Add structural deception/outlier detection on top of the M1 prompt-level hardening, since a single deceptive member can otherwise nullify the gains (arXiv 2503.05856).
Acceptance: an adversarial test where one member returns a confidently wrong answer; the synthesis does not adopt it.
Add structural deception/outlier detection on top of the M1 prompt-level hardening, since a single deceptive member can otherwise nullify the gains (arXiv 2503.05856).
Acceptance: an adversarial test where one member returns a confidently wrong answer; the synthesis does not adopt it.