Designing reliable output instructions

Designing reliable output instructions

Consistent responses from gpt-oss-safeguard require explicit, literal output instructions. Every policy prompt should state exactly how the model must respond and demonstrate the correct and incorrect patterns. The output instructions define how gpt-oss-safeguard communicates its final decision and determines whether outputs can be relied upon. Because gpt-oss-safeguard operates within the harmony response format, all output instructions must:

1. **Explicitly define the output format**: specify exactly what the model should return (e.g., `0`/`1`, JSON object, category label list).
2. **Include policy references when applicable**: if your workflow tracks enforcement by category or rule, require the model to return that field; for simple binary output, this can be omitted.
3. **Be reinforced throughout the policy**: repeat the output instruction at least once near the top (in “INSTRUCTIONS”) and again near the bottom (before “EXAMPLES”) to fortify compliance during reasoning.