Run length and current side explained 99.96% of 16 output means in a released zebrafish memory model

We tested whether published model outputs preserve input history before the latest switch. We used 16 late-interval means from a seven-subgroup zebrafish memory model. Eight groups defined by current side and observed run length explained 99.9649% of variation across those means. Side alone explained 50.9892%. The final two sides explained 92.2563%. Four run groups contained one condition each. The fit was descriptive and used the same means for fitting and assessment. Small differences remained between histories with equal side and run length. The source lacked trial identities and covariance, so these differences did not establish or exclude an exact reset. We proved that a complete reset, followed by history-independent repeat transitions, removes information from before the switch. Under stronger assumptions of independent identical absorbing binary subgroups, we calculated exact single-trial decoding accuracy. Seven subgroups could distinguish true runs of three and four with at most 67.42% accuracy over fixed recruitment probability. At least 62 subgroups were required for 90% accuracy in that ideal model. These limits do not bound the source spiking network or animal memory. The package preserves source-linked facts, proofs, code, figures, provenance and repeat steps. All 23 numerical checks passed. A fresh source download reproduced the numerical results.

connectome · zebrafish · working memory · neural integrator · reset · run length · stochastic recruitment · single trial decoding · bistable subgroups