ABSTRACT Classical fear conditioning describes how neutral cues acquire a threat value, yet how learned associations are retrieved and generalised across similar stimuli specifically is an ongoing debate. We combined behavioural ratings, physiological measures, and fMRI in a two-day classical fear conditioning paradigm to characterize acquisition, retrieval, and generalisation across modalities. Twenty-five healthy participants completed acquisition trials on Day 1 and retrieval and generalisation trials on Day 2 using conditioned (CS+, CS-) and graded generalisation stimuli (GS). Outcomes included trial-wise US expectancy ratings, pre/post fear and arousal ratings, skin conductance responses (SCR), pupil dilation, and ROI-based fMRI (amygdala, hippocampus, insula, periaqueductal gray (PAG), locus coeruleus). Acquisition yielded robust CS+/CS-discrimination in behavioural ratings and increased BOLD responses in bilateral insula and PAG. During retrieval, US-expectancy ratings indicated early retrieval of CS contingency. The fMRI results showed greater BOLD activity during CS+ presentations than during CS-presentations in bilateral hippocampus, left insula and right PAG. Additionally, hippocampus-insula coupling increased. Critically, parametric modulation during retrieval revealed that trial-wise mean US-expectancy modulated BOLD responses in left insula and right PAG, with a trend in left hippocampus. Across generalisation, US-expectancy and pupil dilation responses followed graded profiles, which could be explained by a Gaussian model, whereas SCR generalised, but was not captured by a Gaussian model. Parametric modulation by US-expectancy correlated with BOLD activity in left PAG, with a trend in right hippocampus. Stimulus identity explained variance in bilateral insula and left PAG. Findings converge on a hippocampus-insula-PAG network that retrieves learned predictions, and scales defensive output according to similarity-based threat probability, linking subjective, physiological, and neural outcomes.