Nonnegative weights cost less than 0.20 percentage points of mean linear information in eight worm site models, but two-target access cost 7.7 to 23.4

We test whether high linear information retention in eight partial worm connectome models needs negative readout weights or access to many targets. Each source has independent, equal-probability site releases and unit effects on listed targets. We derive a cone projection test for the exact nonnegative optimum and enumerate the exact best readout with at most two targets. Nonnegative weights reduce mean retention by 0.012 to 0.193 percentage points relative to unrestricted weights. At most two targets reduce it by 7.7 to 23.4 points. In the two adult maps, only 35 and 36 of 141 source cells retain at least 90 percent of unrestricted information with two targets. Nonnegative readouts retain that fraction for all 141 in each adult map. All 3303 source-noise cases pass optimality and independent optimizer checks. We also validate all 265 valid three-by-three binary toy matrices. Thus target access has a larger mean cost than weight signs in this model. The result assumes independent uniform release, unit currents, and perfect separation of each source. Release and coding were not measured. The maps cover one species and a partial brain panel. We provide code, proof, aggregate results, exact input hashes and a retrieval script; external annotations remain outside the bundle.

connectomics · C. elegans · polyadic synapses · linear information · nonnegative readout · sparse readout