2 (CL-2) · As of January 2026
Background
What is a Responsible Scaling Policy?
A Responsible Scaling Policy is a formal commitment that specifies the safety conditions a model must meet before we scale it further. It operationalises the principle that capability and safety work must proceed together, not sequentially with safety retrofitted after capability is achieved.
Our RSP defines four Capability Levels (CL-1 through CL-4), each corresponding to a progressively more capable and potentially more dangerous class of AI system. At each level transition, mandatory evaluations must be completed and safety thresholds must be met before pre-training can continue.
If evaluations reveal that a model is approaching a higher capability level without adequate safety measures, we commit to pausing scaling until those measures are in place, regardless of commercial timeline pressures.
The Framework
Four safety thresholds
Each capability level defines what a model can do and what safety measures must be in place before we proceed.
Advanced Research Assistant
Models at CL-2 demonstrate expert-level performance across most academic disciplines and can carry out complex, multi-step research and reasoning tasks. They may provide meaningful assistance to domain experts but do not yet exhibit autonomous capability to cause large-scale harm without human direction.
Required safeguards: Constitutional alignment · Multilingual safety evaluations · Adversarial red-teaming · Output monitoring
General Purpose Assistant
Models at CL-1 are broadly capable across language, code, and reasoning tasks but do not exceed expert human performance in any single domain. Safety risks are manageable with standard deployment safeguards. Clark's Echo and Cipher class models operate at this level.
Required safeguards: Usage policy enforcement · Basic red-teaming · Refusal behaviour validation
Autonomous Capable System
Models at CL-3 can autonomously execute extended multi-step tasks, including tasks with real-world consequences, without continuous human oversight. At this level, misuse potential for large-scale harm increases significantly. We will not ship a CL-3 model without independent external evaluation and regulator notification.
Required safeguards: Independent third-party evaluation · Regulator notification · Enhanced containment · Deployment restriction review
Transformative Frontier System
Models at CL-4 represent a qualitative capability jump: the ability to conduct novel scientific research, develop new capabilities autonomously, or potentially undermine human oversight mechanisms. We will not scale to CL-4 under any commercial timeline. Reaching CL-4 triggers a mandatory pause and full independent safety review before any further development.
Required safeguards: Mandatory development pause · Government notification · Full public safety case · International peer review
Methodology
How we evaluate
Our capability evaluations run continuously throughout training, not only at major checkpoints. We maintain a suite of probes that detect early signs of capability uplift across: scientific knowledge, autonomous task completion, persuasion and manipulation, cybersecurity, and multilingual deception.
Evaluations are conducted by our Safety team independently of the training team. Results are documented in a capability log that is reviewed quarterly by our independent safety advisors. Any indication of unexpected capability emergence triggers an immediate pause and escalation.
External Accountability
Independent review commitment
We commit to independent external review at every capability level transition. Our review panel includes academic researchers in AI safety, legal scholars in technology governance, and civil society representatives, none of whom are compensated by AICONSORTIUM in ways that create conflicts of interest.
Review panel findings are published in full. Where we disagree with a finding, we publish our reasoning. We do not edit or withhold panel findings.