Drosophila Connectome Topology Provides No Measurable Language-Model Advantage: A Preregistered Negative Experimental Program

Abstract

Whether the synaptic wiring of an animal brain can serve as an inductive bias for language models is an empirical question, not a metaphor. We tested that question in a preregistered experimental program that transferred Drosophila melanogaster anatomy—and, later, a Drosophila-shaped memory interface—into small language-model and memory computers. Win rules, controls, and seed counts were written before each queue. Across the closed tests, no protocol showed that a fly-derived operator reduced pair-token negative log-likelihood (NLL) or cross-entropy relative to both a no-graph baseline and a non-brain control at the preregistered margin. A washed diffusion prior built from top-degree FlyWire nodes never beat scrambled or random graphs on lexical, factual, or morphological probes and is architecturally near-identity after a forced unit diagonal. Raw syntax kernels of size 64 × 12 and 256 × 12 beat a non-brain graph on holdout area under the receiver operating characteristic (AUROC) on three of three seeds, and the 64 × 12 kernel beat the same control on TinyLM pair-NLL on three of three seeds, but never beat the no-kernel model by the required margin of 0.02. Replacing the TinyLM feed-forward block with a frozen antennal-lobe projection neuron → Kenyon-cell expansion and hard k-winner-take-all was worse than a dense feed-forward network and worse than Erdős–Rényi and degree-matched expansions on three of three seeds. A trainable projectome residual first failed to enter computation under a degenerate zero initialization. After a non-degenerate reparameterization the communication channel was used (relative residual norm ρ ≈ 0.02–0.06), yet fly topology won Quality, Anatomy, and Specificity on zero of three seeds. Using the same expansion as a retrieval hash lost to random expansion plus the same winner-take-all (primary recall@1 at noise 0.2 ≈ 0.31 versus 0.37) and to dense cosine lookup (≈ 0.85). Isolated central-complex heading persistence did not meet preregistered attractor criteria. A silent-engram interface is realizable by construction; a learned recovery operator failed already on training data; a family-bank expression computer that beat weak controls was matched by a generic constrained selector (mean success 0.80 on all learned arms, gap 0, five seeds). We conclude that trainable family-level sparse communication is viable, but the Drosophila projectome topology provides no measurable pair-NLL advantage over vanilla, random or degree-matched top

Keywords