Choosing the Right Policy Length

Choosing the Right Policy Length

Policy length is a key control over how deeply gpt-oss-safeguard can reason about your rules. Longer policies add nuance to handle complex cases, but can impact the output and responses. When using the harmony response format, the model can process longer policies more reliably because reasoning happens in the hidden analysis channel, not in the visible final output.

Use [https://platform.openai.com/tokenizer](https://platform.openai.com/tokenizer) to determine the length of your prompt. **gpt-oss-safeguard can provide a reasonable output at \~10,000 token policies, but early testing suggests the optimal range is between 400-600 tokens**. It’s important to experiment and see what works best for you as there is no one-size-fits-all approach. Think of the policy length like a “context budget.” Too short, and the model lacks detail; too long, and the model risks confusion. This is similar to writing policy for people to understand as well. In the same way you should account for giving the model enough output tokens to generate a response. Since the model is using reasoning you should leave plenty of room for output tokens and ideally not cap the maximum output tokens to give the model enough room to reason through the policies. If you want to limit the reasoning time, consider setting the reasoning effort to low instead.

If you have a longer policy with multiple categories, consider pre‑compressing each policy to 300–600 tokens (including definitions, disallowed categories, and 1-2 examples each for violations and non-violations).

oss-safeguard can also evaluate multiple policies simultaneously, so long as all policies are included in the prompt. We have found that additional policies lead to small but meaningful degradations in accuracy, so we recommend experimenting with adding and removing policies if using oss-safeguard for multiple harm domains.