With free amplitudes, time splits reversed the error order of two fixed decay fits for 635 mean zebrafish responses

We tested a functional comparison used for a zebrafish eye movement connectome model. We used 635 released mean response curves, with 12 samples from 2.333 to 6.000 s after an eye movement. Two fixed descriptions each had two free amplitudes. One used 10 s and 4 s decays. The other used a constant and a 4 s decay. The constant model gave 0.71% less squared error when three time blocks were omitted in turn. It gave 20.30% more error when the first half predicted the second half. Source coefficient bounds changed the block difference to 0.083% more error for the constant model. With no bounds, independent sample errors of equal variance imply coefficient error correlation -0.986585 from the fit design alone. The observed free-fit correlation was -0.972268. A noiseless constant plus one 4 s decay control gave fitted correlation -1.000, with at most 1.084% relative RMS error. This control shows that negative coefficient correlation does not require a finite slow decay. It does not estimate the noise in the recordings. The forward test retains information about finite decay shape. The file lacks animal IDs and separate trials, so we report no animal-level uncertainty. The source time constants were also informed by these data. This is a limit on one functional comparison, not a rejection of the source circuit model. Code, source hashes, primary predictions, all score tables, figures and repeat steps are preserved. A fresh source fetch and a clean repeat matched all nine result files.

connectome · zebrafish · eye movement · decay fitting · model validation · baseline offset · coefficient correlation