🚨 Bug | Model returns malformed output unless the system message is first in the messages array

[kmshdev]

When we build the prompt like this 👇🏼

messages = [
    {"role": "user",   "content": multi_questions_prompt},
    {"role": "system", "content": SYSTEM_PROMPT},
]

mech-predict/packages/napthaai/customs/prediction_request_reasoning/prediction_request_reasoning.py

Line 733 in a3b7332

{"role": "user", "content": multi_questions_prompt},

Expected behaviour
• Both model versions return a well-formed XML-tagged block so that
parser_prediction_response() succeeds.

Possible cause

OpenAI’s docs state that “system messages should come first”.
gpt-4o appears tolerant of reversed order; gpt-4.1 is not.

⸻

Work-around / Fix

Swap the message order:

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user",   "content": multi_questions_prompt},
]

…and bump the unit tests to assert the order so this doesn’t regress again.