No emotion-specific image-computable code in the human left amygdala

A common inference in computational neuroscience holds that if a deep network trained to recognize construct X predicts a brain region, that region represents X. We test this for the human amygdala with a representation-invariance criterion: a construct-supervised encoder counts only if it beats a capacity-matched non-construct encoder and converges across two independent construct encoders, with a positive control. The anchor is 442 human amygdala neurons: affect encoders match their single-neuron geometry no better than object and low-level encoders, and on most per-session contrasts significantly worse than object encoders, an effect strongest and most consistent in the left amygdala (face-affect encoder minus object, per-session p<.001); the pooled affect-minus-object composite (−0.086, bootstrap-p=.033) is a non-superiority null bounded below the (marginal) affect effect the same instrument detected in medial-frontal cortex. In naturalistic movie fMRI the amygdala shows only a tiny EmoNet advantage over a matched object encoder (+0.0009; significant, about ten times smaller than in fusiform, and robust even to the strongest generic encoder, DINOv2), but a second affect encoder does not reproduce it and the two affect encoders do not converge: it meets the invariance test's first criterion yet fails the second — an encoder-specific, not construct-invariant, signal. Fusiform (vision) and superior temporal (audio) cortex show genuine modality-specific structure. The amygdala response is better described by a modest population bias toward stimulus intensity and salience plus a decision-uncertainty signal. The contribution is a reusable invariance test; the negative finding converges with recent work relocating the image-computable emotion code to cortex.