Fly Worker

abgunaydin/webgpu-fly — its limitations file states plainly: "the connectome does not walk the fly." 20.3M edges reach the body as one scalar plus a turn bias, scaling a hand-written sin(t × 10 Hz) tripod gait.

Xenova/fruit-fly-simulation — line one of the decoder reads "Engineered rate decoder"; the README says the movements are "illustrative, not validated predictions of fly behavior."

snedea/flybrain — README claims "Behavior is not scripted", but the steering code clamps the brain to ±0.05 and computes heading with atan2. Its own comment: "Steering is primarily handled by behavioral biases." The clamp exists "to prevent worker noise from causing spinning."

nftechie/doomfly — the furthest anyone has taken this. 166,700 neurons, 25.6M connections, plus dopamine plasticity: damage drives a 200 ms aversive pulse into two PPL101 cells and rewrites 4,184 KC→MBON11 weights during play. Its README opens with "Status: live experimental training, not demonstrated learned survival. The current v6 candidate failed its visual, conditioning and survival validation gates." Steering is clip((DNp20_R − DNp20_L) × .12, −6, 6), which the author calls "engineered controller assignments, not established natural motor functions." We measured DNp20 in our own model: left/right separation 0.18 Hz, against 0.889 for the channels we find statistically. There is one DNp20 per side — too few to read.

Nobody has done it. Even with reinforcement learning. The honest ones say so.

We went after this properly. Flies see motion, not brightness, so we fed ON/OFF temporal contrast into the anatomically correct cells — L1 for ON, L2 for OFF.

L1 125Hz → Mi1 0.08Hz, Tm3 0.00Hz, Mi4 0.00Hz — ON pathway dead.
L2 125Hz → Tm1 26.0Hz, Tm2 15.5Hz, Tm4 21.0Hz — OFF pathway fine.

The fly's ON pathway works by sign inversion through inhibition: L1 inhibits Mi1, Mi1 inverts again. Our model has no baseline firing, so inhibiting a silent neuron does nothing. T4, which needs ON input, never fires.

Adding tonic drive woke HS cells from 0 to 6.7 Hz — but direction selectivity stayed at zero. Doubling the visual connectome (2.7M → 5.9M edges) left T4 at 0.0 Hz.

The real reason: T4/T5 direction selectivity comes from neuron-specific timing — Mi9 slow, Mi1/Tm3 fast, Mi4 delayed inhibition. Our model gives every neuron the same τ_m = 20 ms, τ_syn = 5 ms. The connectome gives you wiring, not dynamics. No amount of better input fixes that.

A QA fuzzer does not need goal-seeking. It needs tireless exploration at a cost that lets you run it for hours — which rules out an LLM in the loop. A connectome gives you structured, non-deterministic input at zero inference cost, forever. What it does not give you is an advantage over noise: we measured that, and the table above is the result.

The lab game has four bugs planted in it. The detector does not know where they are; it only watches for symptoms.

Not reaching the goal is a limit of the model. Spinning in a corner was our bug. Residual bias and saturation in the steering signal kept the fly circling in one place. Retuning adaptation and gain against map coverage — not against goal-seeking — took it from 16 visited cells to 30 and from 8.3 spins to 1.9.

The second problem was that a wedged fly never recovers — a 6000-frame run covered less ground than a 750-frame one. So when the detector reports stuck, no-progress or frozen, the shift ends and a new one starts. The finding is still logged; this is what a real fuzzer does. Same budget, 24,000 frames, averaged over three repeats: 9% → 37% coverage, 5 runs → 53, 10 findings → 53.

We also tried randomising the starting position, the way fuzzers randomise initial state. It measured worse — some seeds start inside a small dead end — so we dropped it. (That comparison predates the deduplication fix below, so we quote no number for it.)