Figure 2: Persistent support evidence and SSCA-centered calibration in PULSE. In Stage I, HED is fine-tuned on the target support set, and the resulting encoder snapshot is frozen to construct the evidence bank E. In Stage II, the query detector is initialized with the Stage I weights and produces query features Q, base class logits S, and predicted boxes B. SSCA matches Q with the cached evidence E to obtain support agreement s and refines the base logits as Sā€² = S + G āŠ™ āˆ†, while the box coordinates bypass score calibration and remain unchanged. The evidence bank remains fixed throughout Stage II training and inference. Flames indicate trainable parameters during training, whereas snowflakes indicate frozen modules or fixed evidence. All model parameters are fixed during inference. Test images from the DeepFish dataset (Fu et al., 2024) are adopted for illustration and do not correspond to the sampled adaptation episode; the displayed output box is schematic rather than an actual model prediction.