Today the retina encoder injects rates straight into T4 and LC4 — i.e. C# computes what the optic lobe should compute from light. The optic lobe (lamina → medulla → lobula/lobula plate) is in the checkpoint; whether the uniform-LIF dynamics make it work is unknown and cheap to test.
Deliverables
- Test the optic lobe (CLI, ~1 hour): drive the photoreceptor / lamina layer (R1–R6 or L1/L2) with a moving pattern via
dotfly run --stimulate and check whether T4a–d become direction-selective and LC4 expansion-tuned in this model.
- If it does: photoreceptor-level retina encoder — luminance per ommatidium → photoreceptor rate — and let the connectome compute motion and looming itself (removes
Retina.cs's dark-fraction / image-shift logic).
- Landing triggered by visual expansion (LC4 / Giant Fiber) gated by odor, as in flying flies (van Breugel & Dickinson 2012), instead of "odor high and sugar below".
- Object-approach readouts: LC10 / LC11 → their descending targets, if they respond — a measured "approach dark objects" behaviour (Wu et al. 2016; Keleš & Frye 2017).
- New findings written up in
PLAN.md §12 and docs/findings.md.
Acceptance: no C# computes what a neuron in the checkpoint computes; encoders reduce to physically defined light / odor / wind / contact transductions; behaviour verified in the room demo.
Context: roadmap, measured vs engineered, PLAN.md §11 (F1). Dombrovski et al. 2023 (LC4 → DNp02/DNp04 gradients); Lappalainen et al. 2024 showed the connectome wiring suffices for T4/T5 selectivity with fitted parameters.
Today the retina encoder injects rates straight into T4 and LC4 — i.e. C# computes what the optic lobe should compute from light. The optic lobe (lamina → medulla → lobula/lobula plate) is in the checkpoint; whether the uniform-LIF dynamics make it work is unknown and cheap to test.
Deliverables
dotfly run --stimulateand check whether T4a–d become direction-selective and LC4 expansion-tuned in this model.Retina.cs's dark-fraction / image-shift logic).PLAN.md§12 anddocs/findings.md.Acceptance: no C# computes what a neuron in the checkpoint computes; encoders reduce to physically defined light / odor / wind / contact transductions; behaviour verified in the room demo.
Context: roadmap, measured vs engineered,
PLAN.md§11 (F1). Dombrovski et al. 2023 (LC4 → DNp02/DNp04 gradients); Lappalainen et al. 2024 showed the connectome wiring suffices for T4/T5 selectivity with fitted parameters.