A Drosophila connectome substrate inside an LLM agent: what it gates, what it does not, and what it costs

Draft v3 - 2026-10-15. Applied case study with full cost accounting, negative results and control arms. All numbers re-derived from the artifact corpus on the authors' machine. We report an applied case study that wires a whole-organism connectome substrate (MaleCNS v1.0: 169,700 neurons, 25,582,938 innate synapses) into an LLM agent loop as a fast, cheap front layer - a digital cerebellum that acts on tasks it has consolidated and escalates to the LLM otherwise. Two experiment lines are reported. In a foraging simulation, closed-loop evolution lifted population survival from 13/24 at generation 0 to 24/24 at generation 9 on two independent lineages (mean energy 6.7x and 5.8x), and the evolved population transferred to an unseen world seed at 24/24 with higher energy still. In an applied task line (8-fly populations, an 18-task operational battery plus 3 held-out tasks, three arms totalling 1,029 generations, 26,678 LLM calls, 17,005,656 tokens), the substrate's one measurable contribution is a gate: under a contrast readout the gate is alive from generation 4 and sustains 40-55 of 144 possible solo (zero-token) task executions per population per generation, with a best fly gating 8 of 18 tasks. All 24 held-out solves came from fresh LLM code generation (320-334 tokens each), and a fly with 365 learned synapses on an essentially empty graph passed 180/180 checks: the solutions come from the LLM plus a verification loop, not from the connectome. We further report the token bill as structural (a fixed 72 calls and ~47k tokens per generation caused by cache eviction, flat across 365 generations), and a per-fly policy governor that cut LLM calls 116 to 57 (-51%; 58 to 24, -59%, on the paired w150 runs) at exactly zero meal loss - the governor's per-fly meals equal the no-escalation baseline's meals row for row across 24 paired fly-runs. In the economy gate test, satiety-gated feeding yields 8/8 survivals where always-feeding yields 0/8. Three candidate roles for the substrate are judged dead by measurement: recursive memory horizon 1-2 steps, input relay equivalent to a linear copy, and zero channel margin under superposed cues. The paper is offered as an honest accounting case study: where a biological substrate helps an LLM agent, where it does not, and what the whole arrangement costs.