Fixed weights retained 80 percent efficiency in both states of 107 of 108 selected Allen models with independent noise

We tested one fixed positive readout across first and late 50 Hz response summaries in Allen cortical slice data. Each stored median was assigned as a model mean. Input noise was independent. A second-order cone method chose weights to maximize the smaller efficiency across states. Efficiency was squared signal-to-noise ratio divided by its state-specific optimum. Fixed weights retained at least 80 percent in both states in 107 of 108 selected input groups. First-state optimal weights fell below 80 percent in the late state in 18 groups. Median fixed-weight worst-state efficiency was 98.7 percent for mouse excitation, 98.6 percent for mouse inhibition and 95.4 percent for human excitation. A second optimizer agreed within 3.3e-8 efficiency units. Noise subtraction left only 25 groups, with three below 80 percent. Median-to-mean error, unknown covariance, separate input trials, different summary endpoints and selected slices prevent a neuronal decoding claim. Original code, aggregate evidence, hashes and a figure support reproduction. External records stay outside the bundle.

Allen Institute · synaptic convergence · fixed readout · short-term plasticity · signal-to-noise ratio · conditional model