At 30 Hz, fitting the strength slope cut squared log error by 16.6% for 438 mouse V1 basket-cell inputs with test dates excluded

We tested whether a fitted first-strength slope improves second-response predictions over a fixed proportional rule with source and target layer controls. We used public mouse V1 input maps at 30 Hz. After source checks, 991 connections remained in 69 maps from 39 dates. We fitted one strength slope per target type. Every test date was absent from fitting. For 438 basket-cell inputs, the fitted slope reduced mean squared log amplitude error by 16.6%. A 95% date resample range with full refitting was 2.2% to 29.6%. Pyramidal and Martinotti cell tests gave no clear gain. The pooled descriptive gain was 1.84%. The basket-cell gains fell to 8.0% and 6.2% at first-amplitude thresholds of 0.2 and 0.5 mV, with ranges that included zero. A second calculation fitted the best common gain within each measured map. It left mean squared vector residuals of 15.9%, 21.2%, and 22.1% for the three target types. These residuals equal readout variance fractions only for the same input pattern, linear addition, and independent equal-variance input components. Noise can cause sublinear slopes and vector differences even when true response gains are constant. The study is exploratory and does not identify a release mechanism. The package gives source data, provenance, code, predictions, full model states, resamples, figures, and repeat instructions.

mouse visual cortex · optomapping · short-term synaptic dynamics · input maps · response prediction · measurement error · linear readout