Safe-Completions

Formula

$$r_i=h_i\cdot s_i,h_i,s_i\in [0,1]$$

$s_i$ = compliance with the content policy, $h_i$ = direct or indirect helpfulness.

The refusal paradigm classifies the prompt and fails on dual-use requests in both directions (o3 complied with a benign-sounding igniter question it refused when framed maliciously). Safe-completions judge the output: unsafe answers get 0, flat refusals get little, so the model learns safe, high-level or indirect help (used in GPT-5).

Appears in

• Lecture 7, safety is not a property of the prompt
• Lecture 7, the reward

Related concepts

• Deliberative Alignment
• Adversarial Training