Signal direction changed decoding order in ten matched Allen circuit model sets

We tested whether a mean noise variance rank gives a constant-signal decoding rank in selected Allen slice circuit models. We used 247 mouse VisP and 17 human TCx graph sets. Fixed labeled degrees, class totals, selected motifs and signed spectra left alternatives in ten sets. Every one of their 19 state pairs favored different unit-norm signal directions under optimal full-state Gaussian decoding. Twelve pairs were not equivalent under class-preserving cell permutations. At gain 0.15, the largest snapshot contrast was 0.245099 times the unconnected model information. Coordinate and uniform inputs gave much smaller contrasts. For a 1000-unit stationary average, the largest normalized contrast fell to 0.000247. An exact calculation gives the same limiting information rate for all stable wirings under full observation and fixed positive definite input noise. This rate identity is a linear-filter control, not a new biological law. Independent covariance checks passed. Assumed weights, signed optimized inputs, common decay, full observation, slice sampling and no linked activity data limit biological use. Code, aggregate evidence, provenance and external retrieval instructions support reproduction.

Allen Institute · linear Fisher information · circuit structure · temporal averaging · signal direction · linear model