Safety
capabilities second
Build AI that reasons from first principles, so science and society can trust what it concludes. That only works if our own decisions are as checkable as we want our systems to be.
Every release is reviewed against the thresholds in our published framework before it ships. A release that crosses one waits until the right safeguards are in place, and every decision is recorded.
“We publish every version of the framework, with dates, and never edit a published version in place.”
Reasoning Safety Framework, version 1.0
“If a model crosses a threshold, the release waits until the safeguards the framework names are in place.”
Reasoning Safety Framework, version 1.0
Verification
Every answer carries the steps behind it, so people can check it rather than trust it.
Accountability
A named person owns each release decision, and every decision is recorded.
Provenance
Outputs can be traced back to the model, version and tools that produced them.
01
Inform
- Published, versioned safety framework
- Model evaluation results for every release
02
Prevent
- Pre-release red-teaming against framework thresholds
- Usage policy with prohibited uses
03
Detect
- Monitoring for misuse patterns in our products and agents
- Reasoning-faithfulness checks on releases
04
Enforce
- Account suspension for policy violations
- A route to appeal every enforcement decision
Risk domains we test, the thresholds that trigger extra safeguards, and what happens when one is crossed.
Version history
- Version 1.0October 9, 2026First published version: risk domains, capability thresholds, review process.
Who is accountable
- Accountable for every release decisionMohammad Imtiaz, Founder