Related work

flymsg asks whether the measured MaleCNS wiring, under the fixed LIF model of Shiu et al. (Nature 2024; parameters in model), recruits circuits whose function is known from experiments. The checks are criterion v4 in validation. Integration stays at dt = 0.1 ms.

MaleCNS counts

Berg et al. (Cell 2026, doi:10.1016/j.cell.2026.08.015) released MaleCNS v1.0. The annotation table holds 211,577 bodies. flymsg build keeps the 165,122 with status == "Traced" and the edges between them: 25,563,197 connections and 124,025,046 synapses. The dropped bodies are orphans, glia, fragments and unassigned segments. Both rows are the compaction table in data.

BANC

Bates et al. (Nature 2026, doi:10.1038/s41586-026-10735-w) map one female central nervous system, brain and ventral nerve cord together (BANC). The paper reports 150,841 neurons brought to backbone-proofread status. That count was read from the article's methods.

flymsg has no BANC loader. Sex comparisons stay on MaleCNS against FlyWire FAFB, as in comparison.

A body on the same wiring

Vaxenburg et al. (Nature 2025, doi:10.1038/s41586-025-09029-4) publish a whole-body Drosophila in MuJoCo. FlyBody (TuragaLab) is the open-source model:

https://github.com/TuragaLab/flybody

Issue #8 compares that model with a one-bit body. The selected task is the looming jump in behavior. FlyBody is not a dependency and stays unselected (issue #19). The body that ran is FlyGym's NeuroMechFly (embodiment). The one-bit protocol is frozen, and its closed loop fails the silence check. Issue #20 left the residual TTMn_R partner unnamed. Silence-scan v2 named DNp02 and DNp06, and v3 replicated that naming (behavior).

N1: specific checks need the measured partners

Fixed 2026-10-10, before the confirmatory comparison. The census below is that one permutation. The comparison was run on 2026-10-11. The same seed was run again the same day so the five passing rows could be named; the census and the pass counts matched.

The question is whether criterion v4's specific checks still pass when each neuron keeps its degree and its outgoing weights but loses its measured partners. One shuffle, on the traced MaleCNS graph (data/edges.parquet after build), same default model (w_syn 0.192 mV per synapse, th_jump 0), criterion v4, seeds 0, 1 and 2.

Shuffle. Permute the post column (wiring.shuffle_targets, numpy Generator.permutation, seed 0). Weights stay on the presynaptic side, so each neuron keeps its out-degree and the multiset of its outgoing weights. The multiset of post ids is unchanged, so the multiset of in-degrees is unchanged. The sign of an edge is the presynaptic transmitter, so it stays with pre. Self-loops and duplicate (pre, post) pairs are counted and kept. The simulator sums the weights of a duplicate pair into one matrix entry. One shuffle. No second seed is drawn after the first is seen. No p-value. There is no flag for another seed, another w_syn, or another th_jump.

flymsg wiring-null --stats prints that census and does not simulate. On the traced MaleCNS graph, seed 0 gives 423 self-loops and 84,560 duplicate pairs (87,365 extra edges) among 25,563,197 edges. flymsg wiring-null runs this comparison on one core.

Confirmatory rule. Every specific check that passes on the real graph fails on the shuffle. A specific check that still passes is a finding: under this null that case does not depend on who connects to whom. w_syn is not retuned. On the shuffle, positive, order, dose and stability are reported and not scored.

Result (2026-10-11, flymsg wiring-null, one core). The real graph passes 29/29. The shuffle passes 5/29, and those five are the stability check of every case: self-sustained activity still dies out. All 7/7 specific checks fail on the shuffle, so the confirmatory rule holds. positive, order, dose and negative fail on the shuffle as well. w_syn stays 0.192. The permutation was not replaced.

Multi-draw families (degree-preserving, cell-type-preserving, weight-within-type, sign) are a separate protocol. They do not replace this one-draw comparison. See nulls.

Already open

T3 was run on 2026-10-11. The chain is supported; the counts and p-values are in dimorphism. T4 was run the same day: silencing AVLP717m drops pIP10 after P1 (p = 0.001). The held-out case ORN_DM1 → DM1_lPN fails specific and stability, and both sensitivity corners fail the same way. The record is in benchmarks. This note does not restate the rules.