Gpt-5.2 is not following the specified response format, and includes weird tokens. These tokens then derail our gpt-5.2 evaluator

API Bugs
1

where: at completions API (hosted in EU)
time: January 23rd
model: “gpt-5.2”,
temperature: 0,
stream: true,
service_tier: “priority”,
reasoning_effort: “none”,
verbosity: “low”
system_message:
“”"
…You will interact … with the MediVoice API.
\n\n## MediVoice API\n``` * function msg(message: string): void: Talk to the caller using the TTS system.\n * function set(facts: Record<string, any>): any: Set facts…
“”"

Unexpected behavior: The model responds with “set({“phone”:“015144389161”}) 乐盈”, which does not follow the specified MediVoice API (which would be msg("乐盈") instead), switches the language, and gives a semantically nonsensical answer.

Furthermore, if we use our rating prompt (also through the API using gpt-5.2), we get the response
“”"
mistake: “#outputCooruptions#”,
reasoning: “After set({\"existingPatient\":true}) the assistant output contains an … \u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b\u000b…
“””
with “infinitely” many “\u000b"

This bad behavior never occurs when we use gpt-4.1.

2

Besides “set({“phone”:“015144389161”}) 乐盈”, we see more and more bad tokens appended after our valid MediVoice API calls, repetitively tokens from the following list (some but not all of these are Polluted Chinese (PoC) tokens:
乐盈
򐔀
\\n{\"
สามสิบเอ็ด
ჯობ
արկ
【อ่านข้อความเต็ม
<
արգել
declare function
{\"
\\nnull
开元棋牌
软件合法吗
\\nេញ
\\r\\n\\n{\"
泰皇
\\n```
Linke Unter
򐡀
။\\n

3

Sorry for the late reply. If the format is currently described only in the prompt, Structured Outputs or strict function calling may help: https://developers.openai.com/api/docs/guides/structured-outputs. That doesn’t confirm a fix for the odd tokens you reported.  Since this topic has been inactive for some time, we are going to close it. If the issue is still occurring, feel free to open a new topic with a current sanitized example and share the related request ID privately with Support.