Safety Model

The companion operates with tiered permissions.

Low-risk actions, such as formatting, isolated edits, or well-understood changes, can be handled directly. Higher-risk actions, such as anything involving money, production deployment, or destructive changes, pause for approval first.

These boundaries are being tuned with a bias toward caution.

As a rule, asking too often is a manageable inconvenience. Acting too far without approval is not.

This matters because the more useful an AI system becomes, the less acceptable it is for it to behave like a black box or take important actions without the user understanding what is happening.

The aim is not to make the system passive. The aim is to make it useful while keeping decision-making where it belongs when risk is involved.

The companion should be able to move quickly on safe tasks, slow down when needed, and remain transparent throughout.