The Escape Circuit

When the predator strikes, your fly jumps on its own. The escape reflex is a small, pure, deterministic neural module modeled on the real fruit-fly connectome — the LC4 / LPLC2 → Giant Fiber looming-detection pathway.

The circuit

looming predator
   ├── angular velocity ──> LC4   (2,442 synapses)
   ├── angular size     ──> LPLC2 (1,366 synapses)
   │                          │
   │                    weighted sum (real or swapped)
   │                          ↓
   └───────────────> GF membrane potential (leaky integrate-and-fire)
                              │
                        potential ≥ threshold → READY
                              │
                        escape jump

In the MaleCNS v1.0 connectome the Giant Fiber is neuron DNp01, and LC4 + LPLC2 supply ~99.6% of its visual input (Ache et al. 2019, FAFB). The synapse counts are research references: they are used here only as relative weights between the two channels. They are not biophysical membrane parameters, conductance values, or firing rates, and this model does not simulate ion channels.

State

The whole neural state is one plain object. Field names are a compatibility surface — the HUD, window.FlyLabAPI and the saved-game format all read them.

Transitions

All transitions are pure functions of (state, input) — no DOM, no globals, no RNG. Same inputs, same state, every time.

Sensory layer

lc4   = clamp01(vel × 0.45)                                  // LC4, leads
lplc2 = 𝟙[size > 0.08] × exp(−(size − 1.4)² / (2 × 0.7²))   // LPLC2, lags

LPLC2 is tuned as a size-selective band around 1.4 rad — a looming object is most threatening at a particular apparent size, not the largest one.

Integration

real:    (2442 × lc4  + 1366 × lplc2) / 3808
swapped: (2442 × lplc2 + 1366 × lc4)  / 3808

real puts the large weight on the leading channel (velocity); swapped moves it onto the lagging channel (size). This is the comparison condition of the wiring check below.

Membrane

pot ← pot + (target − pot) × min(1, dt × leakRate)
armed ← pot ≥ threshold  AND  (now − lastFire) > refractory

Leak rate is 10 s⁻¹ in the playable loop, 15 s⁻¹ in the fixed-seed assay. Threshold is 0.6 (the giant trait lowers it by 0.08).

Layer boundaries

• The game loop computes the environment input — predator distance to angular size/velocity — advances time, and calls stepGF.
• The neural module performs every state transition.
• The player is the escape decision. The circuit says READY; you (or your agent's brain) press Space. fireGF(state, now) is called by the behavior layer when the escape actually happens: it resets the membrane to 0 and opens the refractory window. One stimulus crossing threshold arms the unit once; re-arming requires both a new integration and the refractory window to have passed.

Naming: the code identifier for the swapped condition is the string "shuffled" — a compatibility surface kept for saved runs and tests; in prose we call it what it is: two weights swapped.

The reproducible wiring check

A single-loom assay — one looming strike with seed-randomized initial distance, velocity and acceleration; escape succeeds only if the GF fires with ≥ 0.18 s of lead before contact. The only difference between arms is the connectivity condition. 200 trials per arm, pinned, reproducible:

Stated plainly: this is a demonstration of reproducibility and sensitivity to channel order within this model — the real condition is defined as the one that puts the large weight on the leading channel, so real-beats-swapped holds by construction; the escape rates are functions of the hand-set parameters (VEL_GAIN, SIZE_PEAK, THRESHOLD, leak rate). The exact values are asserted in the test suite — any change to the circuit that moves them fails the build. The assay itself is deterministic per seed: simEscapeTrial("real", 12345) always returns lead = 0.18333…. A parameter-sweep over how wide a range real beats swapped is not yet done; until it is, this check should not be quoted as a headline result.

Three models, one honest boundary

The world contains three models at different grains. They are related by theme, not shared code — pages and docs must not describe any two of them as "the same model":

The unit contract, pinned by tests

• Identical inputs produce identical neural state (deterministic replay).
• Zero stimulus leaks the membrane monotonically to < 0.01.
• Threshold crossing arms once; firing resets the membrane and enforces the refractory window; the unit re-arms only after it.
• Real and swapped connectivity are distinguishable and reproducible.
• The assay and experiment numbers above stay pinned — any change that moves them fails the build.