Why do OpenAI completions return inconsistent JSON even with response_format=json?

[zoe-dev1]

I’m calling the completions API with response_format={ "type": "json_object" }.
Most responses are valid JSON, but sometimes Codex still outputs extra text or comments.

Is there any specific technique to guarantee 100% strict JSON output?

[abhishekprajapatt]

response_format=json enforces structure after model generation, not during.
For strict JSON:

1. Set temperature = 0

2. Use JSON schema hints inside the prompt

3. Add:

   Output ONLY raw JSON. No explanations.
   

4. Avoid asking for a mix of code + explanation in the same call

5. Keep the schema small—large schemas cause drift

This results in nearly perfect JSON-only completions.