A tuned nonnegative correction from early responses raised late mean prediction error by 2.3% in ten fly recordings at 5 Hz

We tested whether early voltage responses improve later mean response prediction within identified fly olfactory connections. We used 93 trials from ten paired recordings of DM1 projection neurons and two lateral horn target groups. Each test trial stayed outside fitting and penalty selection. The baseline learned separate means for each connection, burst, and source spike position. A correction added one shared nonnegative slope times the early response deviation. Mean squared error rose from 0.026144 to 0.026751 mV squared, an increase of 2.32%. Tuning removed the correction in 92 of 93 tests. The only nonzero correction raised error in a recording with three trials. A within-recording early-block exchange test gave a conditional upper-tail value of 0.892 from 999 exchanges. A first-spike check gave a 3.35% decrease, with a descriptive recording range from -6.26% to +11.25%. Three other specified checks gave no decrease. The result limits this correction at ten bursts and 5 Hz. It does not establish absent trial state, predict from anatomy alone, or identify a release mechanism. Animal grouping is unknown, and valid event masks use the completed trial. We preserve code, event data, anatomical group keys, source hashes, all predictions, model choices, results, figures, and repeat steps. Independent raw-source and numerical checks passed.

Drosophila · structure and function · synaptic dynamics · trial variability · paired recordings · lateral horn · response prediction