Observed hand-movement reversals postdictively bias ambiguous motion judgements Authors: Osamu Nomura, Kenichiro Ogawa Published: 2026 ABSTRACT Postdiction refers to the phenomenon whereby the visual system retrospectively integrates information arriving within a brief (approximately 100–200 ms) window after an event to finalize perceptual judgements. We investigated whether observed action cues, specifically mouse-control movements by others that were causally irrelevant to the display, can influence this late integration. In the stream/bounce experiments, participants viewed two identical bars approach, spatially overlap, and separate. Another person (an experimenter) facing the participant operated a mouse mounted on a motorized slider that was causally irrelevant to the bars. The slider executed one of five movement patterns: no movement (control), non-invert (continuous movement in one direction), pre (reversal 150 ms before the bars overlap), at (reversal as the bars overlap), and post (reversal 150 ms after the bars overlap). The participants knew that the other person’s mouse movement could not influence the bars’ motion. Relative to the control condition, the other person’s mouse-control reversal movements increased “bounce,” whereas continuous noninverting movement decreased bounce. Critically, in the post condition, where the other person’s reversal followed the bars overlapping, bounce increased substantially, indicating a postdictive influence. These results show that another person’s movement can serve as a context-conditioned prior that is postdictively integrated to disambiguate motion correspondence. Introduction Human perception is not mere feedforward processing of the present. Many studies have shown that perception can be retrospectively revised by information arriving tens to hundreds of milliseconds later, a process termed “postdiction”. Classic demonstrations include the flash-lag family and related temporal illusions, which challenge purely latency-based and strictly predictive accounts of visual awareness and instead suggest integration within a short retrospective window on the order of 100–200 ms1–8. Within that retrospective window, later evidence can reshape what observers report about earlier events, which is consistent with Bayesian or predictive formulations in which past sensory samples are reinterpreted considering new evidence and prior expectations9–13. In the flash-lag display, a briefly flashed target is misaligned with a simultaneously moving reference. Critically, the percept attributed to the time of the flash depends on the following motion signals, which is consistent with a short postdictive integration window rather than a purely online estimate2. Variants such as flash-grab and flash-drag reveal even larger mislocalizations when motion reverses immediately after the flash, indicating a strong influence of motion information accrued within roughly the next \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}100 ms14,15. Visual masking further shows that a trailing mask can both erase a target and, subjectively, appear to precede the erased target, illustrating that the naive temporal order can be violated by integration across a brief window16,17. Causal capture studies have also demonstrated that an ambiguous 100% overlap “collision” is judged as a causal launch when an unequivocal launching event occurs within a narrow temporal window (on the order of 150–200 ms), again indicating that later cues reshape the perception of the overlap event18–20. Taken together, these findings support the view that visual judgements about instantaneous events are finalized only after short-latency evidence is integrated. The stream/bounce display offers a sensitive probe of such postdictive integration: two identical objects (here, discs) approach, overlap in space, and then separate. Observers report either “streaming through” or “bouncing back”. Brief, precisely timed cues around the moment of overlap can strongly bias this categorical judgement. A classic study showed that a brief sound at overlap robustly increases “bounce” reports21–25. In subsequent work, a narrowly tuned temporal window for this bias in the stream/bounce display has been identified, peaking near synchrony and decreasing beyond approximately ±150–200 ms, which is consistent with a multisensory temporal binding window22,23,26–29. Converging multisensory evidence, including the sound-induced flash illusion30, near-optimal reliability-weighted fusion in cue combination9,10,12, and the neurobiological principles of multisensory integration31–39, supports a short decision window in which the visual system can integrate late-arriving disambiguating evidence about the overlap period. Most stream/bounce demonstrations have employed exogenous sensory cues, most commonly a brief auditory beep21–25. In contrast, observed movements by others have rarely been examined as postdictive priors. In daily life, when communicating with others, we often observe others’ movements, and it is possible that this observation of others’ movements influences our perception. If observing others’ movements influence postdiction, then observing another person’s mouse-control movements, especially a salient trajectory reversal suggestive of impact, should bias the late reconstruction of the overlap event towards a “bounce”. We therefore frame the present question at the influence of observing mouse-control movements by others on postdictive perceptual judgements (i.e., stream/bounce judgements), extending the scope of postdiction from simple exogenous cues to context-conditioned cues. This perspective is related to work on vicarious and illusory agency showing that visual information about another agent’s movement can bias judgements even when that movement is known to be causally unrelated to the outcome40–43. We regard these previous studies as evidence suggesting that observing others’ movements can influence perceptual inference, and we ask whether such signals also influence postdictive perceptual disambiguation in the stream/bounce display. Predictive processing and free-energy frameworks enable analyses of how context-conditioned priors can influence postdictive integration within short temporal windows. According to contemporary formulations, perception and action emerge from hierarchical generative models that minimize prediction error by dynamically weighting priors and likelihoods according to their precision44–48. Action observations involve top-down predictions about others’ movements and goals, which is consistent with predictive coding perspectives on action understanding and the mirror system49–53. Within this framework, the precision weighting among candidate interpretations (“streaming through” vs. “bouncing back”) can be adjusted on the basis of late evidence, and an observer’s action context (self-generated movement vs. passive viewing) could modulate the precision weighting of others’ action priors, while remaining agnostic about the neural implementation of postdiction2. In this study, we investigate whether observed mouse-control movements by others, beyond exogenous sensory cues, can act as postdictive priors in the stream/bounce display. We designed two experiments to test whether a causally irrelevant reversal mouse-control movement by another person can postdictively bias stream/bounce judgements and explore whether this bias varies with the observer’s movement context. In Experiment 1, participants moved the mouse to move the bars until shortly before the two bars overlapped; in Experiment 2, participants passively viewed otherwise identical displays while the bars moved automatically. In both experiments, we manipulated the reversal mouse-control trajectory of another person facing the participant, with this movement time-locked to the bars’ overlap and measured shifts in stream/bounce reports. Exploratory analyses were performed to determine whether the postdictive bias varied with the participants’ movement context (self-generated control vs. passive viewing). Our primary focus was on within-experiment postdictive effects; cross-experiment comparisons were exploratory. Methods Overview Each participant completed two experiments within a single session. In Experiment 1 (self-generated movement), participants initiated the movement of two bars on the monitor using a mouse. In Experiment 2 (no self-movement), participants held the mouse stationary while the bars moved automatically at a prespecified constant speed. In both experiments, another person (the experimenter) facing the participant operated a second mouse mounted on a motorized slider that did not influence the bars’ movements. The slider executed one of five movement patterns: no movement (non-move; control), non-invert (continuous movement in one direction), pre (reversal 150 ms before the bars overlap), at (reversal as the bars overlap), and post (reversal 150 ms after the bars overlap). Participants were explicitly instructed that the other person’s mouse was causally irrelevant to the bars’ movements. We adopted this setup because our primary aim was to test whether manual movements performed by another person—acting as a context-conditioned prior—bias stream/bounce judgements within a short postdictive time window, even when such movements are known to be causally irrelevant. In the analyses, the five-level factor of the other person’s movement (control (non-move), non-invert, pre, at, and post) within each experiment was treated as the primary manipulation. Cross-experiment comparisons between Experiment 1 (self-generated movement) and Experiment 2 (no self-movement) were performed as secondary, exploratory checks. Participants Thirty-five adults participated (8 participants in a pilot study that were excluded from the analyses; 27 participants in the main sample). All participants completed Experiment 1 and Experiment 2 within a single session. The main sample included 27 participants (12 women, 15 men; 18–22 years). All the participants were right-handed and had normal hearing and normal or corrected-to-normal vision. The participants were not informed of the purpose of the study. All the participants provided written informed consent and received compensation. No participants from the main sample were excluded. The study protocol was approved by the University of Fukuchiyama Research Ethics Committee and conducted in accordance with the Declaration of Helsinki (2013). Sample size determination The pilot experiment was performed solely to determine the target sample size for the main experiment, and the data of these participants were excluded from the inferential analyses. For each experiment, we computed within-subject paired differences in the proportion of bounce responses for each condition relative to the experiment-specific baseline (control). These paired differences were standardized as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_z$$\end{document} (within-subject standardized mean difference), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_z = \bar{\Delta } / s_{\Delta }$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\Delta }$$\end{document} = mean paired difference; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{\Delta }$$\end{document} = SD of paired differences). In the power analysis, the focus was the four prespecified within-experiment comparisons (control vs. non-invert, pre, at, and post). Although the familywise error in the analyses was controlled by Holm’s step-down procedure at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.05$$\end{document}, for planning, we used a conservative Bonferroni per-contrast level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha ' = 0.0125$$\end{document} (upper-bounding Holm), two-sided paired \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document} tests, and target power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 - \beta = 0.90$$\end{document}. For the comparisons, the maximum required sample size was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 25$$\end{document} (driven by Experiment 1 non-invert vs. control; pilot \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_z \approx -0.81$$\end{document}, reflecting a decrease in bounce under continuous movement conditions). To cover the most demanding comparison and provide a margin for pilot-effect uncertainty, we targeted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 27$$\end{document} for the main sample. For pilot-based effect sizes and planning assumptions, see Supplementary Information S9.Fig. 1Experimental setup. (a) Side view. On the participant’s side, a mouse platform was placed under the desk, occluding the participant’s hand and mouse from view. During the task, participants fixated on a grey cross mark at the top of the screen so that both the bars on the screen and mouse of the other person facing the participant remained within the visual field. (b) Top view. On the other person’s side (facing the participant), a second mouse was fixed to the moving stage of a motorized slider. The stage moved horizontally along the monitor’s long axis. The other person grasped the mouse on the slider stage and moved their hand in synchrony with the slider’s movement.Fig. 2Experimental conditions. (a) Experiment 1: The two bars approach, overlap, and separate. The participant operates the mouse only before overlap. The movements of the bars are initiated by the participant’s mouse, and then the bars move automatically prior to overlap. For clarity, the other person’s mouse is omitted. (b) Experiment 2: The participant keeps the mouse stationary while the bars move automatically at a prespecified constant speed. The other person’s mouse is omitted. (c) The slider executed one of five movement patterns (control, non-invert, pre, at, and post; see “Procedure” section for definitions). The other person (the experimenter), facing the participant, held the mouse fixed to the slider stage and moved their hand in synchrony with the slider’s movement. Experimental setup. (a) Side view. On the participant’s side, a mouse platform was placed under the desk, occluding the participant’s hand and mouse from view. During the task, participants fixated on a grey cross mark at the top of the screen so that both the bars on the screen and mouse of the other person facing the participant remained within the visual field. (b) Top view. On the other person’s side (facing the participant), a second mouse was fixed to the moving stage of a motorized slider. The stage moved horizontally along the monitor’s long axis. The other person grasped the mouse on the slider stage and moved their hand in synchrony with the slider’s movement. Experimental conditions. (a) Experiment 1: The two bars approach, overlap, and separate. The participant operates the mouse only before overlap. The movements of the bars are initiated by the participant’s mouse, and then the bars move automatically prior to overlap. For clarity, the other person’s mouse is omitted. (b) Experiment 2: The participant keeps the mouse stationary while the bars move automatically at a prespecified constant speed. The other person’s mouse is omitted. (c) The slider executed one of five movement patterns (control, non-invert, pre, at, and post; see “Procedure” section for definitions). The other person (the experimenter), facing the participant, held the mouse fixed to the slider stage and moved their hand in synchrony with the slider’s movement. Apparatus Participants used a USB mouse (M-BL27UBKBK, ELECOM, Japan) to control the bars in Experiment 1. The stimuli were displayed on a 27-inch LCD monitor (KA272A, Acer, Taiwan) with a resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1920\times 1080$$\end{document} pixels and a refresh rate of 60 Hz. A headrest was used to stabilize the participants’ head position. The viewing distance was set to approximately 49 cm, and the horizontal viewing angle was set to approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$40^\circ$$\end{document} (see Fig. 1). On the participant’s side, a mouse platform (height: 61 cm) was positioned under the desk, blocking the participant’s view of the hand and the mouse; therefore, participants could not see either the mouse or their own hand. Guide rails and left/right stoppers were installed to ensure consistent horizontal mouse movement (see Supplementary Information S10). The distance between the left and right stoppers limiting the mouse travel on the platform was 25.5 cm. On the other person’s side (facing the participant), a motorized slider (width: 68 cm; height: 7.0 cm; LEFB25T-500-S1C5171; SMC, Japan), which moved horizontally along the monitor’s long axis, was placed parallel to the monitor. The other person’s mouse was fixed to the moving stage of the slider. The other person used a second USB mouse (MA-BL9R; Sanwa Supply, Japan) that never influenced the bars’ movement (see Supplementary Information S10). The slider received control signals from a sequencer (FX3G; Mitsubishi, Japan) and a controller (JXC-W2A; SMC, Japan) that were synchronized with the bars’ movement via the stimulus computer. To block potential vibration cues, vibration-isolating material was placed under the motorized slider. In addition, vibration-damping rubber was attached to the mounting parts of the headrest, which was the only plausible route by which vibration could reach the participant. Under this setup, we verified that vibrations were not transmitted to the participant via the headrest through pilot testing. The stimuli, including the bars’ movements, were presented using PsychoPy (v2024.2.4). The sequencer and slider were operated via manufacturer-provided software. The slider was programmed to implement the five movement patterns used in the experiments (control, non-invert, pre, at, and post; see “Procedure” section and Fig. 2 for definitions). Synchronization between the movements of the bars and the slider (the other person’s mouse movement) was verified using a high-speed video camera (120 fps; HC-VZX2M, Panasonic, Japan) recording a stopwatch (1/100-s precision; W-753, Casio, Japan) to account for control and communication delays. As a result, synchronization between the bars and the mouse occurred within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10$$\end{document} ms. Stimuli and tasks The two bars were white rectangles (200 mm height, 7 mm width) presented on a black background on the monitor screen. A grey fixation cross-shaped mark (17 mm \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 17 mm; line width 3 mm) was presented at the top centre of the monitor so that both the bars and the mouse of the other person facing the participant remained within the participants’ visual field. The bars started 65 mm from the left and right screen edges, moved towards the centre, overlapped, and then continued towards the opposite edges before stopping (Fig. 2). In Experiment 1, participants moved the mouse from right to left along guide rails on the mouse platform to initiate the bars’ movement (see Supplementary Information S10). A stopper on the platform ensured that the participant’s mouse stopped before the bars overlapped; thereafter, the bars continued to move automatically at the constant speed measured immediately before the automatic movement began. Automation was triggered when the bars were approximately 20 cm apart (approximately 0.5 s before overlap), just before the participant’s mouse reached the left stopper. This control prevented mouse operations near overlap from unduly influencing stream/bounce perception. Following the automation step, the bars’ speed was recorded for later analysis of its potential influence on perceptual judgements in Experiment 1.In Experiment 2, the bars moved automatically from the start at a prespecified constant speed (3.5 mm/frame) without participant control. The participants were clearly informed that the movements of the bars on the monitor screen were caused by their own mouse movements or automatically controlled. Thus, the participants knew that the movements of the other person’s mouse never affected the bars’ movements. Procedure The experiments were conducted in a quiet, darkened room. The participant and the other person sat facing each other across a desk (90 cm apart) with the LCD monitor placed centrally between them (Fig. 1). The participants were instructed to fixate on the cross-shaped mark at the top centre of the monitor throughout each trial. To mask the sound of the slider’s movements, the participants wore earplugs and listened to white noise via headphones. Before the main experiment, participants practised moving the mouse in synchrony with a moving dot to learn the standard speed (3.5 mm/frame) for Experiment 1. They then completed practice trials in the fixed order (Experiment 1 then Experiment 2), each containing the five slider movement patterns repeated three times. During the practice trials, the experimenter confirmed that the white noise fully masked the mechanical sound of the motorized slider’s movements for each participant. In the main experiment, the order of Experiment 1 and Experiment 2 was counterbalanced across the participants. A break of approximately 10 minutes was scheduled between Experiment 1 and Experiment 2. Each experiment included 60 trials (12 repetitions of each of the five slider movement patterns) presented in a randomized order, totalling 120 trials per participant (Fig. 2). After each trial, participants reported “stream” or “bounce” by pressing the corresponding key. The prompt and two options were displayed on the monitor screen. In Experiment 1, participants moved the mouse along the horizontal guide rail from the right stopper to the left stopper on each trial. In Experiment 2, participants held the mouse stationary while the bars moved automatically on each trial. In both experiments, the other person facing the participant grasped the mouse fixed on the slider stage and moved their hand in synchrony with the slider’s movement. The slider followed one of five movement patterns (Fig. 2): Control (Non-move): No movement.Non-invert: Continuous movement in the same direction from onset to end, without reversal.Pre: Reversal 150 ms before the bars overlap, returning to the starting position.At: Reversal at the exact moment of the bars’ overlap.Post: Reversal 150 ms after the bars overlap.The five patterns were presented in randomized order. Control (Non-move): No movement. Control (Non-move): No movement. Non-invert: Continuous movement in the same direction from onset to end, without reversal. Non-invert: Continuous movement in the same direction from onset to end, without reversal. Pre: Reversal 150 ms before the bars overlap, returning to the starting position. Pre: Reversal 150 ms before the bars overlap, returning to the starting position. At: Reversal at the exact moment of the bars’ overlap. At: Reversal at the exact moment of the bars’ overlap. Post: Reversal 150 ms after the bars overlap. Post: Reversal 150 ms after the bars overlap. Data analysis General Analyses were conducted on trial-level binary outcomes (bounce = 1, stream = 0) using logistic regression (logit link). Unless noted otherwise, inference targeted prespecified within-experiment contrasts in the main experiment (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_{\textrm{subj}} = 27$$\end{document}; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_{\textrm{trials}} = 3240$$\end{document}). Effects are reported as odds ratios (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}=\exp (\hat{\beta })$$\end{document}) with Wald 95% confidence intervals (unadjusted) on the logit (link) scale; throughout, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}$$\end{document} denotes the point estimate from the fitted model (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{\textrm{OR}}$$\end{document}). Two-sided P values are shown, and families of planned contrasts were multiplicity adjusted using Holm’s step-down procedure. For model specification details, see Supplementary Information S1. Pilot data (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n = 8$$\end{document}) were used solely for power planning and were excluded from the confirmatory analyses (see Methods, Sample size determination, and Supplementary Information S9). All the statistical analyses were implemented in R (version 4.5.1) using custom scripts. Primary GLMMs (per experiment) Separately for Experiment 1 and Experiment 2, we fit binomial (logit-link) generalized linear mixed-effects models (GLMMs) with the condition as a fixed effect and a participant-level random intercept to account for repeated measurements:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {bounce} \sim \texttt {condition} + (1 \mid \texttt {participant}).$$\end{document}The reference level was the control within each experiment. We evaluated four prespecified planned contrasts (non-invert, pre, at, and post; each vs. control) and controlled the familywise error rate at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha = 0.05$$\end{document} using Holm’s step-down procedure (Holm-adjusted P values are reported). Wald 95% confidence intervals (unadjusted) are reported on the link scale and exponentiated to odds ratios. Factorial robustness (GLMM and GEE) As robustness checks, we also fit (i) a factorial binomial (logit-link) GLMM with fixed effects of the experiment (Experiment 1 vs. Experiment 2) and the condition (control, non-invert, pre, at, and post) and their interaction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {bounce} \sim \texttt {experiment}\times \texttt {condition} + (1 \mid \texttt {participant}),$$\end{document}with Experiment 1 (experiment) and control (condition) as reference levels and (ii) a population-averaged binomial generalized estimating equation (GEE) model with a logit link, clustered by participants with an exchangeable working correlation. For the GLMM, we report Wald 95% CIs (unadjusted); for the GEE, we report robust (sandwich) 95% CIs. Because the main effects depend on how one averages over the condition in the presence of interactions, per-experiment GLMMs were treated as the primary analyses, with explicit averaging schemes defined for cross-experiment summaries (see Supplementary Information S1). Mouse speed (Experiment 1) To test whether trialwise mouse speed modulated bounce in Experiment 1, we decomposed speed into within- and between-participant components via person-mean centring as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textit{speed}_{\textrm{within},ij}=\textit{speed}_{ij}-\bar{\textit{speed}}_{j},\qquad \textit{speed}_{\textrm{between},j}=\bar{\textit{speed}}_{j}-\bar{\textit{speed}}\,,$$\end{document}In addition, we computed z-scores for each component for interpretability. We then fit a binomial (logit-link) GLMM:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {bounce} \sim \text {condition} \times \text {speed}\_{\rm within} + \text {speed}\_{\rm between} + (1 \mid \text {participant})\,.$$\end{document}We tested (i) the within-participant speed slope in each condition, and (ii) four pre-specified slope contrasts comparing each to control. We report odds ratios (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}$$\end{document}) with Wald 95% confidence intervals (unadjusted, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}$$\end{document} corresponds to a 1-SD increase in speed); multiplicity within this interaction family was controlled using Holm’s step-down procedure. Diagnostics and auxiliary analyses We screened for complete/quasiseparation by (experiment \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} condition) cells and summarized GLMM convergence and overdispersion (Pearson \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ^2$$\end{document}/df). As an aggregation-level cross-check, we computed per-participant bounce proportions by experiment and condition, applied an empirical logit transform with a Haldane–Anscombe correction, and fit a participant fixed-effects linear model (see Supplementary Information S5). We also derived minimal detectable effects (MDEs) on the odds ratio (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}$$\end{document}) scale for the four planned contrasts using a two-sided normal approximation with target power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 - \beta = 0.90$$\end{document}. Because Holm’s step-down procedure does not yield a single fixed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} per contrast, we report MDEs using a conservative Bonferroni per-contrast level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha ' = 0.0125$$\end{document} (four-test family; see Supplementary Information S7). Secondary cross-experiment (exploratory; difference and equivalence) In addition to the prespecified within-experiment comparisons, we explored the overall difference between Experiment 1 and Experiment 2 and assessed practical equivalence. Because interactions make main effects depend on the averaging scheme, we defined three contrasts for the experimental effect on the log-odds scale: control-specific (primary for interpretation), equal-weighted, and proportion-weighted marginal averages (see Supplementary Information S1). For equivalence, we used TOST with SESOI \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}\in [0.80, 1.25]$$\end{document}, computing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\textrm{TOST}}=\max \{p_{\textrm{lower}}, p_{\textrm{upper}}\}$$\end{document} for each contrast on the log-OR scale and then applying the Holm adjustment across the three contrasts within the TOST family. We also report conventional two-sided difference tests; for completeness, their P values are likewise Holm-adjusted across the three contrasts. These exploratory contrasts were not powered a priori and were interpreted cautiously. Results Data overview and descriptive pattern A total of 3,240 trials from the 27 main sample participants were analysed for confirmatory inference, which corresponded to 100% of the collected trials, and no trials were excluded (pilot data were used only for power analyses; see Supplementary Information S9). Figure 3 summarizes condition-wise bounce proportions, displaying individual variability. The bars indicate the mean, and the error bars denote the standard error (SE); individual participant dots are overlaid to display the distribution. In both experiments, the control condition is the reference condition (no movement by the other person). The result pattern is consistent with integration within a short temporal window (approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 150$$\end{document} ms) in the previous study21: relative to the control condition, reversals by the other person (pre/at/post) increased “bounce,” whereas continuous noninverting movement (non-invert) decreased it. Critically, in the post condition, where the reversal by the other person followed the bars overlapping, “bounce” still increased substantially, indicating a postdictive influence. Primary GLMMs within each experiment (Holm-corrected planned contrasts) Figure 4 presents the prespecified planned contrasts with Holm-adjusted P values (denoted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}$$\end{document}), shown as odds ratios (ORs) with 95% confidence intervals on a log scale; the dashed vertical line indicates OR = 1. Key numbers (OR [95% CI], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}$$\end{document}) are displayed in the right margin (see Supplementary Information S2). Experiment 1 Compared with the control condition, bounce was robustly increased in the post condition (OR = 6.10, 95% CI [4.12, 9.03], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}), demonstrating postdictive integration. Bounce was also increased in the at (OR = 8.68, [5.83, 12.92], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}) and pre conditions (OR = 6.71, [4.53, 9.94], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}), whereas bounce was reduced in the non-invert condition (OR = 0.26, [0.15, 0.45], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}). Experiment 2 Similarly, compared with the control condition, bounce was strongly increased in the post condition (OR = 9.40, [6.24, 14.18], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}). Bounce was also increased in the at (OR = 11.02, [7.28, 16.70], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}) and pre conditions (OR = 9.57, [6.34, 14.44], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}) but decreased in the non-invert condition (OR = 0.30, [0.18, 0.50], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}<0.001$$\end{document}). Factorial GLMM and GEE (robustness) The factorial GLMM reproduced the pattern relative to the control condition: the odds of “bounce” were increased in the post condition (OR = 5.98, 95% CI [4.07, 8.79]), as well as in the pre and at conditions (ORs 6.58–8.54), whereas the odds of “bounce” were decreased in the non-invert condition (OR = 0.30, [0.18, 0.50]). Interactions with Experiment 2 were small and not significant (ORs 0.95–1.42; smallest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 0.21$$\end{document}). The population-averaged GEE (logit link; exchangeable working correlation) likewise replicated the pattern and again showed increases in the post condition (OR = 4.23, [2.36, 7.58]), as well as increases in the pre and at conditions and a decrease in the non-invert condition. For additional details, see Supplementary Information S3 and S4. Secondary cross-experiment check (exploratory) In Fig. 5, the data for Experiment 1 (circles) and Experiment 2 (squares) for the same planned contrasts are directly compared, highlighting alignment in both the direction and magnitude. Descriptively, for the three reversal-timing conditions (Pre, At, Post), the point estimates of the odds ratios are consistently larger in Experiment 2 than in Experiment 1; Exp1 vs. Exp2: Pre 6.71 vs. 9.57; At 8.68 vs. 11.02; Post 6.10 vs. 9.40, whereas for Non-invert both estimates are below 1. Because these between-experiment contrasts were not pre-specified and the confidence intervals largely overlap, we treat this pattern as exploratory and refrain from inferential claims here. Using a control-specific comparison (Experiment 2 vs. Experiment 1 in the control condition), the factorial model estimated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} = 0.91$$\end{document} (95% CI [0.60, 1.39]), indicating no compelling evidence of a large global shift (two-sided difference test, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}} = 1.000$$\end{document}). TOST with SESOI [0.80, 1.25] did not establish equivalence for this control-specific contrast (unadjusted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST}} = 0.266$$\end{document}; Holm-adjusted across the three definitions, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST,Holm}} = 0.266$$\end{document}). For completeness, marginal comparisons (equal- and proportional-weight averaging over condition; see Supplementary Information S1 for definitions) estimated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} = 1.05$$\end{document} (95% CI [0.87, 1.27]), with unadjusted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST}} = 0.038$$\end{document}; however, this was not significant after Holm correction across the three definitions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST,Holm}} = 0.115$$\end{document} for both the equal- and proportional- weight averaging analyses; see also Supplementary Information S8). Mouse speed in Experiment 1 The speed components and their associations with bounce are summarized in Fig. 6. Decomposed speed analyses revealed a small but reliable within-participant association with bounce (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} = 1.61$$\end{document}, 95% CI [1.31, 1.98], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P < 0.001$$\end{document}), with a nonsignificant between-participant component (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} = 0.78$$\end{document}, [0.46, 1.34], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 0.37$$\end{document}). Crucially, after Holm adjustment, none of the interactions between the condition and the within-participant speed were statistically significant (minimum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}=0.128$$\end{document}), providing no evidence that the speed differentially modulated the condition effects. Consistent with this, the Post vs. Control contrast remained large and significant after adjusting for speed (see Supplementary Information S6). These results indicate that mouse speed does not account for the primary condition effects in Experiment 1. Auxiliary ANOVA and MDE As an aggregation-level cross-check (participant fixed-effects ANOVA on empirical-logit proportions), we observed a large condition effect and no experimental main effect or interaction effect: Condition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(4,234) = 70.04$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P < 0.001$$\end{document}; Experiments \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(1,234) = 0.055$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 0.82$$\end{document}; Interaction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(4,234) = 0.66$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 0.62$$\end{document} (see Supplementary Information S5). Observed standard errors implied per-contrast MDEs (90% power; Bonferroni per-contrast level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha ' = 0.0125$$\end{document}) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{MDE}_{\textrm{OR}} \in [2.13, 2.90]$$\end{document} for Experiment 1 and [2.21, 2.72] for Experiment 2 (see Supplementary Information S7). All the observed pre, at, and post effects exceeded these thresholds in the direction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} > 1$$\end{document}, whereas the non-invert effect exceeded the threshold in the opposite direction (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} < 1$$\end{document}). Participant-level separation patterns are summarized in Supplementary Information S8.Fig. 3Condition-wise bounce proportions (mean ± SE) with participant-level dots overlaid. (a) Experiment 1 (self-generated movement). (b) Experiment 2 (no self-movement). The definitions of conditions are provided in “2.6” section and Fig. 2.Fig. 4Planned contrasts within each experiment (Holm-corrected). (a) Experiment 1 (self-generated movement). (b) Experiment 2 (no self-movement). Panels show odds ratios (OR) with 95% CIs on a log scale; the dashed vertical line indicates OR = 1. Right-side labels display “OR [CI], P”.Fig. 5Planned contrasts relative to Control, fit separately within each experiment; overlaid to visualise cross-experiment concordance. Cross-experiment contrasts (e.g., Exp2 vs Exp1 within Control) are not plotted; see text and Supplementary Information S8. Circles indicate Experiment 1; squares indicate Experiment 2. Right-side labels display “OR [95% CI], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}$$\end{document}”.Fig. 6Mouse speed effects in Experiment 1 decomposed into within- and between-participant components. Odds ratios (OR) with 95% CIs are shown on a log scale; right-side labels display “OR [95% CI], P”. Condition-wise bounce proportions (mean ± SE) with participant-level dots overlaid. (a) Experiment 1 (self-generated movement). (b) Experiment 2 (no self-movement). The definitions of conditions are provided in “2.6” section and Fig. 2. Planned contrasts within each experiment (Holm-corrected). (a) Experiment 1 (self-generated movement). (b) Experiment 2 (no self-movement). Panels show odds ratios (OR) with 95% CIs on a log scale; the dashed vertical line indicates OR = 1. Right-side labels display “OR [CI], P”. Planned contrasts relative to Control, fit separately within each experiment; overlaid to visualise cross-experiment concordance. Cross-experiment contrasts (e.g., Exp2 vs Exp1 within Control) are not plotted; see text and Supplementary Information S8. Circles indicate Experiment 1; squares indicate Experiment 2. Right-side labels display “OR [95% CI], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}}$$\end{document}”. Mouse speed effects in Experiment 1 decomposed into within- and between-participant components. Odds ratios (OR) with 95% CIs are shown on a log scale; right-side labels display “OR [95% CI], P”. Discussion We asked whether another person’s mouse-control movement that is explicitly known to be causally irrelevant to the stimulus influences postdictive visual inference in the stream/bounce display and whether concurrent self-generated actions modulate that influence relative to passive viewing. In keeping with the prespecified inferential scope, primary tests targeted within-experiment planned contrasts indexing the timing-dependent influence of the other’s movements; the cross-experiment check (Experiment 1 vs. Experiment 2) was treated as exploratory with appropriate error control. Figure 5 shows a direct comparison of Experiment 1 (circles) and Experiment 2 (squares) for the same planned contrasts, highlighting alignment in both the direction and magnitude. This exploratory check complements the per-experiment estimates and supports cross-experiment consistency. Across the two experiments and convergent analyses (trial-level GLMM with Holm-adjusted planned contrasts, factorial GLMM/GEE robustness, and a participant fixed-effects ANOVA on logit proportions), our results support a postdictive influence of observed mouse-control movements by others within a short temporal window; full robustness and ancillary analyses are provided in the Supplementary Information S2–S5 and S8. Critically, “bounce” was markedly increased in the post condition, during which the reversal followed physical overlap (Experiment 1: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}=6.10$$\end{document} [4.12, 9.03]; Experiment 2: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR}=9.40$$\end{document} [6.24, 14.18]), consistent with a short postdictive window. “Bounce” was similarly increased in the at and pre conditions, whereas it was decreased in the non-invert condition. Given that the other person’s mouse-control movement never influenced the display, that participants were explicitly informed about this noncausality, and that auditory cues were masked, these effects cannot be reduced to trivial physical causation. The timing (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 150\,\textrm{ms}$$\end{document} around overlap) is consistent with the short postdictive window established in stream/bounce and related phenomena3,21,30. Thus, observed mouse-control movement acts as a context-conditioned prior that can be retroactively integrated to disambiguate “bounce” versus “stream” within a narrow temporal window. Classic work with bouncing sound showed a timing profile that peaks at synchrony and shows pre over post (at > pre > post), while post still increases “bounce” relative to no cue21. Our mouse-control hand movements shows the same profile: at produced the largest increase, pre was slightly smaller, and post was smallest but significant. We interpret this match as evidence for a shared short postdictive window, with cue type shaping the window’s effective profile. In particular, an observed mouse-control hand reversal functions as an external control signal. Observers apply an action-effect prior, which predicts maximal influence at synchrony and when the action precedes the effect; a smaller post effect likely reflects retrospective re-alignment within the same narrow window rather than a different mechanism. Compared with bouncing sound, which is physically coupled to the bounce of the bars21, observed mouse-control hand reversals are context-conditioned and causally decoupled cues. Namely, bouncing sounds are readily construed as lawlike, physically coupled consequences of impact (tightly causally coupled; relatively high precision to the observer, even if their actual source is uncertain), whereas the observed mouse-control reversals in our task were explicitly stated to have no influence on the display and, furthermore, require contextual interpretation with respect to the on-screen motion (causally decoupled; relatively lower precision). Our results show that even such noncausal context-conditioned mouse-control movements can bias postdictive interpretation on subsecond scales, indicating that postdiction is not limited to multisensory physical evidence but extends to others’ movements that carry interpretive weight. This perspective complements classic demonstrations of perceptual causality and launching20 while highlighting a distinct context-conditioned grounded route to late disambiguation. An alternative explanation is that the observed bias reflects generic attentional capture or arousal rather than postdictive integration. Two empirical observations are inconsistent with this account. First, continuous, non‑inverting movement (non‑invert) did not increase “bounce” and significantly decreased it relative to control, which is inconsistent with a simple attentional-capture or arousal account that should increase “bounce” across timing conditions. Second, the timing profile was similar to the beep‑onset results (at > pre > post) rather than a nonspecific increase in arousal21. In addition, participants were explicitly informed that the other person’s mouse could not influence the bars, yet reversals still biased their judgements, arguing against explanations based on mistaken beliefs about physical causation or strategic responding. Finally, even if temporal grouping contributes, post-overlap reversals, which occur after the ambiguous overlap and thus cannot reflect purely anticipatory grouping, indicate a postdictive influence. Although we cannot definitively exclude transient salience or temporal grouping, the available evidence supports postdictive integration of a distinctive movement cue within a short temporal window. One possible explanation is based on joint action accounts in which an observer can co-represent the action of the other person54,55. From this perspective, when the participant’s action and the other person’s mouse-control movement are incongruent with respect to movement direction, the resulting prediction violation could bias ambiguous motion judgements towards “bounce.” This account is compatible with the finding that direction reversals increased “bounce” relative to continuous movement, which provides a concrete hypothesis about how observed mouse-control movements can influence postdictive disambiguation. Related work has also reported Simon-type interference even when the co-actor is replaced by a moving object56. This suggests that whether similar effects are observed when the mouse is moved by a motor without the presence of another person should be tested in future work. More broadly, this discussion can be linked to the BRAC framework57. According to this framework, irrelevant features of an observed action (e.g., movement direction) are bound into an ‘event file’ that, when retrieved, could modulate the postdictive integration of visual motion signals if their features are incompatible. Self-generated action appears to act as a contextual modulator rather than an on/off gate for postdictive perception. As noted above, our primary inferences relied on within-experiment contrasts; the cross-experiment comparison of Experiment 1 vs. Experiment 2 was exploratory and not part of the a priori power plan. In the factorial GLMM, the overall cross-experimental contrast was modest (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{OR} = 0.91$$\end{document}), and the corresponding two-sided difference test was not significant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{Holm}} = 1.000$$\end{document}; unadjusted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 0.670$$\end{document}; see Supplementary Information S3, S8). TOST with SESOI [0.80, 1.25] did not establish equivalence after the Holm adjustment across the three predefined comparisons (control specific: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST,Holm}} = 0.266$$\end{document}; equal-/proportional-weighted: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{\textrm{TOST,Holm}} = 0.115$$\end{document}; see Supplementary Information S8). Thus, while we find no compelling evidence for a large global shift between self-generated and passive contexts, strict equivalence remains inconclusive at the realized power. A cautious interpretation is that self-generated action may modulate, rather than globally override, the precision-weighting of context-conditioned priors within the same short postdictive window. The influence of motor signals on visual perception has been well-documented in the context of self-motion and manual control. For example, manual control of stimuli has been reported to modulate stream/bounce judgements in ambiguous motion displays58. Similarly, research on the flash-lag effect has shown that manual control can either reduce59 or increase60 the spatial offset depending on the task settings and temporal constraints. These findings suggest that internal motor signals provide a significant context for the temporal and spatial integration of visual information. Consistent with this literature, our exploratory findings in Experiment 1 suggest that self-generated movements may bias stream/bounce judgements. Although these results remain preliminary, they provide a theoretical basis for our main finding that the visual system’s postdictive window is sensitive not only to one’s own motor signals but also to the observed actions of others. By demonstrating that observed, causally irrelevant movement reversals can bias judgements even when they occur after the critical event, the present study extends the principles of motor-visual integration into the domain of context-conditioned postdiction. Under predictive-processing and free-energy views, perception over a short window can be modeled as precision-weighted fusion of priors and likelihoods, where precision (inverse variance) modulates the gain on prediction-error units. In our display, a reversal movement by a person facing the participant provides a temporally precise context-conditioned prior that favors “bouncing,” whereas continuous movement favors “streaming.” The timing-specific shifts are therefore compatible with precision-weighted integration of context-conditioned priors with sensory likelihoods in resolving motion correspondence. We do not claim that the present data uniquely identify these mechanisms. Rather, they motivate targeted tests that the framework predictions should modulate the bias, e.g., time-locked TMS to premotor/mirror/STS candidates around overlap or MEG/EEG decoding of action templates aligned to the critical event44,49,51. In terms of linking agency attribution to perceptual reconstruction, previous works have shown that observing movements by other people can bias sense-of-agency judgements even when those movements are known to be causally inert, which is consistent with comparator and multifactorial two-step accounts of agency41,42. Notably, our previous study revealed that such biases can arise even when the other person’s movement lags the visual outcome, which is consistent with a short retrospective (postdictive) integration window40. The present data extend this logic from metacognitive attribution to perception: observed action cues can reshape not only causal beliefs about control but also the visual segmentation of motion itself (bounce vs. stream). A unified—although not necessarily identical—mechanism, namely, precision-weighted integration of context-conditioned priors within a short postdictive window, may underlie both phenomena. An open question is whether the present effect depends on agency/animacy per se or generalizes to non-agent objects that exhibit movements matched in terms of velocity, trajectory, and motion energy. If human agents are special, context-conditioned priors may invoke person-specific models and expectations that others’ actions are informative and referential; otherwise, there may be more generic movement priors. In future work, “agent” could be considered as a continuum, from the human hand to a robotic arm to a geometric bar, the communicative context (cooperative vs. incidental movement) could be manipulated, and whether the temporal sensitivity within a short postdictive window (approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 150\,\textrm{ms}$$\end{document}) shifts with these manipulations could be tested. This agenda has implications for human-robot interactions: if human gestures influence postdiction more strongly than robot motions do, interfaces might harness or guard against late reinterpretations during collaborative tasks. Limitations We did not include a control condition in which only the mouse moved without the experimenter’s hand being visible or a condition without the presence of another person. Because task-irrelevant motion cues can modulate ambiguous motion judgements18, these control conditions should be included in future work to test whether similar effects arise when only the mouse moves. In addition, factors related to the presence of another person in the room, such as arousal, as well as participant-experimenter gender pairings and levels of familiarity, were not measured in the present study and should be examined in future work. We sampled three discrete offsets around the critical overlap (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-150, 0, +150$$\end{document} ms) using slider movements. Whether the postdictive perception based on others’ movements exhibits a broader or asymmetric temporal profile or depends on acceleration profiles and gesture morphology remains unknown. Furthermore, we did not record data that would enable the computation of the acceleration and deceleration of the mouse movement. In addition to mouse speed, future work should include continuous logging of mouse movement to assess these measures. Naturalistic contingencies (e.g., variability in movement onset, peak velocity, or multijoint trajectories) were not manipulated and will be important to map in future work. Conclusions Other persons’ mouse-control movements, although causally irrelevant and context-conditioned, can retroactively bias ambiguous motion perception within a short postdictive window. Framed as a compatibility claim, these findings fit predictive-coding/free-energy views in which context-conditioned priors are integrated with sensory evidence to minimize prediction error within a short temporal window. Our primary inferences relied on prespecified within-experiment contrasts; the cross-experiment Experiment 1 vs. Experiment 2 comparison was exploratory. Consistent with this scope, we found no strong evidence for a large global shift between self-action contexts, yet strict equivalence was not established at the realized power. Supplementary Information Supplementary Information. Supplementary Information.