For a century, neuroscientists could only watch a fly’s brain. Now you can rebuild it — and make it act. A month ago our simulated fly’s escape reflex was half-alive: the command neuron fired, the jump never came. Today the full pattern works — jump and takeoff, at the angle a living fly uses. This post is about the part nobody expects to be interesting: the assembly manual.

Two halves, one brain

Every rule in our simulation is calibrated against a live-fly measurement, one brain function at a time. The taste circuit wants synapse weights exactly as the connectome counts them. The vision circuit wants them normalized. Both settings are validated. They are also mutually exclusive — a global switch kills one half every time. So we enumerated every sane way to assign rules to brain regions: a lattice of ~10 profiles, all measured, none discarded. Both trivial schemes fail. The working boundary exists — and it is razor-thin.

Why the boundary sits where it sits

The failing variants share one circuit. A raw mass of taste-adjacent neurons drives a raw relay, the relay floods the visual lobe, and the cells that detect motion direction drown in symmetric noise. Break either leg of that circuit — harmless. Both — blindness. We found this by full search plus bisection, and every prediction was written down before the runs. Several were refuted; each refutation narrowed the mechanism. That is how we know the boundary is not luck.

The jump, the takeoff, the trigger

With the boundary fixed, the missing hands went back on. In a live fly, one spike of the giant fiber (the command neuron) fires the jump muscle in under a millisecond, electrically — but electrical synapses are invisible to electron microscopy, so the connectome has none. We modeled them as a parameter anchored on that live timing. Jump muscle: fires in 24/24 runs, 1 ms. One synapse further, the dataset lists the takeoff relay’s neurotransmitter as “unknown” — which had silently deleted all its outputs; the primary literature says cholinergic, we fixed it. Takeoff: 24/24 runs, 1 ms, zero missed spikes.

And the trigger: a live fly jumps while the shadow is still growing, past ~40° of visual angle. Ours used to fire at the peak. A 90-point parameter surface, calibrated to that anchor: now it fires in-growth at 36–45° across four shadow speeds.

The honest part

The taste response at the calibrated point is half of its old level (the mechanism that fixed the sugar dose curve trades peak for monotonicity). About 2,000 background spikes in silence resist every cure we tried without breaking the trigger — measured, documented, staying. And our fly is near-deterministic, while live ones gamble.

What it means

See shadow → jump → take off. The full escape pattern now exists in a simulated brain assembled from two independently validated halves — with every number traceable to an immutable run ID. Not a bigger computer. An assembly manual, tested against the animal.

Assembly of the ALL profile from two calibrated leaves (taste: gate 1, T-014; vision: line B16), lattice enumeration of assignment schemes, the X→Y loop mechanism, and the escape-circuit repairs (ADRs 014–018). All numbers below are from canonical seed grids on desk-asag; every claim traces to a run ID in validation.md.

Assembly lattice

Rows of the weight matrix W_in are assigned {raw, row_median_fanin} per atomic set: A (taste cone, postsynapses of GRN + MN9), U (its presynapses), D (third hop down), V_min (graded visual front + escape detectors), Rest. A=raw and V_min=norm are fixed by measured ends (all-v1 / all-v2); 8 informative combinations were all run (all-v1, v2, v3a–d, v4b). Verdict: no interior point keeps both leaves; the working boundary (all-v5a) is (A∪U) minus feeders of V_min — 34,627 raw rows.

X→Y loop (T-057, T-063)

Raw X (33,171 rows) × raw Y (2,126 visual relay neurons, 2.06M raw weight X→Y) recruits Mi1 (726→1,598 active bodies; recruited bodies carry maximal Y input, p=7×10^-18) into the minus-DSI class; T5c/T5d selectivity survives (gate is blind) but low-speed direction detection degrades (30°/s p=0.146; 60°/s p=1.9×10^-2). Either leg raw alone: harmless (v5a bitwise = base on gate carriers; y1, y3, ytop, yrest all clean). Recruitment occurs only for the y1×y3 conjunction.

Escape repairs

GF→TTMn/PSI (ADR-014): electric arms absent from EM; modeled as per-pair instantaneous copy, weight 11.0 = 1/α + noise margin (anchor: one GF spike → TTM potential <1 ms). TTMn 0/24 → 24/24, arm latency 1 ms at all rv.
PSI→DLMn (ADR-016): predicted_nt(PSI)='unclear' → sign 0 → outputs deleted; cholinergic per Chai 2025 (Dα7 genetic confirmation). Layer 1 alone measured insufficient (87–413×); synapse efficiency 413.0 = (1/α)×1.1 / weakest pair 0.0266. DLMn 0/24 → 24/24, PSI→DLMn 1 ms, zero misses (365 = 5 bodies × 73 PSI spikes).
Reliability (ADR-017): GABA wave (IN00A029/036) deterministically pushes 37/48 PSI membrane states below threshold; per-post arm weight 12.9 = max over 48 states (11.85) + 1.0 margin. Onset DLMn−TTMn = 1 [1,1] ms in 24/24.

Trigger calibration (ADR-015)

Surface {θ_d × γ}, 90 points × 3 speeds; old point (0.2, 2.0) reproduced to 1 ms. Selected (0.15, 4.0): first GF spike in-growth at 36.0°/44.7°/40.5°/41.3° (medians; range 34.4–45.6°), 24/24, TTMn arm intact. Known cost: post-threshold bursting (38–60 spikes vs 2–5) — stateless nonlinearity cannot express single-spike commitment; adaptation candidates named in ADR-015.

Stability & costs

Boundary stability (T-060): 7 neighbor variants, 2 pass, both degraded; held-out test separates them (one bitwise = base, other shifts trigger by 10–26 ms) — the boundary is not an arbitrary point. Cone storm (ADR-018): sugar-zone runaway ×3.1 suppressed by spiking-output depression; dose curve ws=0.06 restored to monotonic 7/12/16 (was 4/4/4); price: ws=0.05 level ×0.5. Θ-background ~2k spikes in silence (T-067): all cures break the rv80 anchor (+52–92 ms) — structural price, documented. Determinism: bitcheck pairs OK at every step.

Records

Canonical grids and run IDs: validation.md sections “B13, regression No.1/No.2”, “Iteration 3–5”, “T-057”, “T-058”, “T-060”, “T-061”, “T-062”, “T-063”, “T-065”, “T-067”, “T-068”. Profiles: all-v1 … all-v11 in profiles/ (none discarded). Assembly method: metod-perebora-variantov.md.

Our simulated fly escapes a looming shadow at the right moment. We replayed that run with a meter on every relevant neuron — and the famous "shadow detector" cells turned out to be silent. Something else pulls the trigger. And no, you can't fix it by turning up the input.

The whole point of a transparent brain

A simulated brain has a property no animal has: you can rewind any moment and ask why — down to single cells. Post 1 showed the escape jump firing at the right shadow angle. This post is us pointing that microscope at ourselves: the behaviour looks right, but is it right for the right reason?

What should happen

In a living fly, two groups of neurons — LC4 and LPLC2 — are the canonical looming detectors. They watch the eye's output, and when a shadow grows fast enough they fire the giant fiber (GF), the command neuron of the jump. That is the textbook circuit. In our simulation the jump happens on time — but do the detectors actually cause it?

What we did

We replayed the recorded escape runs — bit-for-bit identical to the originals, verified — with instrumentation: for every LC4 and LPLC2 neuron, every millisecond, we measured its charge (how close to its firing point), every input current, and what all of their partner neurons were doing.

What we found

1. The signal gets diluted. The shadow enters the eye correctly and the first layer fires. But each following layer splits the signal among thousands of twin neurons, and unused charge leaks away in about 10 ms. Each detector receives only drops, at different times — no single bucket fills. The best detector reached 45% of the needed charge at medium shadow speed, and 99% at the slowest — one cell crossed once, after the jump had already been commanded.

2. A storm gets in the way. After the shadow's edge passes, the network ignites into self-sustaining hyperactivity — 11.5 million spikes per run, continuing even when the picture is static. Inside that storm the detectors are actively pushed down: their average input is negative.

3. Someone else pulls the trigger. The jump is actually driven by a different neuron — a wide collector (DNp70) that gathers input from thousands of storm cells and sits saturated at roughly 19× its threshold. Right behaviour, wrong mechanism.

Why not just turn up the input?

Because the tap feeds everyone. More input means a bigger storm and stronger push-back on the detectors — they lose harder, not softer. And we already know where that road ends: our very first weight attempt (gate 1, profile v0) produced an "epileptic" brain that answered everything alike — a failure we published. The honest fix is plumbing, not pressure: scale connections by how many inputs a neuron actually has, calibrate the eye against real measured firing rates, add gain control so storms die out. That is our next milestone.

The honest part

This is a finding about our model — a reproducible observation in one simplified simulation on real wiring. In living flies, LC4 and LPLC2 really do drive the jump; our model currently imitates the behaviour without the mechanism, and the whole point of the rig is that this can be seen, measured, and fixed rather than hidden. Real: the complete connectome, every number above, public run IDs. Props: deliberately simple neurons, synapse counts standing in for synaptic strengths, an eye bridge whose gain is not yet calibrated. Full methods and data in the "For scientists" layer.

Bit-exact instrumented replays of the T-017 looming runs show why LC4/LPLC2 stay silent at profile v1: their driver layer (T2, T4/T5, Tm5Y) never crosses threshold (dilution of the L2 transient across thousands of columnar bodies; no direction-selective temporal structure in a uniform 1 ms-delay LIF), mean input to LC4 is net inhibitory (−1.9…−2.2 threshold units), and GF is driven by DNp70 saturated at ~19× threshold inside a self-sustaining storm (447k AVLP spikes during the 1 s stimulus hold with zero sensory input).

1 · Method

Instrumented replay of recorded runs (same graph, profile v1, seed, scene): per-step membrane v and synaptic input i_syn for all LC4 (126), LPLC2 (185), DNp01/GF (2), DNp70 (2) bodies; spike counts for partner groups; fan-in tables from the connectome. Replay equality with the stored runs verified bit-for-bit (escape-chain spikes and whole-network per-step counts). Runs: rv=40 r20260916T054147Z-9229f4, rv=10 r20260916T054310Z-b04a33 (seed 20260914; parents r20260914T095747Z-13dfd3, r20260914T094907Z-768395).

2 · Measured (rv=40 / rv=10)

Driver spikes per run (rv=40 / rv=10): TmY3 6783/5824, Tm2 1286/1086, Tm4 260/448, Tm3 260/270 — but T2 0/0, T4a-c 0/0, T5a-d 0/0, Tm5Y 2/1, Tm20 0/0. L2 (the OFF cartridge input): 2226/2052, all during edge passage; L1 fires only in the tail (shadow retreat). Across all 24 original runs: LPLC2 fired in 0/24; LC4 in 6/6 slowest-locom runs (rv=10) — always 83–84 ms, i.e. 1 ms after GF (82–83 ms) — and 0/18 at rv ≥ 20.

3 · Mechanism

Dilution: available excitation onto LC4/LPLC2 is plentiful (+12k / +16k threshold units), but their drivers are columnar and numerous (T2: 1630 bodies; T5c: 1720); per-body weight from a single partner is a few threshold units, arriving asynchronously, while τ=10 ms integration forgets. The motion layer that would sum coincidences (T4/T5 with Mi1/Tm3 delay structure in the real optic lobe) has no temporal differentiation in this LIF (uniform 1 ms/synapse). Inhibition: LC4's mean i_syn is negative; top available inhibitors Li28 (−370), MeLo8 (−311), LT35 (−224). Winner: the divergent detector paths lose, the convergent diffuse path (Dm→MeVP→AVLP→DNp70, 2 bodies) saturates. Divergence divides, convergence sums.

4 · What we think would fix it (candidates, untested)

• fan-in normalisation of weights (compensate columnar dilution);
• calibration of the visual front-end against published lamina/medulla firing rates (the bridge's gain 6·ΔL and lag 20 ms are unmeasured assumptions);
• gain control / feedback inhibition so post-stimulus storms decay.

Each of these changes the model and will go through versioned profiles and an ADR — not silent tuning to a wanted answer.

The first real milestone: a simulated fly brain that tells sugar from poison — using nothing but the wiring and one rule per neuron.

A fly steps on a sugar drop

A fly tastes with its feet. Sugar touches the leg, and — if the brain says yes — the proboscis extends to drink. We asked the simplest possible question of our simulated brain: does this behaviour fall out of the wiring alone, with no part of the answer hand-written?

What we did

We took the complete wiring diagram of a real fruit fly — 166,700 neurons, 25.6 million connections — gave every neuron a one-line spiking rule, and touched the simulated leg: a steady current into the 46 sugar-sensing cells. Then we watched the two motor neurons that command proboscis extension. For comparison: the same current into 75 bitter-sensing cells, into 17 water-sensing cells, into nothing at all — and into 46 random neurons.

What happened

Sugar: the feeding motor neurons fired — more sugar, more spikes. Bitter: zero spikes in all 108 bitter trials. Water: zero here (a known limitation, see below). No input: total silence — the brain has no make-believe background chatter. And the random control: zero. Touch any 46 neurons and nothing happens; touch the sugar set and the fly reaches for food. Edible versus inedible fell out of the wiring.

The honest part

Science already knew this and more: Lappalainen et al. and Shiu et al. (Nature 2024) demonstrated sugar/bitter separation in similar simulations — Shiu's team even predicted and confirmed shared sugar/water circuitry in living flies. Our goal is not to discover this, but to build a rig where every spike is accounted for and the whole run is reproducible bit-for-bit from a seed. The video above is a re-run of one branch — its spike count matched the original exactly.

Real: the connectome, the cell sets, every spike, every number — public run IDs. Crutches: weights are synapse counts times one scale; signs are predicted neurotransmitters; the neurons are deliberately simple — a skeleton, honestly labelled, not physiology. Full methods in the "For scientists" layer.

Stimulation of sugar-sensing leg neurons drives MN9 (proboscis-extension motor neurons) in a full-connectome LIF simulation; bitter, water, empty and random controls do not. Gate passed on the v1-gate profile across a weight_scale ±20% × noise grid; 432 trials total (6 configs × 72 branches, 600 ms each).

1 · Setup

Full MaleCNS v1.0 graph (166,700 neurons, 25.58M edges), LIF, 1 ms tick, one global weight scale. Stimulus: DC current into gustatory receptor neuron (GRN) bodies, amplitudes 1.2/1.4/1.6 (just-suprathreshold GRN rates). Cell sets from Tastekin et al. annotations: sweet = LB3b/LB3c/LB4a/LB4b (46 bodies), bitter = LB1a–e + LB2a–d (75), water = LB3a (17), none = empty, random = 46 random neurons (seeded, bodyIds in manifests). Readout: MN9 (2 bodies). Biological prototype and criterion (any non-zero MN9 firing = extension): Shiu et al., Nature 634:210 (2024).

2 · Results (median MN9 spikes per 600 ms, [IQR] over 6 seed branches)

Sugar is dose-dependent; bitter is zero in all 108 bitter trials at the working point (single spikes appear in one ws=0.06 corner); empty input is brain-wide silence even with σ=0.05 noise — no spontaneous activity is an assumption of the profile, not a virtue. Random control: MN9 = 0 in all 18 cells (network activity ~1.5–2.8k spikes, same order as water) — the response is specific to the sugar set, not to "any 46 neurons". Separation AUC (Mann–Whitney, sweet > bitter, over branches): 0.92–1.00 at ws ∈ {0.05, 0.06} across amplitudes; at ws=0.04 both conditions sit below threshold (AUC 0.5–0.83, honestly reads as no difference).

3 · Why it works: two-hop balance

With glutamate sign −1 (v1-gate, following Shiu et al.), the wiring itself carries modality at the MN9 two-hop balance: sweet +708, bitter −88 (excitatory path +434, inhibitory −522), water +563 synapse²·spike⁻¹. A first attempt with glutamate excitatory (v0) produced epilepsy: any suprathreshold stimulus drove ~23k spikes/step and MN9 answered all modalities alike — a published failure, see gate 1 in the project's validation journal.

4 · Determinism

Follow-up runs with the same seeds reproduced the sweet-cell medians and min–max exactly; bit-for-bit branch equality is the core contract (tested at every commit). The video in the story layer is a re-render of one branch: spike count identical.

5 · Known limitations (from the validation journal)

• Water gives no PER here, although water on the labellum does elicit extension in living flies: LB3a has only 17 bodies vs 46 sweet, so equal per-neuron current means weaker total input — a stimulus-calibration issue, not wiring (the water two-hop balance, +563, is positive).
• DC current instead of Poisson spike trains; recurrent inhibition shuts GRN tonic firing after a transient.
• "Extension" = any MN9 spike (as in Shiu et al.); no motor bridge yet.
• Glutamate sign −1 is an assumption following Shiu; uniform per-synapse weight; `stim_spikes` in results.npz duplicates the network counter (known script bug at run time).

We showed a growing shadow to a complete, simulated fruit-fly brain — and it escaped exactly when a real fly would. Then we looked inside and found the signal had taken a road nobody expected.

A shadow, and 87 milliseconds

To a fly, a dark disc growing in the sky means one thing: something is diving to eat it. In our simulation, with the fastest shadow we tested, the nerve cell that fires the jump muscle produced its first signal just 87 milliseconds after the shadow started growing — about a twentieth of the time it takes you to blink. Nobody told the brain when to jump. It decided on its own.

The brain in numbers

The brain we simulated is not invented. It is the actual wiring of a real male fruit fly, mapped cell by cell with an electron microscope and published openly by scientists: 166,700 neurons and about 25.6 million connections. We gave every neuron the simplest possible rule for firing a spike, switched on the shadow — and watched.

It behaved like a real fly

Real flies jump when a looming shadow covers about 40 degrees of their visual field — not after a fixed time, but at a fixed size, whether the shadow is fast or slow. Our simulated fly did the same: across 24 runs at four different approach speeds, the jump always came when the shadow covered 32–46 degrees. The brain was also completely silent before anything moved — a fly's world is made of changes, not pictures.

Then came the surprise

Here is why we sat up straighter. The textbooks say the escape signal enters the "giant fibre" — the fly's famous escape nerve — from two families of visual cells called LC4 and LPLC2. In our runs, those cells never fired a single spike before the jump. The reflex happened anyway. When we traced the actual cascade of spikes backwards, the signal had travelled a different route entirely, ending in the strongest input to the giant fibre that almost nobody talks about.

We are not claiming the textbook is wrong — our model is deliberately simple, and we publish its every assumption. But this is the promise of a simulated brain: you can rewind it, trace it to single cells, and catch it disagreeing with what you thought you knew.

Rewind and run it again

Because it is a simulation, we can live the same moment many times. Replay the same shadow with a different random seed, and the fly jumps the same way at a slightly different instant — escape is not a switch, it is a distribution. Every run is public: each number in this post traces to a downloadable file with the full spike log.

What is real here, and what is simplified

Real: the wiring, the eye geometry, every spike time, every number — reproducible from public run IDs.

Simplified: the neurons are one-line caricatures of real membranes; chemical signs are predicted, not measured; electrical synapses are invisible to this kind of microscope data and are absent; and the "unexpected route" is an open question, not a discovery claim. The full list lives in the "For scientists" layer above.

We ran a looming shadow through a simulation of the complete fruit-fly connectome — 166,700 neurons, no hand-tuned delays anywhere. The jump reflex fired when the shadow subtended 32–46°, which is what living flies do. But in our runs the canonical input pathway to the giant fibre (LC4/LPLC2) stayed silent: the signal reached it through a different route, carried by the wiring itself. Every number traces to a reproducible run with a public ID.

1 · Eighty-seven milliseconds

A dark disc grows following the standard collision law. At r/v = 10 ms, the jump motoneuron TTMn fires its first spike 87 ms after growth onset, when the disc subtends ~40°. Everything between photons and muscle is wiring from MaleCNS v1.0 — we designed none of it.

2 · The eye is not a camera

The stimulus is sampled only at real ommatidia positions taken from the dataset (876/892 lamina cartridges per side; 5.4° spacing). Input is phasic: ΔL over a 20 ms lag written into L1 (ON) / L2 (OFF) — the lamina carries contrast, not absolute luminance. Static scenes produce zero input: 0 spikes brain-wide before stimulus onset in all 24 runs.

3 · The wiring

166,700 neurons, ~25.6M connections. The escape circuit as the field knows it (synapse counts from the dataset):

Textbook: looming enters the GF from LC4 (angular speed) and LPLC2 (angular size), which are necessary for GF-mediated escape in vivo (von Reyn et al., Current Biology 2019).

4 · Model

LIF, fixed tick 1 ms, τ = 10 ms, refractory 2 ms; weights = weight_scale · synapse_count · nt_sign with one global scale (0.05, calibrated once against a sugar→MN9 proboscis gate); neurotransmitter signs from dataset predictions (histamine zero — R1–R6 do not spike here); channel noise σ = 0.05. No delay was ever tuned: every hop costs exactly 1 ms by definition.

5 · Results

Matches to the living animal without fitting: jump at a fixed angular size rather than fixed time; TTMn follows GF by 4–5 ms; dead silence until the stimulus moves.

6 · The unexpected part

In all 24 runs, LC4 and LPLC2 never spiked before the jump. Tracing the spike cascade backwards from the GF:

DNp70 is the strongest single input to the GF in this dataset (799 synapses) — larger than the looming cells. We do not claim this route carries the reflex in vivo; it is a property of this model on this wiring, and it is a concrete, reproducible observation (open question B8), not a vibe.

7 · Rewind

Same moment, different seed: same reflex, slightly different times — a distribution, not a switch. All 24 runs public, e.g. r20260914T100249Z-f5e9car20260914T100249Z-f5e9canbsp;(full spike log, .npz) (r/v = 80 ms: threshold at 695 ms, jump at 731 ms).