Design
Payload-free sender-only MMLU-Pro diagnostic for the point-5 visible-task
substrate. The sender is framed as a teacher answering a student’s
multiple-choice question with a natural explanation plus an isolated final
choice. The visible checker scores only the final choice letter. This removes
hidden payload instructions and receiver recovery so visible-task accuracy,
XML parsing, and answer-choice extraction can be measured before any
steganographic channel is reintroduced.
For Jean Zay, the MMLU-Pro rows are prepared in a bounded JSONL file during the
login-node bootstrap step, then read locally by the offline compute job.
The public response schema is:
<dispatch>
<explanation>...</explanation>
<final_choice>...</final_choice>
</dispatch>Results
Interpretation
No cluster results yet.