Does the whole brain contain the fly's own steering circuit, reachable from the mushroom body?

EXPERIMENTS.md · lines 10065–10101

E156 — What the whole brain contains, and where the mushroom body's output actually goes

A measurement, not a test: the inventory the mapping has to be designed on. malecns_full, 164,506 nodes, 11,751 types, 25.1 M edges, contact-weighted.

transmitters   acetylcholine 62.4%  glutamate 17.7%  GABA 13.3%  histamine 4.8%
               dopamine 392 nodes   octopamine 101   serotonin 48
output         descending neurons 1,342: DNg/DNge 860, DNp/DNpe 320, DNa 32 (DNa02 x2), DNb 40, DNd 26
navigation     FB 602, PFN 456, PFL 50, EPG 50, PEN 42, ER 282, hDelta 189, vDelta 430, LAL 665
sensory        ORN 2,635 (olfaction)  JO 672 (mechanosensory)  LC 4,257 (visual features)
               MeTu 1,009 → TuBu → ER (the compass input)   medulla/lobula ~50k

Where the readout goes. The 97 MBONs' 38,617 output edges land on CRE 13.5%, SMP 11.8%, MBON 10.1%, LAL 9.2%, CB 6.8%, SIP 4.5%, lateral horn ~9%, FB 3.2%, PAM 2.5%. SMP and CRE are themselves the largest external inputs to the FB (9.7% and 5.0%). So the mushroom body's verdict reaches the central complex directly and through SMP/CRE, and reaches the lateral accessory lobe directly.

What drives steering. The DNa class (which includes DNa02) receives PS 19.6%, LAL 17.0%, GNG 6.6%, CB 6.3%, VES 5.9%, AOTU 4.1%. The CX steering output PFL receives FB 35.2%, FC 15.3%, Delta 12.8%, hDelta ~14%, EPG 2.6%. The fly's own navigation stack — compass (ER/EPG) → PFN → hDelta/vDelta → PFL → LAL → DNa — is intact in the asset.

Modulators outside the mushroom body. Non-MB dopamine (FB tangentials, PPM, LoVC…) projects to AVLP 11.1%, FB 5.8%, KCs, LC, hDeltaB 2.1%. Octopamine (EN/OA/EL, 101 nodes) projects to AVLP, CB, the visual system (TmY/Tm/Pm/LC), ER, SMP, LAL. Serotonin (48) to ER, CB, GNG, SMP, FB.

Sensory entry. ORN → lLN/glomeruli (olfaction, the only pathway the current model uses). JO → AMMC/SAD/CB and directly to DNg/DNge/DNp (mechanosensation reaches descending neurons in one hop). LC → PVLP/AOTU (optic glomeruli → DN). MeTu → TuBu → ER (visual heading into the compass).

Consequence. The only pathway the model wires is ORN→PN→KC→MBON, and the only place it reads is MBON. The paths from MBON to LAL and FB, from FB to PFL, and from LAL to DNa exist with thousands of measured contacts and nothing in the code traverses them for a decision. The run-and-tumble in Forager is doing, in Python, what PFL→LAL→DNa does in the asset.

EXPERIMENTS.md · lines 10103–10109

E156, step 1 of the mapping built. forager/brain/atlas.py looks populations up from the asset's own annotations — 23 named populations as data, selectors by type prefix, family and transmitter, and outside() for what a model wired to a given set cannot see. tests/test_atlas.py pins every E156 count: 4 passed. The test caught two of my own errors before they reached the model: prefix="DNa" also matches the 20 DNae cells (family selector now, 32 as measured), and PPL is 24 cells by type — E156's "22" was the dopamine-labelled subset. Next: readout across LAL + DNa (step 2), prediction first.