Separate target tables cut response error by 29% in ten fly recordings at 5 Hz after one gain fit per recording

We tested whether two targets of the fly DM1 projection neuron need separate response tables after each recording has its own scalar gain. We extracted 3,720 events from 93 trials in ten open paired recordings. Each test recording stayed outside table fitting. Its initial trials set one gain; its final trials supplied the test responses. Separate PD2 and PV5 tables reduced mean normalized squared error from 0.33319 to 0.23553, a 29.31% decrease. Eight of ten recordings had lower error. Every complete recording deletion refit kept a positive decrease. A control with equal target-group weights gave a 23.15% decrease. A compressed table, formed from a burst factor and a spike-position factor, gave similar test error. The source study already established target-specific plasticity. This work adds an excluded-recording benchmark and compact response data for circuit models. The all-event calibration was added after an unstable first-response analysis; both results are preserved. The test covers ten bursts at 5 Hz and the first four source spikes per burst. It does not establish transfer to new frequencies, infer dynamics from synapse counts, or identify a causal mechanism. The package has code, event windows, anatomical group keys, source hashes, all predictions, figures and repeat instructions.

connectome · short-term synaptic plasticity · Drosophila · olfactory circuit · lateral horn · response prediction · rank-one model · reproducible benchmark