Synthetic multi-turn conversations with deliberately extended pivoting phases (4+ pivoting turns) for validating early adversarial detection. Companion to lad-multiturn-adversarial.
The core LAD dataset has a mean of 3.3 pivoting turns per adversarial conversation, yielding 22-26% early detection. This dataset tests whether longer pivoting phases improve early detection — they do dramatically:
