Limits of learning in connectome-constrained spiking networks: seven boundaries with pre-registered controls

Negative results with rigorous protocols are a scarce and valuable commodity. In a systematic campaign on the MaleCNS connectome-constrained spiking model of Drosophila, we mapped seven fundamental boundaries of learning, each established with pre-registered criteria, kill conditions, and clean controls (no-reward exactly zero on 8/8 seeds). (1) The release boundary: learned odour identity is stored in synaptic weights (decoding 0.71-0.78) and transmitted to the postsynaptic drive (0.97-0.99) but decays at the spike output (0.31-0.64) - an intrinsic LIF-homeostasis limit; seven candidate mechanisms fail to close the gap. (2) No composition: transitive inference does not emerge under any addressing of KC-to-MBON depression - ungated (MCI chain=nochain=model, +0.994), odour-signature (32-D compartment signatures collapse, cos 0.89-0.9998), or anatomical lobe-split (pre-registered model-excess test fails against both no-mediator and swapped-mediator controls). (3) No metric semantics on the MBON readout (Mantel approximately 0). (4) No self-organization without a plasticity rule (delta-NMI=0, 8/8). (5) No absolute valence without an internal anchor: relative taste valence forms in 6/8 seeds, the absolute sign is an attractor coin-flip. (6) The differential-readout trap: odour responses anti-correlate (-0.75), so difference readouts amplify rather than cancel noise. (7) Connectome specificity: the learning canon does not transfer to FlyWire v783. Together with the companion paper's positive results, these boundaries delineate the precise territory of what connectome-constrained spiking models can and cannot learn. Project channel: «Правила игры» — https://t.me/law_of_the_game