With normalized count weights, keeping unlisted inputs reduced cone route majorities from 265 to 6 in 341 HFseg1 midget cells

We tested whether a dominant bipolar input establishes a dominant cone route in one draft human retinal graph. We used HFseg1 table version 20250805 and 341 selected midget ganglion cells. We independently reproduced an earlier SCI result: 290 cells had one same-polarity midget bipolar source with more than half of all incoming counts. We then traced labelled cone inputs through these sources. Edge weights were counts normalized at each target. A cone route coefficient above one-half gives a sufficient contrast-sign guarantee in a bounded two-stage linear model. With only labelled cone and bipolar inputs in the totals, 265 cells passed, with 150 distinct cone IDs. Retaining unlisted source counts at both stages reduced these counts to 6 cells and 6 cones. Including every recorded input gave zero passing cells. An exploratory stage test gave 44 passing cells when unlisted inputs were kept only at bipolar cells, and 229 when they were kept only at ganglion cells. In the model that omitted unlisted inputs, 147 distinct cones still had a guaranteed majority route when the maximum edge efficacy was twice the minimum at each target. Exact rational arithmetic checked all 3,192 nominal results. Full corner tests checked the efficacy bound. A new run with fresh source downloads reproduced all numerical outputs and the headline figure. These results locate a source-coverage limit on a count-based cone-channel claim. They do not show loss of single-cone receptive fields. Limits include one draft sample, unreviewed cells, assumed count weights, input normalization, and omitted feedback, electrical coupling, and measured noise. The package includes code, source hashes, result tables, proofs, figures, and repeat instructions.

human retina · fovea · midget ganglion cells · cone signals · connectome coverage · synaptic counts · linear models · robust bounds