A Drop Below Baseline: Can a Romantic Partner's Predictive Model of Chat Style Detect Account Replacement by a Bot or Impostor?

This paper proposes, rather than claims to demonstrate, that romantic partners form a usable internal predictive model of each other's written replies in computer-mediated communication (CMC) [1–4, 9]. Existing behavioral and psycholinguistic studies indicate that individuals forecast upcoming utterances above chance [5, 9], that this predictive advantage is significantly larger for an intimate partner than for a stranger [3, 4], and that romantic couples' texting syntax, lexicon, and timing converge systematically over time [2, 6, 8, 10]. The hypothesis advanced here is narrower and operational: if one partner's messaging account is surreptitiously or experimentally replaced by an automated dialogue model (a bot) or by a human impostor writing through that account, the remaining partner's accuracy at anticipating the next reply should fall significantly below that individual's own established baseline. That within-person drop below baseline—rather than an isolated, uncalibrated impression that *"this feels off"*—is formalized as a testable behavioral signal of account replacement. A prospective, blinded, within-person experimental protocol is outlined. Following sober baseline calibration, each participant completes 20 randomized conversation blocks disguised under their partner's messaging interface: 10 blocks in which replies originate from their actual partner, and 10 blocks in which replies originate from a style-prompted bot or a trained human impostor. A pre-registered **8-of-10 pass bar** ($\ge 8/10$ replacement blocks dropping below personal baseline or explicitly classified as non-partner, paired with $\le 2/10$ false alarms on real-partner blocks) serves as the practical binary classification rule alongside continuous within-person accuracy contrasts. Furthermore, a between-subjects pharmacological factor compares a caffeine-abstinent control arm against an acute caffeine/pre-workout arm targeting approximately $200\text{ mg}$ of caffeine (corresponding to a standard half-scoop of a commercial pre-workout blend labeled at $400\text{ mg}$ caffeine per full scoop, matching the author's habitual intake). Because multi-ingredient pre-workout supplements (MIPS) contain co-active compounds that confound pure adenosinergic interpretations [11], full quantitative ingredient disclosure—or a parallel USP caffeine-only capsule arm—is mandated in the protocol. Methodological confounds, expectation biases, assumed similarity [7], and ethical safeguards for relationship trust and data privacy are critically examined. No causal or empirical conclusion is claimed prior to execution of the trial.