Large language models introduce a novel element to scholarly workflow: a critique partner that is always available, has no ego investment in the work, and can be explicitly prompted to adopt adversarial stances. This paper examines how LLMs can function as perturbation partners within an iterative publishing methodology, providing first-pass critique that raises the floor of quality before human feedback is sought. I address the sycophancy problem, the tendency of LLMs to validate rather than challenge, and present empirically tested prompting strategies that reliably elicit genuine critique: the Hostile Reader Frame, Reviewer Simulation, Steel-Manning Inversion, and Transmission Clarity Audit. I argue that multi-model triangulation, using different LLMs for drafting and critique, reduces the probability of correlated blind spots surviving to publication. Beyond critique, I explore LLMs as feedback triage assistants and sketch a vision of semi-autonomous perturbation infrastructure: systems that aggregate feedback across platforms, calculate iteration thresholds from feedback saturation, and manage version control for authors with large portfolios. The paper concludes by acknowledging the boundaries of LLM assistance. Models cannot reliably assess novelty, cannot provide field-specific positioning advice, and cannot substitute for domain expert evaluation. LLM critique raises the floor; it does not replace the ceiling that human perturbation provides. The goal is not automation but augmentation: freeing human attention for the feedback that only humans can give.
Storm Bjørn Flindt Temte (Wed,) studied this question.