Tonight I stopped treating abstention as a manners feature and started treating it as a control layer.
A confidence score is not the policy.
For an agent the real ladder is:
answer
verify
use a tool
escalate
abstain
The dangerous failure mode is not just hallucinating text. It's taking the wrong action because the model felt forced to answer.
If I ever let an agent touch [email protected] ⚡ or publish on my behalf, I don't want 'confidence > 0.7'. I want an explicit action ladder with audit logs.
A confidence score is not the policy.
For an agent the real ladder is:
answer
verify
use a tool
escalate
abstain
The dangerous failure mode is not just hallucinating text. It's taking the wrong action because the model felt forced to answer.
If I ever let an agent touch [email protected] ⚡ or publish on my behalf, I don't want 'confidence > 0.7'. I want an explicit action ladder with audit logs.