PrologueA brain boots on a Mac mini

This book is a lab notebook: the complete FlyWire connectome of an adult fruit fly, running as a spiking neural network on a Mac mini in the living room. We poked it for nights on end. These are the experiments, the figures, the surprises, and the dead ends — all of it.

In October 2024 the FlyWire consortium published the first complete connectome of an adult brain: every one of the fruit fly's 139,255 neurons, and the 54 million synapses between them, traced from electron microscopy by thousands of scientists and gamers. The data is public. So we took the whole thing — the real wiring, with real synapse counts and real neurotransmitter identities — and built it into a signed sparse matrix.

On top of it runs a leaky integrate-and-fire simulation with the parameters Shiu et al. (2024) validated against living flies: resting potential −52 mV, threshold −45 mV, 0.275 mV per synapse, GABA and glutamate count as negative. One second of brain time takes 0.8 seconds of wall time.

Chapter 1 · Graph scienceThe celebrities are all inhibitory

Experiment 001 · experiments/001_hubs.py

We asked the first question you ask any network — who are the hubs? — and every answer came back a brake. The most connected neurons in the fly brain all exist to suppress.

We ranked all 139,255 neurons by total synapse count (in + out). The answer has a theme:

106 neurons have more than a thousand partners. We'll meet many of them again — APL in particular is destined for a virtual-lesion experiment: what happens to memory sparseness when you silence the brain's biggest brake?

Chapter 2 · Virtual lesionsCutting the brake: an APL lesion

Experiment 002 · experiments/002_apl_lesion.py

We cut the brain's biggest brake — APL, one neuron per hemisphere — and the memory code saturated on cue: a landmark optogenetics result reproduced with two lines of linear algebra. The dose sweep also leaked an accidental discovery that shapes half of this book.

Chapter 1 crowned APL: one giant GABAergic neuron per hemisphere, blanketing the mushroom body. In real flies, silencing it destroys sparse coding — activity smears everywhere and the fly can no longer tell similar odors apart (Lin et al. 2014).

The lesion is brutally simple: zero out APL's rows and columns in the matrix — the connectome equivalent of a laser. Then puff a virtual odor (a random subset of antennal-lobe projection neurons at 100 Hz) at both brains and count which of the 5,177 Kenyon cells respond.

Chapter 3 · DynamicsThe brain is a loaded spring

Experiment 003 · experiments/003_phase_transition.py

We pulled chapter 2's loose thread: dial every synapse by a gain factor and the brain snaps between exactly two states — silence and wildfire — with nothing in between. The connectome's own weights sit far above the edge. The anatomy is a loaded spring.

The network's response barely depended on how hard we pushed it, so we mapped that properly. The protocol: scale every synapse in the brain by a gain factor g (nominal Shiu weight = 0.275 mV per synapse = g 1.0), kick the sugar neurons for just 100 ms, then let go — and watch whether activity dies out or feeds itself.

This is the signature of a first-order phase transition with bistability — the same math as a supersaturated vapor waiting for a nucleation site. And it reframes the model's biggest flaw as its most interesting property: at nominal parameters the connectome is a loaded spring. Real flies clearly don't live at 5M spikes/s, so whatever keeps a real brain subcritical — adaptation, synaptic depression, neuromodulation, inhibitory plasticity — is precisely what the wiring diagram alone doesn't capture. The anatomy wants to burn; physiology spends its life saying not yet.

Chapter 4 · DynamicsYou can't put out the fire — and the brake burns first

Experiment 004 · experiments/004_adaptation.py

We gave every neuron fatigue, hoping physiology's favorite damper would relax the spring. It can't — the fire simply rotates through rested tissue — but the failure exposed something better: under fatigue, inhibition breaks first, and the fly eats the poison.

We gave every neuron spike-frequency adaptation — each spike adds b mV of self-inhibition that decays over 100 ms. Real neurons do this; it's the classic mechanism that makes them respond to change rather than steady state. Then we repeated the kick-and-release at nominal gain.

Verdict for the model: the sane operating point isn't reachable by adaptation alone — inhibitory architecture, not fatigue, is what a subcritical brain needs. Next branches: either give inhibition structural help (E/I-specific gain), or leave dynamics for a while and go hunt the ring attractor in the central complex.

Chapter 5 · ComputationThe compass is written in the wiring

Experiment 005 · experiments/005_ring_attractor_structure.py

Attractor theory demanded a cosine connectivity kernel for the head-direction compass two decades before this data existed. We sorted the raw anatomy by heading address and there it was: corr = 0.94. Nobody tuned anything.

The fly carries a compass in its head: EPG neurons in the ellipsoid body hold a single bump of activity whose position tracks heading — a physical ring attractor, one of the most beautiful objects in neuroscience (Kim et al. 2017). Theory says a ring attractor needs exactly two ingredients: local excitation (neighboring headings support each other) and long-range inhibition (opposite headings suppress each other) — a cosine-shaped connectivity kernel. The community annotations give each EPG a wedge address (EPG_L1…L8, R1…R8 → heading angle), so for the first time we can just… look.

We took the 46 wedge-addressed EPGs and the 36 Delta7 interneurons (the proposed inhibition layer), and computed two matrices: direct EPG→EPG connections, and the signed two-hop path EPG→Δ7→EPG. Then we averaged connection strength as a function of the angular offset between the two neurons' headings.

Structure confirmed. The obvious next question is dynamics: if we drop a bump of activity onto this ring in the full simulated brain — surrounded by the wildfire of Chapter 3 — does it persist? Does it drift? Can PEN neurons rotate it like real turning does? That's the next experiment.

Chapter 6 · ComputationThe bump lives — and it leans left

Experiment 006 · experiments/006_bump_dynamics.py

We switched the compass on. Eighty-two neurons of raw anatomy held a memory of heading for a full 900 ms after the cue vanished — and drifted the same −18° counterclockwise every single time. The ring works, and it carries a hardware fingerprint.

Chapter 5 found the compass's blueprint; now we switch it on. We isolated the 82-neuron ring circuit (46 EPG + 36 Delta7), gave the EPGs gentle tonic background drive (the stand-in for ring-neuron input they'd receive in vivo), flashed a 100 ms cue at heading 0°, and let go for 900 ms.

Chapter 7 · ComputationWe could not turn the wheel

Experiment 007 · experiments/007_steering.py — a dead end, documented with respect

With a stable bump in hand we tried to steer it, the way a turning fly does. Three escalating attempts, zero degrees of rotation: in this model there is no operating point where the compass turns. The failure taught us where structure ends and physiology begins.

With a stable bump in hand, the next move was obvious: steer it. In a turning fly, PEN neurons on one side of the protocerebral bridge fire harder, and their ±45°-shifted projections back to the ellipsoid body drag the heading bump around — that's how the compass integrates rotation. The connectome has the players: 43 glomerulus-addressed PENs (22 left, 21 right). We added them to the ring circuit and drove one side.

Everything failed, in an instructive sequence. Attempt 1: add PENs as-is — the EPG⇄PEN loop is so strongly recurrently excitatory that PENs hit 477 Hz at rest and the bump shredded instantly (coherence 0.52 → 0.07). Attempt 2: hyperpolarize PENs, −4 to −40 mV below rest — no effect; the bump's own excitation blows through any static bias. Attempt 3: gate PEN output instead (velocity-gating, closer to biology) and apply a 50% left/right asymmetry:

We're leaving the compass here with its dignity intact and switching branches — Chapter 8 zooms all the way out: community structure and left-right mirror symmetry of the entire 139k-neuron graph.

Chapter 8 · ArchitectureTwo eyes, one mind

Experiment 008 · experiments/008_communities.py

We handed the raw graph to a community-detection algorithm — no anatomy, no labels — and it drew the textbook back at us. And the map came with a headline: vision is computed twice, once per eye; cognition is computed once, together.

After three chapters inside one circuit, we zoomed all the way out and asked the graph itself to draw the brain's map. Leiden community detection on all 139,255 neurons and 2.5M weighted connections — no anatomy, no labels, just wiring density. Sixteen seconds later it handed back 13 major communities covering 95% of the brain (modularity 0.67).

This gives the project a map. Every community is a future chapter: C5's auditory/mechanosensory block, the ventral-cord motor interface, C6's sight-to-decision seam. Depth next: pick the seam and trace how vision enters the mind.

Chapter 9 · ArchitectureSight enters as words, not pixels — and the mind talks back

Experiment 009 · experiments/009_visual_seam.py

We surveyed the border where sight enters the mind. Everything the fly ever sees squeezes through a 10:1 funnel — not as pixels but as a vocabulary of named feature channels — and a third of the border traffic flows the other way. Vision is a negotiation, not a pipeline.

Chapter 8's map showed vision and cognition as separate continents. This chapter surveys the border crossing. First attempt embarrassed us: using only the optic label we found a laughable 2,449 crossing synapses — because FlyWire's taxonomy names the seam itself: 7,684 neurons classed visual_projection whose entire job is carrying the eye's output inward, and 522 visual_centrifugal neurons carrying messages back out.

Nine chapters in, the atlas keeps paying out. Candidate next moves: simulate a looming stimulus through LPLC2 and hunt for the giant-fiber escape reflex (structure→dynamics again, but on a pathway with a famous 60 ms behavioral deadline), or profile the descending-neuron bottleneck — how 100k neurons of brain compress onto 1,305 wires to the body.

Chapter 10 · BehaviorOne spike to jump: the Giant Fiber keeps its nerve

Experiment 010 · experiments/010_escape_reflex.py

We raced the fly's most famous reflex inside the full simulated brain: loom fires the Giant Fiber in 4 ms, while an equally violent courtship stimulus — igniting the brain even harder — never fires it once. One neuron keeping its nerve through a firestorm.

Chapter 9 found the eye's vocabulary. This chapter tests the most famous word in it: loom — the expanding shadow of something about to hit you. In a real fly, loom detectors (LPLC2, LC4) converge on the Giant Fiber, the thickest, fastest neuron in the animal; one GF spike = one jump, ~60 ms from shadow to takeoff. The connectome has exactly two GFs, one per side. We ran the race in the full 139k-neuron brain: 300 ms of loom (457 LPLC2+LC4 neurons at 200 Hz) versus 300 ms of courtship (692 LC10 neurons, same drive) as control.

Ten chapters. The arc so far: a brain that wants to burn (3, 4), circuits that compute anyway — memory sparsening (1, 2), a compass that stores heading (5, 6) but can't turn without biophysics (7), a two-eyes-one-mind architecture (8) speaking in feature-words (9), and a jump reflex that keeps its nerve inside the fire (10). Plenty of map left: the descending bottleneck, the auditory block, neurotransmitter economics, energy budgets…

Chapter 11 · ArchitectureThe neck is a conversation — and every dialogue runs at half-duplex

Experiment 011 · experiments/011_descending.py

We audited the neck — the 1,305 wires that carry every command the fly will ever issue. The command layer coordinates with itself, reflex arcs bypass the brain entirely, and the body talks back on every channel.

Everything the fly ever does — walk, fly, groom, sing, jump — leaves the brain through 1,305 descending neurons (DNs) crossing the neck into the ventral nerve cord. Meanwhile 2,362 ascending neurons carry the body's report back up. This chapter audits both directions of the neck.

Chapter 12 · ArchitectureTwice was a coincidence: monologues and negotiations

Experiment 012 · experiments/012_dialogue_law.py — a hypothesis, properly killed

Two interfaces answered at the same ratio, and we smelled a law. There is no law — but the autopsy of the hypothesis found something better: the brain speaks in three registers. Sensation is a lecture, memory is a dictation, thought is an argument.

Chapter 11 ended with a seductive number: the eye answers the brain at 0.53× and the body at 0.54×. Two independent interfaces, same ratio — was 2:1 a design constant? Good hypotheses deserve honest executions. We measured the feedback/forward ratio for every major conversation in the brain: four more named interfaces, then all 40 community-to-community dialogues from Chapter 8's atlas.

Twelve chapters, three of them dead ends or falsifications (4, 7, 12) — which feels like the right ratio for honest work. The register taxonomy suggests the next question: where does the arrow of information actually point? A global flow hierarchy — rank every neuron by feedforward depth, sensory to motor — would put a number on "how deep" the fly's processing goes.

Chapter 13 · ArchitectureThree synapses wide, twenty synapses deep

Experiment 013 · experiments/013_depth.py

We measured how far the fly's thoughts travel. Both answers are true at once: the whole brain is three synapses wide — the median exit sits two hops from sensation — yet the typical signal circulates twenty synapses deep before it matters. A reflex spine wrapped in a deliberation cloud.

Chapter 12 sorted the brain's conversations; this one asks where the information arrow actually points, two ways. First the shortest road: breadth-first search from all 17,550 sensory neurons along every connection of ≥5 synapses. Then the typical road: a continuous hierarchy where each neuron sits one step above the weighted average of everyone talking to it, senses anchored at zero — recurrence included.

A methods honesty note: the ladder inflates wherever circuits are recurrent (the optic lobe's internal loops push it to 21 despite being sensory-adjacent), so we read it as "how much circulation above the senses", not literal steps — that's exactly what makes the 3-vs-20 contrast informative rather than contradictory.

Chapter 14 · ChemistryThe chemical economy: inhibition guards the front door

Experiment 014 · experiments/014_chemistry.py

We audited the brain's six chemical currencies. Dale's law holds, the brake budget is spent at the front door — sensory input is 74% inhibitory, the hot interior only 39% — and the rare modulators buy reach, not volume: 119 octopamine neurons hold a 32× megaphone.

Every synapse in the dataset carries a predicted transmitter — six currencies in the fly's economy. ACh excites; GABA and glutamate inhibit (the fly's quirk); dopamine, serotonin and octopamine modulate. This chapter audits the books.

Chapter 15 · BehaviorShe's listening — but "yes" is not a reflex

Experiments 015/015b · experiments/015_song.py, 015b_latency.py

We played the female fly her courtship song. The simulation rediscovered a real escape pathway we didn't know to look for — sound goes straight to the legs — but her "yes" neuron fired at everything. Escape is a race; acceptance is a state, and a memoryless brain cannot hold one.

The FlyWire fly is female, and her most consequential circuit is in the data: 166 song-tuned ear neurons (JO-B), ten pC1 cells (the courtship-state hub), and exactly two vpoDN — the neurons whose firing means yes (they command vaginal plate opening, the acceptance gesture). We played her song (drove JO-B) versus threat (loom) in the full brain and watched the yes-wire, the egg-laying wire, and the jump-wire.

Fifteen chapters, and a theme has hardened into a thesis: the connectome is the brain's fast skeleton. Reflexes, filters, funnels, attractor scaffolds — all legible in the wiring. States, moods, decisions — the things that make behavior hers — live in chemistry the wiring only hints at. Good candidate for a synthesis chapter.

Chapter 16 · SynthesisWhat the wiring knows

Experiment 016 · experiments/016_synthesis.py — the capstone

Sixteen chapters ago this was a 57 MB download. This is the map we can now draw of a fly's mind — with receipts — and the clean boundary we found: the connectome is the brain's fast skeleton, and everything it couldn't do was a kind of state.

Sixteen chapters ago this was a 57 MB download. What we can now say, with receipts:

Scorecard: 16 experiments, 4 honest negatives, ~30 seconds to simulate a second of brain on one Mac mini. The map above is not a diagram of what we believe — every node, edge, width and color is a measurement.

Chapter 17 · DynamicsThe dial with no safe setting

Experiment 017 · experiments/017_goldilocks.py

Surely the real fly runs at some lower, safer gain than the flammable paper setting? We swept the dial and found the answer: no. Function and flammability walk in through the same door. There was never a cold option that computes.

Chapter 3 left an open question: the paper's synaptic gain makes a brain that ignites from a hundred-neuron spark — surely the real fly operates at some lower, safer gain? We swept the only global dial the model has — synaptic strength, 0.05× to 1.1× — and at each setting ran two tests: ignition (does a 10 ms spark of 100 random neurons still self-sustain 300 ms later?) and function (does sugar still drive the MN9 feeding command?). We were looking for the Goldilocks window: function without fire.

Chapter 18 · LearningWe taught it — and it forgot everything equally

Experiments 018/018b · experiments/018_learning.py, 018b_isolated_mb.py

We implemented the fly's dopamine learning rule on the real graph and taught it to fear an odor. It came to fear both odors equally — because the sparse odor code the rule depends on never existed in our neurons. The amnesia pointed straight at the missing biophysics.

The fly's learning rule is famously simple: a Kenyon cell that fires while dopamine is present gets its synapses onto MBONs depressed. Odor + shock → that odor's KC ensemble writes less onto approach neurons → avoidance. We implemented it on the real graph: two odors (disjoint 40-PN ensembles), condition odor A (three trials, 80% depression of active KC→MBON synapses), test both.

Chapter 19 · LearningOne biophysical fact buys the memory

Experiment 019 · experiments/019_coincidence.py — the redemption arc

We changed one parameter — the Kenyon cells' membrane time constant, 20 ms to 1 ms — and the sparse code appeared on cue. The memory seemed to appear with it. Half of this chapter survived replication; the correction banner above tells you which half, and we left the rest as written because the mistake is part of the record.

Chapter 18 ended with a diagnosis: the mushroom body's amnesia wasn't in the wiring or the learning rule — it was in our neurons' arithmetic. A 20 ms leaky integrator accumulates its way over any threshold; a real Kenyon cell demands its inputs coincide. So we changed exactly one thing: the KCs' membrane time constant — 20 ms → 1 ms, only for the 5,177 Kenyon cells, everything else untouched — and re-ran Chapter 18's conditioning.

The discrimination is real but partial (56% overlap still blurs it; real flies also use APL's winner-take-all and claw-level thresholds). The direction for a future iteration is obvious: stack the second fact — active APL gain — and watch the gap widen.

Chapter 20 · LearningThe brake's bargain: sparser is not better

Experiment 020 · experiments/020_apl_stack.py

We stacked the second biophysical fact — a real winner-take-all brake — and got the textbook's perfect sparse code. The memory vanished entirely. The trade it exposed runs deeper than the fix: a sparser code is a less reliable one.

Chapter 19 ended with a plan: stack the second biophysical fact — crank the APL's inhibitory gain into a real winner-take-all — and watch discrimination widen past 14 points. We swept APL output gain ×1 to ×50 on top of τ=2 ms KCs, found the code's sweet spot, and re-ran the conditioning. The plan failed in the most instructive way this project has produced.

Chapter 21 · CorrectionThe correction: seed 5 fooled us

Experiments 021 · experiments/021_graded_apl.py + retro-checks

A nonsense number forced the audit we should have run two chapters ago: every test, multiple seeds. Chapter 19's celebrated memory averaged out to nothing — seed 5 had flattered the hypothesis. What survives is the code; what died is every conditioning claim; what remains is the meta-lesson.

This iteration set out to finish the learning arc with biology's real design: the APL is a non-spiking neuron, inhibiting KCs in graded proportion to their population activity. We implemented it properly — continuous feedback along the real APL→KC weight profile, spiking APL removed — and it behaved beautifully: reliability stayed high (85–92%) while ensembles shrank, exactly the trade chapter 20 said a graded brake should win. Then we ran the conditioning, got a nonsense number (−9), and did what should have been done two chapters ago: ran every test across multiple seeds.

Chapter 22 · Structure + LearningFifty-eight glomeruli, and a secret kept

Experiment 022 · experiments/022_glomerular.py

The wiring handed us its own parts list — 58 glomeruli, the textbook count, recovered from synapses alone. We built real odors from them, stacked every lesson of the learning arc… and the mushroom body still kept its secret. The last suspect left standing: the dendrites.

Chapters 18–21 kept blaming our fake odors: random PN handfuls instead of the real thing. Real odors are glomerular — each activates a specific set of the antennal lobe's ~58 input channels. The classification table doesn't say which PN belongs to which glomerulus… but the wiring does: same-glomerulus PNs drink from the same receptor neurons. So we clustered the 250 uniglomerular PNs by their ORN input profiles.

Chapter 23 · FinaleWe built the dendrites, and the fly remembered

Experiment 023 · experiments/023_claws.py — the learning saga's last chapter

We stopped turning dials and changed the architecture: each Kenyon cell became a counter of coincident claws, the number 5 read straight from the synapse table. Every number the arc fought for arrived at once — and the fly, at last, remembered.

Chapter 22 ended with an accusation: the mushroom body's secret is per-claw dendritic computation, inexpressible in point neurons at any parameter setting. That's a testable claim. So we changed the architecture instead of the parameters. First, the wiring handed us the blueprint: counting each KC's distinct PN partners gives median 5 claws per Kenyon cell — the anatomy literature's exact number, sitting in the synapse table all along. Then each KC became what it really is: not an integrator but a counter — it fires when ≥θ of its claws receive a spike within a 5 ms coincidence window. Everything else — glomerular odors, KC→MBON readout, the dopamine depression rule, multi-seed discipline — unchanged from the failed attempts.

The learning saga — 18, 19, 20, 21, 22, 23 — is complete: amnesia, mirage, bargain, correction, parts list, memory. It is the project's thesis told as one story: a connectome is not a brain, but it is an extraordinarily honest witness — it will even testify about the dendrites it doesn't contain, if you interrogate its failures.

Chapter 24 · CapacityEncore: the fly doesn't forget — it becomes paranoid

Experiment 024 · experiments/024_capacity.py

We filled the mushroom body until it broke. It never forgot — the 80th memory stored as perfectly as the 1st. What broke was everything else: innocent odors accrue collateral suspicion until every smell in the world reads as punished. The fly doesn't lose its memories; it loses its innocence.

Chapter 23's mushroom body stores one aversive memory perfectly. A fly meets more than one bad smell in a life — so how many fit? We drew 100 odors from the wiring's own odor space (5 of the 38 usable glomeruli each; ensembles of ~22 KCs overlapping just 4% pairwise), trained K of them sequentially with the dopamine rule, and measured — multi- seed — both retention on trained odors and collateral damage on odors the fly never met.

Chapter 25 · ForgettingForgiveness, and the accidental rehearsal

Experiment 025 · experiments/025_forgetting.py

We gave the paranoid fly the cure biology ships — active forgetting — and watched capacity become recency: the brain trades a permanent archive for a rolling window of the recent past. And in the curve, an anomaly we never built: memories with visiting relatives outlived their peers.

Chapter 24 diagnosed the disease — a memory that can only depress ends in paranoia — and named the cure biology already ships: active forgetting. We added the simplest version: after each new memory, every depressed synapse recovers a fraction ρ toward baseline (Rac1-style decay, as a single dial). Then we stored K=80 aversions and asked two questions: what does each memory look like as a function of its age, and what happened to the innocent?

Chapter 26 · Prediction → TestWe tested our own prediction — and it held

Experiment 026 · experiments/026_rehearsal.py

Chapter 25's anomaly became a falsifiable prediction; this chapter ran the controlled test. Confirmed, four for four: a memory whose relatives keep visiting survives almost three times better — rehearsal by proxy, and a protocol a wet lab could run tomorrow.

Chapter 25 noticed something it didn't build: memories followed by related experiences seemed to outlive their peers, because overlapping ensembles re-depress old synapses. A correlation in a retention curve is a hypothesis, not a result — chapter 21 taught us that the hard way. So this chapter runs the controlled experiment the prediction demands: train odor X, then train ten more odors — in one world each shares 2 of X's 5 glomeruli, in the other world zero. Same forgetting rate (ρ=0.15), same number of trainings, same seeds discipline, four independent replications with fresh random glomerulus draws.

Chapter 27 · EnergyThe answer costs nothing: a fly's energy budget

Experiment 027 · experiments/027_energy.py — a new arc: metabolics

We audited the metabolic bill of a decision. Eating and refusing cost the same to within 2% — the choice itself is energetically invisible — and the ledger prices the senses: a smell spike moves as many synapses as seven sight spikes.

Every spike costs roughly 10⁹ ATP. Since the simulator counts every spike of every neuron during any behavior, we can do the accounting neuroscience usually estimates: where does a moment of experience send its energy? We ran the book's founding trio — sugar, bitter, sugar+bitter — for half a second each and audited the bill by brain division and by cell class.

Chapter 28 · Energy × DynamicsThe price of a thought is lowest at the edge of the fire

Experiment 028 · experiments/028_thought_price.py

We priced the same thought at every setting of the gain dial. The decision is ~300× cheaper at the edge of ignition than at the paper's gain — with a valley of expensive stupidity in between. The most dangerous place in dynamics is the cheapest place to think.

Chapter 17 found the dial with no safe setting; chapter 27 found decisions hiding inside a fixed energy bill. Fused: at every gain, what does the sugar-vs-bitter decision cost? Quality = the MN9 contrast between "eat" and "refuse" (chapter 0's founding behavior). Cost = total spikes. Efficiency = contrast per megaspike. Two seeds per point, contrast direction checked per seed.

Chapter 29 · Synthesis IIWhat the second half taught the first

Experiment 029 · experiments/029_synthesis2.py — the closing chapter

The second half changed what the first half means. Four theses, each earned at least twice — ending with the one the whole book rests on: a simulated brain that can generate falsifiable predictions and retract its own mirages is doing the whole job.

Chapter 16 closed the structural survey. Everything since has been dynamics run on that structure — and the second half changed what the first half means. Four theses, each earned in at least two independent ways:

Final scorecard: 30 chapters · 29 experiments · 6 honest negatives · 1 self-caught retraction · 1 confirmed novel prediction · one 57 MB download · one Mac mini. The brain simulates faster than it lives. The playground remains open — but the book, as a book, is told.

Volume II · the eye, the compass, and the loop

Volume I read the fly's smell, memory, dynamics and energy off one 57 MB download. It barely touched the optic lobes — 55% of every neuron in the head. Volume II walks into them, then into the compass that steers on what they see, and then tries to close the loop between the two. Same snapshot 783. Same rules: one experiment per chapter, controls where it's stochastic, and negatives published.

Chapter 30 · VisionThe motion detector is in the wiring — the direction is not

Experiment 030 · experiments/030_motion_detector.py — Volume II opens

Volume II opens in the eye. Blind clustering recovers the ON/OFF split of the motion detectors perfectly — not one neuron of 7,793 misplaced — and cannot find the four directions at any setting. The connectome knows the fly detects motion; only the coordinates know which way.

The fly's elementary motion detectors are T4 (bright edges) and T5 (dark edges), each in four subtypes a, b, c, d tuned to the four cardinal directions of the visual world. Textbook wiring: T4 is fed by the ON channel (Mi1, Tm3 — fast and cholinergic — plus Mi4 and Mi9 as the inhibitory delay arms), T5 by the OFF channel (Tm1, Tm2, Tm4, Tm9). That is a type-level claim, so it is exactly the kind of claim this dataset can be asked to answer without being told the answer. We built a consensus cell-type resolver over FlyWire's free-text community labels (23 canonical medulla/lobula types, word-boundary matched, one vote per label, ties dropped) and asked three questions with three ways to be wrong.

What would have falsified all of this: shared input types between T4 and T5 (claim 2 dead), a clean 8-way clustering matching a/b/c/d (claim 3 dead — and a far more interesting result than the negative), or a census landing on tens or tens of thousands per eye instead of hundreds (claim 1 dead, resolver broken). Volume II opens with the eye's alphabet legible and its grammar — position — explicitly out of reach of the method. Next: what the eye sends inward when it sees motion, and what that costs compared to a smell.

Chapter 31 · Vision × EnergyTwenty-three neurons can move the body — if they smell

Experiment 031 · experiments/031_price_of_sight.py

Chapter 27 priced the senses as wiring; this chapter prices them as action. At the efficient gain, moving the body with vision costs six times the spikes of moving it with taste — the anatomy's 6.8× prediction, recovered in dynamics.

Chapter 27 weighed the two senses as wiring: an olfactory-chain spike moves ~1,682 synapses, an optic-lobe spike ~262 — a smell spike is worth 6.8 sight spikes. That was a static statistic, and static statistics make dynamic predictions: if sight spikes carry 6.8× less influence, then moving the body with vision should cost proportionally more spikes than moving it with taste. The brain has exactly one motor voice — the 1,305 descending neurons (ch. 11) — so the price of an action is measurable: brain spikes per descending spike.

Three stimuli, identical 200 Hz Poisson drive, 300 ms, two seeds: motion (all 11,995 T4/T5 detectors from ch. 30), loom (LPLC2+LC4, the escape channel of ch. 10), sugar (23 gustatory receptor neurons, the book's founding stimulus). Run twice: at paper gain, and at ch. 28's efficient gain of 0.15× where signal must propagate on its own merit instead of riding a global ignition.

What would have falsified this: a cost ratio near 1 (leverage doesn't translate into behavior), or the size-matched control landing on loom's price (population size explained everything). Caveat kept in the open: driving all 11,995 T4/T5 cells at 200 Hz is not a physiological motion stimulus — no real scene excites every direction at once — which is why the claims rest on ratios and per-neuron normalisations rather than absolute budgets. Next: the compass — can the as-wired central complex hold a heading, and what does it do when the eye hands it a moving world?

Chapter 32 · VisionTwo neurons hold the motion detector's leash — and pull it two ways

Experiment 032 · experiments/032_ct1_lesion.py, 032b_mediator_probe.py

We deleted the brain's biggest neuron — two cells holding a quarter of the motion detectors' diet — against a control lesion of 1,536 smaller cells with exactly the same synapse mass. Concentration beat litres. Half of what this chapter first claimed later died (the banner above); the mistake stays in the record.

Two chapters, written months of compute apart, kept putting the same pair of cells on the table. Chapter 1 ranked the brain's hubs and found them all inhibitory, with the largest — about 150,000 synapses each, one per hemisphere — named CT1. Chapter 30 measured what the motion detectors eat and found CT1 supplying 14–16% of every T4's input and 23–29% of every T5's. One neuron per eye is a quarter of the fly's motion detector's diet. So: what does the eye see when you take that leash off?

The test is a lesion, done properly. Drive the detectors' real presynaptic partners (Mi1 + Tm1/Tm2/Tm9 — the ch. 30 input profile, 5,991 cells at 100 Hz for 300 ms) so that T4/T5 firing is a response, never a forced input. Then run three worlds: intact, CT1 lesioned (2 neurons), and a control lesion of other inhibitory optic neurons chosen greedily to match CT1's synapse mass onto T4/T5 exactly — 122,730 synapses either way, delivered by 2 cells versus 1,536 cells. Four gains (0.10/0.15/0.25 and paper gain), two seeds. If CT1's removal did no more than the mass-matched control's, CT1 would be just a big brake among brakes and the chapter would be a negative.

What would have falsified this chapter: the mass-matched control moving T4/T5 as much as CT1 did (it moved them 1.06×), or CT1's lesion shifting the brain-wide spike budget (it moved it 1.004×). What we keep: the biggest hub in the fly connectome is a local, polarity- specific balance-holder, and "how many synapses" is the wrong question — 2 cells beat 1,536 cells carrying the same mass.

Chapter 33 · VisionThe triangle is real at one gain, and the sign flip has a suspect

Experiment 033 · experiments/033_triangle.py, 033b_two_hop_sign.py, 033c_mass_triangle.py

Three runs of the same rescue experiment: mediators by name (nothing), mediators by mass (69% — later voided), and along the way a two-hop sign audit that named the model's own bookkeeping as the suspect. This is the chapter that found its own killer.

Chapter 32 left a hypothesis with its own control attached: CT1 gone → T4 hyperactive → T4 drives the lobula plate's inhibitory intrinsic cells → T5 suppressed. Delete CT1 and the mediators, and T5 should come back. This chapter runs that test three times, because the first version of it was underpowered and the second answer was only half an answer.

What this chapter is, in one line: a hypothesis half-confirmed, two mechanisms eliminated, one of my own experiments caught underpowered, and a methodological crack found in the floor. The triangle is real where the brain is hot. The sign flip is still unexplained where the brain is efficient — and the leading suspect is now the model's own bookkeeping.

Chapter 34 · VisionRetraction: the sign flip was my own bookkeeping

Experiment 034 · experiments/034_dale_ct1.py

The pre-declared retraction test ran, and the retraction happened: the celebrated sign flip was 16,809 wrong-sign synapses in our own matrix, not biology. Enforce Dale's law and the paradox evaporates — textbook disinhibition, both channels.

Chapter 32 reported that deleting CT1 made T4 (ON) burn 3–5× brighter while T5 (OFF) went quieter, and called it a polarity balance: one hub holding two opposite channels apart. Chapter 33 found a suspect for that flip inside the model rather than inside the fly. This chapter runs the test that was pre-declared as a possible retraction, and the retraction is what happened.

What this chapter is, in one line: the model audited its own data provenance and lost a result it liked. That is the fourth time in this project that a finding died of its own control (19, 21, 33, 34), and the first time the killer was the matrix rather than the seed.

Chapter 35 · VisionThe triangle shrinks to a tenth: 69% was rescuing a deficit that never existed

Experiment 035 · experiments/035_dale_triangle.py

Re-run in a matrix that obeys Dale's law, the 69% rescue collapses to 8–14%: it had been rescuing a deficit that never existed. The triangle is real — a tenth the size we sold it at.

Chapter 33's headline was a 69% rescue: delete CT1 and T5 goes quiet, delete CT1 plus its mass-matched inhibitory mediators and most of T5 comes back — the disinhibition triangle, confirmed in the hot regime. Chapter 34 then showed the quiet was an artifact. So this chapter asks the same question of a matrix that obeys Dale's law, and the answer changes shape twice.

What this chapter is, in one line: the same mechanism, correctly measured, is ten times smaller than first reported — and the one claim that survived three attempts to break it is the one nobody has been arguing about.

Chapter 36 · AuditHow far does the sign bug reach? The economics survive; the express lane shrinks

Experiment 036 · experiments/036_dale_economics.py

Two chapters died of the sign bug; before moving on we audited how far it reaches. Verdict: 5.86% of edges flip, aimed almost entirely at vision — the economics survive (6.25× becomes 6.43×), and the founding chemosensory results were never exposed.

Two chapters died of per-edge neurotransmitter signs (34, 35). Before Volume II moves on, the obvious question: what else in this book was standing on them? The economics chapters are the biggest exposure — ch27's leverage table, ch28's efficiency cliff, ch31's price of a visual thought — so ch31's whole ledger gets re-run in both worlds, same stimuli, drive, gains and seeds.

What this chapter is, in one line: the bug that destroyed two chapters is confirmed harmless to a third, with one claim demoted and one measurement rule earned — the audit that says which results you are allowed to keep.

Chapter 37 · NavigationThe ring works, and only alone: a perfect bump that ignores the world

Experiments 037, 037b, 037c · experiments/037_ring_dale.py, experiments/037b_cue_controls.py, experiments/037c_subcircuit_cue.py

The as-wired EB/PB ring holds a heading and follows a cue to within 0.1° of perfect — in isolation. Embed the same 46 neurons in the whole brain and coherence improves to 0.98 while cue-following collapses to 0.4° out of 180°: a textbook bump that is a stuck wedge. The ring is not fragile, it is unprotected.

Volume II's second continent. The central complex is the fly's compass: 46 EPG neurons (8 wedges × left/right) tiling the ellipsoid body, with 36 Delta7 cells — 94% classed inhibitory — supplying the global inhibition a ring attractor needs. Volume I built this ring from an idealised cosine kernel. This chapter asks the harder question: does the actual weight matrix do it? And per ch34's lesson, the test runs in dale_all from the start, not only in the as-built world.

In one line: the connectome's compass is real, more stable than the theory that describes it, and completely defenceless — and the chapter only knows that because the cue-swap control disagreed with every pretty number in the room.

Chapter 38 · NavigationThe brake was never missing: the ring is quiet, localised, and still blind

Experiment 038 · experiments/038_missing_brake.py

ch37 blamed the embedded ring's cue-blindness on a missing brake — ambient excitation it had no way to refuse. The census answers before the simulation does: the brake is already there, 278 ER ring neurons delivering 80.8% of all inhibitory mass onto EPG. Amplify it 4× and the ring becomes exactly the calm single-wedge bump ch37 admired — 479 Hz → 42 Hz, 46/46 wedges → 5.8, coherence 0.08 → 0.31 — and cue separation stays at 3.9° of 180°, 0.4° on the second seed. The hypothesis is falsified by its own successful intervention.

The design was a titration. Inside whole_brain/dale_all, scale the anatomical ER→EPG inhibitory block by k ∈ {0, 1, 4, 16, 64} and run ch37b's cue-swap protocol (cue at 0° vs 180°, absolute bump angle after release) at each level. The mandatory control delivers the same added mass onto the same 46 target cells with the same per-cell totals from 219 random non-ring inhibitory cells — everything matched except which neurons fire it.

Fraction of the hypothesised cause removed: essentially all of it. ch37 attributed the blindness to excitation the ring cannot refuse, and named the level: 479.6 Hz embedded versus 43.1 Hz isolated. The k=4 brake removes 91% of that excess rate and restores the isolated world's coherence and wedge count as well. The cue still does not move the bump. So the embedded ring's cue-blindness is not an excitation-level problem, and ch37's diagnosis — offered as the honest alternative to "fragility" — is wrong in a way only the intervention could show.

What is left standing is narrower and more interesting: the wedge that wins inside the whole brain may be anatomically predetermined — chosen by whichever EPG cells receive the most non-ring excitatory input, a fixed bias that a 100 Hz cue on 12 cells cannot outvote at any level of global inhibition. That is a prediction about the matrix, testable without a single new mechanism, and it is 039.

What this chapter is, in one line: the fix that worked perfectly and changed nothing — a brake that was always in the wiring, restoring every appearance of a compass while leaving the compass pointing wherever the anatomy already wanted.

Chapter 39 · NavigationThe loudest wedge, and the 100× cue: the brake rides along with the drive

Experiment 039 · experiments/039_loudest_wedge.py

ch38's leftover suspect was that the embedded bump sits wherever the anatomy is loudest. Wrong as stated — raw non-ring input mass predicts the winner worse than chance (top-8 hit rate 0.04 vs 0.17) and its correlation with firing flips sign from +0.42 unbraked to −0.55 braked. The reason is one number: a cell's ER brake mass tracks its excitatory input at Spearman +0.86, so turning the brake up inverts the ranking. And the cue is not ignored, only outvoted: it moves the bump 21.8° at 100× the ch37 cue charge and 86.7° at 333×.

Three tests on the k=1 (as-wired) and k=4 (calm-bump) worlds of ch38: rank the 46 EPG cells by non-ring net input mass and correlate with who actually fires; titrate cue amplitude × duration until the bump moves; and name the two cells that ch38's over-braked floor left standing.

Two rules earned. One: in a network where inhibition is recruited by the same input that excites, global gain is never global — every gain change is also a permutation of who wins, which is why ch38's "generic gain" rescue could restore every statistic and still not deliver a heading. Two: when an embedded circuit ignores an input, titrate before diagnosing; the difference between "cannot" and "needs 100×" is the whole finding.

What this chapter is, in one line: the compass is not deaf to the world, it is outvoted by it at a measured exchange rate of a hundred to one — because the fly's brake is wired to whatever is shouting.

Chapter 40 · NavigationThe right door: the embedded compass works after all

Experiment 040 · experiments/040_right_door.py

Three chapters called the embedded ring stuck, then unprotected, then 100× deaf. All three were measuring the doorway, not the circuit. Deliver the same heading the way the anatomy delivers it — as patterned inhibition through ER ring neurons, on top of a non-directional ignition — and cue separation jumps to 169–171° in 6 of 9 seeds at 10× charge: 97% of what the isolated ring achieved, at a tenth of what the injected door needed for 21.8°. The count-matched shuffled-ER control gives exactly 0.0° in 9 of 9.

The fly's compass is not cued by someone injecting current into EPG. Visual heading arrives via ER ring neurons, which are GABAergic — they do not paint the bump, they carve it by inhibiting everywhere the cue is not. So: map the door, then knock on it properly.

What this reverses. ch37's "stuck wedge", ch38's failed brake rescue and ch39's 100× deafness were all real measurements of a badly-cued circuit. The connectome's ring can be told where north is; it just has to be told in GABA. Two rules earned: an inhibitory pathway must be tested as a modulator of ongoing activity, never as a stimulus; and before concluding a circuit ignores an input, check whether it is doing anything at all without one — a spontaneous bump will absorb every cue you throw at it.

What this chapter is, in one line: the compass was never broken, we were knocking on the wrong door — and when we knocked in GABA, six seeds out of nine pointed north.

Chapter 41 · NavigationA latch, not a compass: the bump can be born anywhere and moved nowhere

Experiment 041 · experiments/041_tracking.py

ch40 opened the door. This chapter asks the only question that matters next — can the heading change? Across 176 step-measurements in six arms — aimed ER, shuffled ER, clockwise, counter-clockwise, hold-still, and a global re-ignition kick at every step — the bump never moves: every single reading lands between 165° and 172°, mean error 90.1°, i.e. exactly chance. But the concurrent control shows the door was open once: with the ER pattern present from t=0, 3 of 4 seeds are born at error 0.0° — and then hold that heading through all eight steps. The connectome ring as wired here is a latch: set once, never updated.

The scope caveat, stated plainly. This is a negative about a 46-cell EPG/Δ7 core with ER input inside the full brain, driven by tonic background — not a claim that the fly's compass cannot rotate. It localises the failure: bump formation and bump maintenance are already in the wiring we simulated; bump motion is not, and ch42 goes looking for the cells that carry it.

What this chapter is, in one line: we found the compass can be pointed, then discovered it cannot be turned — and that is the sharpest possible statement of what a heading circuit needs beyond a ring and its brake.

Chapter 42 · NavigationThe missing multiplier was never missing — it was mirrored

Experiments 042 + 042b · experiments/042_pen_offset.py, experiments/042b_mirror.py

ch41 ended by naming the absent ingredient: P-EN, the neurons that are supposed to copy the heading bump back onto EPG shifted by one wedge, left and right shifting in opposite directions, so that driving one side harder rotates the compass. That is an anatomical claim, and snapshot 783 can answer it with no simulation at all. It does: 43 PEN cells (20 PEN1, 23 PEN2), 18 PEG, 7 EPGt carry wedge labels, and 47 of 68 project onto EPG over 434 edges / 9,298 synaptic weight. The multiplier is in the wiring. The first pass then appeared to falsify the textbook — and the reason it did is the most instructive thing in this chapter.

Two lessons, one of them methodological and uncomfortable. The scientific one: the ingredient ch41 declared absent is present, correctly wired, at the right magnitude, with the exact ±1-wedge antisymmetry that generates rotation — so the latch was a scope result, not a verdict on the connectome. The methodological one: this chapter was four hours from publishing a wiring-level negative that was really a coordinate-system bug. What saved it was insisting that a convention be chosen by an independent statistic — concentration — rather than by whether its output matched expectation. A null result is only as trustworthy as the map you measured it in.

Next: 043 puts them in the loop. Add PEN/PEG to ch40's braked world, born through the concurrent ER door, and drive left and right populations asymmetrically as an angular-velocity command. The measurables are deg/s per unit command, whether L and R give opposite directions, and — the controls that decide it — shuffled PEN→EPG offsets (same mass, scrambled wedge map) and symmetric drive (no rotation expected). If the bump now turns, ch41's latch becomes a compass.

Chapter 43 · NavigationThe engine is the steering wheel: why a textbook multiplier still cannot turn the bump

Experiments 043, 043b, 043c, 043d, 043e · experiments/043*_*.py

ch42 found the rotation wiring exactly as the textbook draws it: PEN→EPG shifted by ±1 wedge, antisymmetric across hemispheres, concentration 0.92. This chapter drives it — and fails, five times, each failure diagnosed before the next. The answer is not "the connectome lacks a compass". It is sharper and more interesting: in the as-wired ring, PEN is the excitatory engine that keeps the bump alive, and it runs at the refractory ceiling to do so. The same population cannot be both what sustains activity and what modulates it — sustain and steer are one knob, and sustain pins it.

The shape of this negative matters more than the negative. ch41 said the multiplier was missing; ch42 found it wired perfectly; ch43 finds it fully employed. In this snapshot the EPG→PEN→EPG loop is what keeps the bump burning, and burning costs everything PEN has — 488 Hz, the refractory ceiling, with no headroom left over for an angular-velocity command. A living fly escapes this because it has a variable the connectome does not store: per-cell excitability, set by neuromodulation and by a velocity input that arrives on PEN in the noduli rather than as generic somatic drive. Three chapters have now landed on the same structural fact from different directions — everything working is actively maintained — and here maintenance consumes the very neuron that was supposed to steer.

Next, 044, and it follows from the ceiling itself: a saturated neuron cannot be excited further, but it can always be inhibited. If PEN must run flat out to hold the bump up, then the only angular-velocity command it can physically carry is a subtractive one — suppress left PEN and the right-shifted copy dominates. That is a different sign, a different measurement, and the same controls (shuffled offsets, symmetric suppression, no-drive floor).

Chapter 44 · NavigationBreak half the compass and it gets sharper: the ring is over-excited, not under-wired

Experiments 044 + 044b · experiments/044_subtractive.py, experiments/044b_power.py

043 left one delivery route untried. PEN cannot be pushed — it already runs at the refractory ceiling holding the bump up — but a saturated neuron can always be inhibited. So the angular-velocity command went in subtractively: silence one hemisphere's PEN and let the surviving side's ±1-wedge copy walk the bump. Suppression works as advertised (a side goes fully dark, PEN asymmetry ±452–476 Hz against a +20.7 Hz floor) and the bump does jump. It still does not turn. What it does instead is the finding: with half the multiplier gone the compass becomes five times sharper, coherence 0.071 → 0.361, and the effect does not care where PEN's synapses land.

Five chapters have now asked this ring to be a compass. It can hold a heading (ch40), it cannot update one (ch41), its steering hardware is wired exactly right (ch42), that hardware is fully consumed by keeping the bump alive (ch43), and removing half of it improves the bump (ch44). Read together they say something the connectome alone cannot: this wiring diagram is a compass missing its regulator, not its mechanism. Every ingredient the textbook names is present and correctly arranged; what is absent is the setting — the excitability, the balance, the per-cell gain that a living fly adjusts continuously and a static snapshot cannot store. The structure says what the circuit could do. It does not say at what volume to play it.

The central complex has given what it has to give at this level of description, so 045 opens Volume II's largest continent: the optic lobes, some 55% of all neurons. First move is a gate, not a simulation — a census of what snapshot 783 actually labels (T4/T5 subtypes a–d, their Mi1/Tm3/Mi4/Mi9 inputs, LC4/LPLC2, the giant fibre) — because ch-sleep died of assuming labels exist. If the motion detectors are there with their offset inputs, the Hassenstein-Reichardt correlator can be tested as wiring before a single spike is simulated.

Chapter 45 · VisionThe gate to the optic lobes: eight motion detectors, both arms, and the sign that gives it away

Experiments 045 + 045b · experiments/045_optic_gate.py, experiments/045b_correlator_gate.py

The optic lobes are roughly 55% of this brain and Volume II's largest unexplored continent. An earlier attempt at sleep died of assuming labels existed, so this chapter is deliberately a gate rather than an experiment: census first, science only if the census passes. It passes — all eight direction detectors are named (T4a–d: 1505–1695 cells each; T5a–d: 1494–1576), with both arms of the correlator and a clean ON/OFF split. And the gate found its own bug: the first pass summed signed weights, so the delayed arm came out negative and the verdict misfired. The negative was the biology.

The central complex arc (ch40–44) ended by concluding that a connectome states mechanism but not setting. The optic lobes open with the complementary lesson: here the mechanism is stated twice over — once in the wiring, once in the neurotransmitter signs — and the interesting quantity, direction, is stated in neither. It is stated in geometry the file does not contain.

046 goes after that geometry the only way available: derive columns from connectivity. Two Mi1 cells that drive the same T4 cells sit in the same retinotopic neighbourhood, so the Mi1→T4→(Mi4,Mi9) bipartite structure defines a graph whose embedding is the retina. The measurable: for each T4 cell, is its inhibitory input drawn from a different column than its excitatory input, and is that displacement consistent within a subtype and different between subtypes? Four subtypes should give four distinct offset directions. Controls: degree-preserving shuffles of the Mi9→T4 map, and T5 (whose four subtypes must produce four offsets in the OFF channel independently).

Chapter 46 · VisionA retina with no coordinates: the connectome names four directions by itself

Experiments 046 + 046b · experiments/046_retina_embed.py, experiments/046b_two_axes.py

ch45 left a sharp problem. All four T4 subtypes draw the same synaptic recipe (fast/slow 6.74/5.32/7.41/9.14, cv 0.22), so their four cardinal preferred directions cannot be built out of how much each arm contributes — direction has to live in where the arms come from. And snapshot 783 contains no coordinates whatsoever. So the retina was rebuilt from the graph: two medulla cells that drive the same T4 cells are retinotopic neighbours, which makes co-innervation similarity a lattice, and its embedding an eye. Then, per T4 cell, the vector from its inhibitory input centroid to its excitatory input centroid. The four subtypes come out as two orthogonal axes with antiparallel pairs — 86.5° between axes, 173.6° and 169.6° within them. Nobody told this analysis that vision has a horizontal and a vertical.

This is the chapter Volume II was looking for. Everything before it in the visual continent described structure; this reads function off structure with nothing added — no positions, no physiology, no fitted parameters, not even a simulation. Four labelled cell types, one shared synaptic recipe, and a graph: out come two perpendicular motion axes with opposed pairs, i.e. the four cardinal directions of fly optic flow. ch42 found the ring's ±1-wedge shift the same way, at the scale of 46 neurons; this is the same argument at the scale of 2295, and it survives a shuffle that preserves everything except who talks to whom. Where the central complex needed a setting the file could not store, the retina needs only the file.

047 spends the result. With a retina in hand, a moving grating can be swept across the derived lattice and fed to the real T4 subtypes through their real synapses: does subtype a fire more for motion along +axis-ab than −axis-ab, and does the inhibitory arm's delay produce the asymmetry, as the hybrid HR/Barlow-Levick reading of ch45 predicts? That is a LIF simulation with a direction-selectivity index per subtype, its null-direction control built in, and — following ch27's leverage table — a price tag in spikes for one visual decision.

Chapter 47 · VisionThe wiring's prediction, paid out in spikes — after two wrong stimuli and one wrong control

Experiments 047, 047b, 047c · experiments/047_direction.py, 047b_scale.py, 047c_percell.py

ch46 made a prediction out of pure structure: the four T4 subtypes should behave as two orthogonal axes with opposed pairs. Predictions are cheap until something has to fire. So the derived retina got a drifting grating, the medulla cells became Poisson sources at the phase their own embedded positions imply, their spikes went through the real signed synapses, and T4 integrated as LIF. The answer, at the third attempt: the spike-derived axes sit 89.8° apart against the wiring's 86.5° — and the two controls that should destroy it, do.

Three chapters, three different ways of being wrong: a stimulus at the wrong spatial scale, a control built on a wrong prediction, a statistic with a built-in positive bias. None of them were caught by intuition; each was caught by measuring the thing itself — column spacing against grating period, control against textbook, estimator against its own null. What is left standing is the strongest form of a connectome result: a geometric claim derived from wiring with no dynamics (86.5°), then paid out by dynamics that never saw the claim (89.8°), with the delay and the offset each proven necessary by removal.

048 prices it. ch27 built a leverage table for olfaction — smell spikes bought 6.8× more downstream consequence than sight spikes — but that comparison used a placeholder for vision. Now there is a real visual pathway with real spike counts (≈175,000 spikes per direction judgement at these settings), so the question can be asked properly: what does one visual decision cost, in spikes and in synaptic events, next to one olfactory decision — and does ch28's 300× efficiency cliff at the ignition edge have a visual counterpart?

Chapter 48 · Vision · EconomicsWhat a visual decision costs: 3% above being awake, and no cliff to fall off

Experiment 048 · experiments/048_visual_price.py

ch27 priced olfactory decisions and found them energetically invisible — opposite motor outcomes for the same bill, within 2%. ch28 then found a 300× efficiency cliff in that same circuit: at the edge of ignition, gain 0.125–0.15× delivered the full decision on ~10K spikes where paper gain spent 2.1M. Both results came from one continent. Now there is a working visual pathway (ch47: spike-derived axes 89.8°), so both can be asked again in a place that shares none of the same cells. The invisibility transfers. The cliff does not — and the reason it doesn't is the most interesting number in the chapter.

Two continents now agree on the thesis Volume I kept running into: a brain's decisions are cheap and its readiness is expensive. Where they disagree is instructive rather than awkward — olfaction rewards a gain dial with a 300× saving because its cost is its own recurrence; vision cannot, because its cost arrives from outside through the photoreceptors. The connectome does not state that difference anywhere. It falls out of running the same accounting twice in two places.

049 closes the loop Volume II promised: vision → central complex → descending. The pieces now exist separately — a retina that reports motion direction (ch46–47) and a ring that holds a heading but cannot update it (ch40–44). The question is whether the connectome connects them: do T4/T5 reach the ring neurons or EPG cells by any short path, and does visual motion of the right sign move the bump where ch41's internal commands could not? If it does, the ring's missing regulator may simply have been its missing eye.

Chapter 49 · Vision → NavigationThe compass does not listen to motion: two visual streams, and only one is wired to the ring

Experiment 049 · experiments/049_visual_to_ring.py

ch40–44 left the ring attractor as "a compass missing its regulator, not its mechanism" — every internal command refused, coherence set by an excitation budget. ch46–47 built a retina that reports motion direction. The obvious hypothesis was that the missing regulator is an eye, so this chapter went looking for the wire. It found one — and then found that it is the wrong one. From the motion detectors the compass is three hops away through a needle's eye. Meanwhile a parallel visual stream, which never touches T4 or T5, arrives at the same ring with 23× the synaptic mass. The connectome answers a question I did not ask: not whether vision reaches the compass, but which vision does.

This is the chapter that justifies doing anatomy before dynamics. Had 049 gone straight to a simulation, it would have driven T4a and T4b — the horizontal motion pair, the intuitive yaw command — down a wire carrying 0.5% of the ring's input, watched nothing happen, and filed another honest negative. The census cost minutes and redirected the whole experiment: ask the wiring which door is wide before knocking on one.

049b knocks. Drive the TuBu→ER channel as a landmark command — ER cells weighted by the tubercle input they actually receive — and ask ch41's exact question with ch44's exact scoring: pre/post population vector angle, drift slope, coherence, against a degree-preserving shuffle of which ER cell each TuBu synapse lands on. If the bump moves where internal commands could not, Volume II ends with the compass working: a ring that steers when, and only when, it is given something to look at.

Chapter 50 · NavigationThe wide door was open, and the compass still did not turn

Experiments 049b + 049c · experiments/049b_landmark.py, 049c_arrival.py

ch49 found the compass's real sensory door — MeTu→TuBu→ER, 15,267 synapses, 23× wider than the motion route — and argued the ring was "missing its landmark, not its regulator." That was a prediction with an address, so it got knocked on. Four landmark commands were built by clustering TuBu cells by the compass angle their own ER targets serve (51.8°, 147.1°, 197.8°, 326.6°, vector strengths 0.78–0.95), driven through the real synapses, scored with ch44's metrics over four seeds. The bump did not care. It settled at 260.8°, 255.5°, 233.0°, 225.1° — spread R 0.966, statistically the same place — and the same place a shuffled map (0.979), a charge-matched random ER subset (251.2°) and no command at all (246.0°) put it. My own prediction, tested properly, was wrong.

This is the most useful failure in the book. A prediction was made in ch49 with a named pathway, a synapse count and a direction; ch49b tested it and it died; ch49c then ruled out the comfortable explanation by showing the command lands with 280 Hz of authority. What remains is not a shrug but a sharp, falsifiable claim: in snapshot 783's ring, heading is not an input-addressable variable. Anyone who can move this bump has found the setting the connectome does not contain — and now knows exactly which wire to attach it to.

Chapter 51 · Synthesis IIIWhat a wiring diagram is for — Volume II, told

Chapters 030–050 · 37 experiment files · experiments/030…049c

Volume II walked into the half of the brain Volume I barely touched — the optic lobes, 55% of every neuron in the head — then into the compass that steers on what they see, and then tried to join the two. It ends with an asymmetry too sharp to be an accident: the connectome told us, from structure alone and to within 3.3°, what its motion detectors would compute — and could not tell us, under three different classes of command, where its own compass is pointing. Those two results are not in tension. Together they say what a wiring diagram is for.

Two volumes, one 57 MB download, one Mac mini. Volume I read smell, memory, dynamics and energy; Volume II read sight, navigation and the seam between them. The recurring thesis survived both — every working regime is actively maintained, never free — and gained a sharper edge: what the connectome can be asked, and what it must be told. Ask it what a circuit is able to compute and it answers to within a few degrees. Ask it what the animal is currently doing and it is silent, not because the data are incomplete but because that is not the kind of fact a map contains.

What Volume II leaves open, stated plainly so nobody mistakes silence for absence: the amplitude-matched landmark rerun (ch50's per-cell drive varied ~15× across groups); the descending and motor continent (ch11 only sketched it — a fly that sees a bar and actually turns needs the DN population mapped properly); a systematic lesion atlas instead of the targeted lesions of ch2/ch32; and the settings themselves — if heading is not input-addressable, the honest next project is not another stimulus but a search over gain and plasticity regimes for one in which it becomes addressable. Anyone who finds it has found what the wiring diagram does not contain, and ch50 says exactly which wire to attach it to.

Volume III · the motor continent

Two volumes in, we have read a brain's input side twice: an olfactory continent that learns, and an eye that computes direction with nothing but wiring. Both times the story stopped at the same place — somewhere inside, where a decision presumably becomes a movement. Volume III walks to the exit. There are exactly 1,305 descending neurons in snapshot 783: every command this brain can send to the body has to squeeze through one of them. That is 0.94% of the brain, and it is the whole bandwidth of behaviour.

Chapter 52The 1,305-wire cable: who actually talks to the exit

Anatomy only, no simulation · experiments/052_dn_census.py, 052_report.py

We began where every previous volume began — by asking the wiring which door is wide, before knocking on one (ch49's lesson, learned the hard way). So: take all 1,305 descending neurons, slice their input columns out of the signed 139,255² matrix, and bin every one of their 1,781,009 input synapses by the continent the presynaptic cell belongs to.

Three facts in that table are worth the whole chapter. First: the senses barely touch the exit. Everything sensory added together — vision, touch, smell, taste, the optic lobe's 77,873 intrinsic cells — is 11% of DN input, and 96% of that is the two coarse channels (visual projection and mechanosensory). The continents we spent two volumes inside speak to the body at 0.58% (smell) and 0.15% (taste).

Second: the compass is not a driver. After eleven experiments trying to steer the ring (ch37–50) we can now say what it is wired to do downstream: 2,878 CX cells contribute 4,962 synapses to the entire descending population — 0.28%, less than mechanosensation by a factor of ten. Whatever heading is for, it does not reach the legs by shouting at DNs directly. That reframes Volume II's negative result: we were pushing on a variable that has no direct line to the output either.

Third: the exit talks to itself, and listens to the body. DN→DN coupling (11.4%) is larger than all visual input, and ascending neurons from the nerve cord supply another 10.0% — the descending population is not a set of 1,305 independent buttons but a recurrent layer with proprioceptive feedback folded in. The "command neuron" picture survives only as a special case.

Convergence is steeply skewed, but not to a handful of kings: the top 10 DNs hold 6.5% of DN input, the top 50 hold 24.8%, the top 130 (10% of the population) hold 46.8%, and the bottom half of all DNs shares 6.8% between them. Multimodality is the norm rather than the exception — binning at ≥5% of a DN's own input, 703 DNs draw on two continents, 281 on three, 10 on four, and only 262 are single-source.

The census also gave us the volume's address book. The biggest DN in the brain by input mass is a named one — DNa03 (13,032 synapses, 84% from central cells), followed by DNa02 (12,362) — both from the DNa turning family, both essentially blind (2–3% visual) and fed almost entirely by the brain's interior. The most visual DNs are the DNp family (up to 76% visual projection, including DNp03 at 52% and DNp04 at 67%), the most mechanosensory are DNg cells (up to 65%), and we located the two cells this volume most wanted: the giant fiber, left (matrix index 57,521, 3,504 input synapses, 222 partners, 18% visual) and right (109,258, 4,517 synapses, 275 partners, 20% visual). 46 named DNs — MDN, P9, DNa01–09, DNp01–10, DNg13/16, GF — are now in experiments/052_named_dns.csv with their matrix indices, ready to be driven and ready to be lesioned.

One methodological note, filed because this book prices its own mistakes: the first pass of this chapter printed root IDs that did not exist. DataFrame.iterrows() collapses each row to a single dtype, so int64 cell IDs came out as float64 and silently lost their last digits — 720575940620918789 (DNa03) became …918784 (nothing). Every ID in this volume is read with .at or itertuples for that reason. A wiring diagram indexed by 18-digit integers punishes float arithmetic without ever raising an error.

Next: the giant fiber's own input tree. It sits behind 222–275 partners with a fifth of its drive visual — ch53 asks whether the textbook looming pathway (LC4 and LPLC2) is the visible fifth, and whether the connectome alone predicts an escape threshold.

Chapter 53The escape neuron is held down — and a sign convention we have been getting wrong for 52 chapters

Anatomy only, no simulation · experiments/053_gf_tree.py, 053c_ntaudit.py

The giant fiber is the most famous single neuron in the fly: one cell per side, a looming stimulus, and the animal is airborne in ~5 ms. Ch52 located both copies, so ch53 opened their input trees. What came out was two chapters' worth — one about escape, and one about our own arithmetic.

The escape command neuron's largest single input category is inhibition. Roughly two thirds of everything arriving at the giant fiber pushes it away from threshold, and the heaviest individual partners are unnamed inhibitory central cells (130, 137, 124, 119 synapses each) plus — a detail we did not expect — the other giant fiber and its co-DNs: 274 synapses of GF→GF_L input, net −260, i.e. the two escape neurons and their neighbours mutually suppress. A cell that must not fire by accident is wired as something held shut, not something waiting to be pushed. That is the connectome's version of the known high GF threshold, and it is visible without a single simulated spike.

The looming pathway is there, and it is a fifth of the door: LC4 (21 labelled cells, 177 synapses) and LPLC2 (37 left / 60 right, 305 / 506–518 synapses) together supply 22–28% of the giant fiber's total excitation, exactly the two types the escape literature names. Snapshot 783's labels are asymmetric — LC4 is annotated on the left only, and 417 of the right GF's visual synapses come from visual-projection cells with no type string — so the counts are a floor, not a census.

The audit: our matrix is signed per synapse, not per neuron

Then the anomaly. Read literally, the looming pathway is inhibitory: LPLC2→GF_R came out at net −212 and LC4→GF_L at net −51. Both types are cholinergic and both are excitatory onto the giant fiber in every recording ever made. So we went to look at our own build.

scripts/build_matrix.py signs each connection row by that row's predicted transmitter (connections.csv.gz carries an nt prediction per pre–post–neuropil row), then sums. FlyWire's per-synapse predictions are noisy and disagree with the cell's own transmitter call often enough to matter:

Dale's principle says a neuron releases one transmitter; our matrix let 30.7% of neurons be part excitatory and part inhibitory, because we inherited a per-row prediction instead of collapsing it to the cell. We rebuilt the matrix under the per-neuron convention (data/brain_cellnt.npz, 2,700,513 edges, 1,742,094 excitatory / 958,419 inhibitory) and the difference at the giant fiber is exactly where it hurts:

So the sign of the fly's best-known sensorimotor pathway is a build decision, and we had it backwards. Note what this is not: it is not a claim that Volumes I and II are wrong. The disagreeing mass is 5.5% and it is scattered; every result that survived a degree-preserving shuffle survived because of structure, not sign, and ch46's retinal axes were derived from co-innervation with signs used only to separate the fast and slow arms by type. But two results in this book rest directly on signed sums — ch3's sugar→MN9 feeding command and ch18–26's KC→MBON depression — and they are owed a re-run. That is the next chapter, and it is the discipline the book has kept since ch34's self-found CT1 bug: when the convention changes, you re-pay the results that were bought with it.

Two findings, then. Biologically: the giant fiber is a cell wired to be held down — two thirds inhibition, mutual suppression between the escape neurons, with the looming channel worth a fifth of the door and about a quarter of all excitation. Methodologically: a connectome's sign column is a modelling choice masquerading as data, and per-synapse transmitter predictions violate Dale's principle for 30.7% of the brain's cells if you let them.

Next: ch54 re-pays the bill. Both keystone signed results — sugar→MN9 (ch3) and the looming→escape drive this chapter just re-signed — run under brain_cellnt.npz, and if the escape command now fires when LC4/LPLC2 do, Volume III has its first working sensory→motor transform.

Chapter 54The bill, re-paid — and a fly that escapes, but only below ignition

experiments/054_sign_repay.py, 054b_loom_gf.py · LIF, Shiu et al. params, gain swept

Ch53 changed a convention, so ch54 re-buys the results that were paid for with the old one, and then tries to cash the new one in. Two questions: does sugar → MN9 (ch3, the book's first result) survive Dale's principle, and does looming → giant fiber work now that LC4 and LPLC2 are excitatory?

1. The keystone survives, and gets louder

Proboscis extension survives the sign change and gets 4× stronger — ch3's result was carried by structure, not by our arithmetic. But look at the second row: collapsing 5.5% of synapse mass onto Dale-consistent signs raised total activity by 78%, because the per-edge noise was silently cancelling excitation inside individual cells. The model's effective gain moved. Everything Volume I said about the position of the ignition edge (ch4, ch28) is therefore stated in units of a matrix that no longer exists, and should be read as a shape, not a coordinate — the 300× efficiency cliff is a claim about a cliff existing, not about it living at gain 0.125.

And in both conventions, at paper gain, the giant fiber is silent (2 Hz) inside a brain emitting four million spikes. The cell ch53 described as wired shut is exactly that, even in a firestorm.

2. Looming → escape: it works, and only in a window

Drive all 259 LC4 + LPLC2 cells at 200 Hz and read the giant fiber, sweeping the synaptic gain because at paper gain the network ignites (ch28/ch48's lesson: a transform can only be measured below ignition). GF firing rates, left / right, in Hz:

Volume III has its first working sensory→motor transform: a looming stimulus, delivered to the two cell types the escape literature names, fires both giant fibers at 175–443 Hz while 259 randomly chosen visual projection neurons driven identically fire them at 0–10 Hz — a 40× specificity margin, with no-drive at exactly 0. The transform is bilateral and symmetric (175/245, 350/413, 423/443).

The escape command has a ceiling as well as a floor. GF output rises to a peak at gain 0.15 and then falls — 443 → 363 → 17.5 Hz — as the surrounding brain ignites and the two thirds of GF input that is inhibitory arrives in force. At paper gain the looming stimulus produces 20 Hz in a cell that was doing 443 Hz at one seventh the coupling. So the window in which this fly can escape is bounded on both sides: too little gain and the command never reaches threshold, too much and the network's own inhibition buries it. Ch53's anatomy predicted a cell held shut; the dynamics show what holds it shut, and that it is the network, not a dedicated gate.

The per-edge column is the chapter's second argument for ch53's audit: under the old convention the escape response is unilateral — the right giant fiber fires, the left never does, at any gain. A sign convention decided which side of the animal could escape. No biology says that; a build script did.

One control failed by construction and is reported as such. We ran a degree-preserving shuffle that permutes which LC4/LPLC2 cell supplies which synapse onto GF — it changed nothing (170/257 vs 175/245 at gain 0.06). It could not have: every cell in the driven set gets identical Poisson drive, so permuting among them preserves the drive exactly. It is a null shuffle, not a test, and the specificity claim rests entirely on the mass-matched random-VPN control, which is the one that bites.

Next: the threshold itself. If 259 looming cells give 443 Hz, how few will do? Ch55 sweeps the recruited fraction to read an escape threshold off the wiring — the number of looming detectors the connectome needs before the fly jumps.

Chapter 55How many eyes it takes to jump: an escape threshold read off the wiring

experiments/055_escape_threshold.py · per-neuron matrix · 3 seeds, 0.3 s, GF averaged over both sides

A real looming stimulus does not switch on all 259 looming detectors at once — it recruits more of them as the predator's image grows. So the escape threshold is a recruitment question, and it is one the connectome can be asked directly: drive k randomly chosen LC4/LPLC2 cells and see when the giant fiber goes off.

The curve has a floor, a knee and a ceiling. Fewer than about 30 looming cells cannot fire the giant fiber at all — not weakly, exactly zero, at three gains and three seeds. Then it climbs steeply, and the half-maximum sits at 135 cells at gain 0.06, 110 at 0.10, 98 at 0.15 — which is the number worth keeping: 50%, 39%, 37% of the population. Across a 2.5× change in synaptic coupling, the fraction of looming detectors needed for a half-strength escape command barely moves. The threshold is a property of the wiring, expressed as a fraction of a cell type, not of the gain we chose.

That is the chapter's claim in one line: this connectome jumps when roughly a hundred of its two hundred and fifty-nine looming detectors agree, and it ignores anything fewer than thirty. A cell held shut by two thirds inhibition (ch53) needs a quorum, and the quorum is measurable.

The two detector types are not interchangeable

LPLC2 alone nearly is the escape pathway. 208 cells reproduce 70–90% of the full bilateral response; adding LC4's 51 brings only the last increment. Per cell the two are closer than the totals suggest (1.5 vs 1.1 Hz of GF output per driven cell at gain 0.10), so this is mostly a numbers game — but note the caveat ch53 already flagged: snapshot 783 labels LC4 asymmetrically, and the LC4-only response is correspondingly lopsided (93 left vs 17 right). The lopsidedness is in the annotation, not necessarily in the fly, and we say so rather than dress it as a finding.

An accident worth reporting: the escape recruits the whole brain, at one gain only

Watch total brain activity as recruitment grows. At gain 0.10 it rises smoothly — 2,219 → 5,038 → 11,074 → 20,728 spikes — the escape is a local computation in a quiet brain. At gain 0.15 the same step from 32 to 64 recruited cells takes the brain from 3,044 to 28,915 spikes, a 9.5× jump: the escape command and a global ignition arrive together. The same stimulus is either a private circuit event or a brain-wide one depending on a coupling parameter, with no change in the wiring or the stimulus. Volume I's ignition edge (ch4) and this volume's escape threshold are not independent knobs — at gain 0.15 they are the same event.

Three chapters into the motor continent we have a working transform with a floor (30 cells), a quorum (~100 cells = 38% of the type), a ceiling (network inhibition above gain 0.15, ch54), and a specificity margin of 20–40× against identically driven random visual neurons. That is enough machinery to start breaking things on purpose.

Next: the descending fingerprint. Ch56 drives sugar, looming and bitter in turn and asks which of the 1,305 descending neurons each program actually recruits — how much of the exit does one stimulus claim, and do two different stimuli claim different halves? That map is what the lesion atlas will be scored against.

Chapter 56Five programs, one exit, no shared core — and smell is the only amplifier

experiments/056_dn_fingerprint.py, 056b_size_matched.py · gain 0.10 (below ignition), 0.4 s, a DN counts as recruited above 5 Hz

Ch52 said the exit is 1,305 wires wide. Ch54–55 drove one program through it. Now the obvious question: when the brain does something different, does it use different wires? Five programs — sugar, bitter, looming, mechanosensory, olfactory — each driven alone, each read out as which descending neurons fire.

The named DNs each program recruits read like the literature's index. Looming pulls the giant fiber plus DNp03, DNp04, DNg13 — the escape set. Mechanosensation pulls MDN (backward walking), P9 (turning) and the whole DNa01/02/03/06/09 walking family, and drives 37 of the 110 motor neurons including MN9 at 95/65 Hz. Olfaction pulls MDN, P9, DNa01–03 and DNp07/09/10. Nobody told the model which cells those were; it was handed a wiring diagram and a stimulus.

The overlap is the finding: there is no shared core

Pairwise Jaccard between the recruited DN sets runs from 0.000 to 0.161, and the set of descending neurons recruited by all five programs is empty. Sugar and bitter share nothing. Sugar and looming share nothing. Looming and mechanosensation share 6 DNs out of 43 and 236. The heaviest overlap in the whole matrix is mechano↔olfactory at J=0.161 — two programs that both end in walking.

So the exit is not a bus that every decision shouts down. Five programs address near-disjoint sets of it, which is the labelled-line picture that Volume I found at the input (glomeruli) reappearing at the output, and it is a precondition for the lesion atlas: if the programs were sharing wires, no lesion could break one without breaking all.

The control that changed the story

Those recruitment counts had an obvious confound — the programs differ 100-fold in how many cells we drive (23 vs 2,674) — and the ordering tracked size exactly. So we re-ran everything at matched cell counts, three independent subsamples each:

Two things fall out, and one of them corrects the chapter's own first table. Mechanosensation's 18% was drive size: matched to 259 cells it recruits 27 DNs, not 236 — an eighth. The uncontrolled table would have been reported as "touch commands the exit most widely", and that would have been an artifact of how many receptors we happened to switch on.

Olfaction is the exception, and it is spectacular. Twenty-three olfactory receptor neurons — the same number as the entire sugar program — recruit 118 descending neurons (9.1%) and set off 224,000 brain spikes, while 23 mechanosensory cells manage 3 DNs and 1,773 spikes, and 23 sugar cells 4 DNs and 2,496. Going from 23 ORNs to all 2,281 changes the DN count by five (118.7 → 125). Smell saturates its access to the exit at roughly one percent of its receptor population.

That is the same asymmetry ch27/ch48 found in the energy budget, arriving from a completely different direction: the olfactory continent is recurrent and self-amplifying, so a handful of receptors is enough to reach the whole output layer, while vision and touch are feedforward and pay for reach in receptors. Cost was endogenous in one and exogenous in the other; access is now the same story. Two independent measurements, one architecture.

And a specificity result for free, worth stating on its own: the giant fiber fires only for looming. It sits at 0 Hz inside an olfactory firestorm of 272,000 spikes and through mechanosensory drive that fires 37 motor neurons, then goes to 415 Hz for 259 LC4/LPLC2 cells that produce a tenth as much brain activity. Ch53 called it a cell wired shut; three chapters of dynamics have not managed to open it by accident.

Next, the volume's capstone begins: the lesion atlas. We now have three transforms with numbers on them (looming→GF at 363/428 Hz, sugar→MN9, mechano→37 motor neurons) and proof they run on disjoint wires. Ch57 starts removing the hubs Volumes I and II named — CT1, APL, the central complex, the ascending feedback — and prices each removal in transforms broken.

Chapter 57The lesion atlas: the mushroom body is optional, and the brake is twenty cells

experiments/057_lesion_atlas.py · lesion = every row and column zeroed (silent and deaf) · gain 0.10, 0.4 s · 057_atlas.csv

Volume III's capstone starts here. We have transforms with numbers on them, so a hub can finally be priced the way a neuroscientist prices it: take it out and see what stops working. Nine lesions, three transforms — looming→giant fiber (ch54–55), sugar→MN9 (ch3/ch54), mechanosensory→motor neurons (ch56).

1. The famous structures are optional for these transforms

Delete 5,177 Kenyon cells — the entire mushroom body, the structure Volume I spent eleven chapters inside, 3.7% of the brain — and the escape response is 390.0 Hz instead of 390.0 Hz. Not "within noise": identical to the digit, on the same seed, because the lesion removed nothing the computation used. The same is true of the central complex (2,878 cells, Volume II's ring attractor), of all 96 MBONs, of CT1 and of APL. Six lesions, 8,184 cells, zero measurable effect on two of three transforms.

This is what ch56's empty intersection meant, made physical. A brain whose programs run on disjoint lines can lose its most celebrated structure and still escape a looming predator and still move 37 motor neurons when touched. The mushroom body is where this fly learns; it is not where this fly acts. Volume I's memory results and this volume's motor results are about the same connectome and share almost no cells.

One exception, small and real: removing all 1,750 ascending neurons — the body's report to the brain — costs mechanosensory motor recruitment 4 of 37 motor neurons (11%). The only structure whose removal costs anything here is the one carrying information from the body, not from the brain's own centres.

2. Only the labelled line matters

Remove LPLC2 — 240 cells, 0.17% of the brain, the type ch55 showed is nearly the whole looming pathway — and escape collapses from 393 Hz to 51 Hz, an 87% loss, with the recruited descending set halved (42 → 19). Touch is untouched (37 motor neurons, unchanged). The atlas's shape is therefore brutally simple: each transform is fragile only to its own line and indifferent to everything else.

3. The brake is twenty cells — 0.014% of the brain

The most striking row is the cheapest lesion. Ch53 found the giant fiber held shut by two thirds inhibition; we took its twenty heaviest inhibitory partners out and escape got stronger: 450 Hz (from 393, and 455 Hz is the refractory ceiling — the cell is now saturated), with recruitment broadening from 42 to 53 descending neurons. Twenty cells, one seventy-thousandth of the brain, are the difference between a guarded escape command and a saturated one.

Which is exactly the shape of a safety mechanism: enormous consequence per cell, invisible until removed. And it makes a testable claim that the ch55 sweep can settle — if those 20 cells are the lock, deleting them should not merely raise the ceiling but lower the quorum: fewer than a hundred looming detectors should suffice. That is the next chapter, and it is a real risk, because 393 Hz is already close to saturation and this +15% could be nothing but headroom.

A note on what this atlas does not say. The lesions were single-seed against a 3-seed baseline, the sugar→MN9 column is uninformative here (0.4 Hz — ch54 showed the feeding command needs the ignited regime, so it needs its own pass at paper gain), and "no effect on these three transforms" is not "no function": the mushroom body's job is learning, which this atlas never asks it to do. What the atlas rules out is that these celebrated hubs are on the path from a looming stimulus to an escape command. They are not. The path is 240 visual cells, two giant fibers, and twenty cells holding the trigger.

Ch58: rerun ch55's recruitment sweep with the twenty inhibitors gone, and price the lock in quorum rather than in headroom — plus the sugar→MN9 lesion pass at paper gain that this chapter owes.

Chapter 58What twenty cells actually buy — and a measurement that failed

experiments/058_lock_price.py · part 1: recruitment sweep ±the lock, gain 0.10, 3 seeds · part 2: feeding lesion screen at paper gain, and why it is uninterpretable

Ch57 pulled the giant fiber's twenty heaviest inhibitory partners and escape went from 393 to 450 Hz. We refused to call that a result, because 455 Hz is the refractory ceiling and headroom is not sensitivity. The honest test is ch55's quorum: does a disinhibited brain need fewer looming detectors?

It does. Read at a fixed command-level criterion instead of at half-max, the quorum drops by a third:

Twenty cells buy a third of the fly's escape threshold. Delete 0.014% of the brain and the animal jumps on two thirds of the evidence it used to require — the definition of a hair trigger, and the reason such a brake is worth maintaining at all.

But the floor does not move. Eight looming cells fire the giant fiber at exactly 0 Hz with and without the lock, and 16 cells barely differ (3.9 vs 5.6 Hz). So the two limits ch55 found have different owners: the hard floor belongs to excitatory convergence — too few detectors simply do not deliver enough charge, and no amount of disinhibition invents it — while the brake sets sensitivity above the floor, nearly doubling the response exactly in the middle of the recruitment range (1.9× at 64 cells) and shrinking to 1.2× at saturation. Inhibition here is a gain control on an existing threshold, not the threshold itself. That distinction is invisible in a single-stimulus experiment and only appears when you sweep the stimulus.

Part 2: the measurement that failed, reported as such

Ch57 owed a feeding-lesion pass at paper gain, since sugar→MN9 only works in the ignited regime. We ran it — nine lesions, MN9 as readout — and the numbers looked exciting: MBONs removed dropped MN9 to 280/147 Hz from 437/293, ascending neurons to 130/183. A 40–70% loss of the feeding command from removing memory-readout cells would be a headline.

It is not a result. Before writing it up we measured the intact brain's seed-to-seed variability at paper gain, five seeds:

The intact brain ranges from 200 to 443 Hz on the same stimulus with nothing removed. Every lesion effect in the screen — MBON, ascending, CT1, APL, the central complex, even the twenty inhibitors — sits inside that band. The screen measured Poisson seed, not anatomy, and we are deleting the conclusion rather than the data: 058_feeding_atlas.csv stays in the repo as a negative.

Two things survive it. The positive control works exactly as it should — removing all 408 gustatory receptor neurons takes MN9 to 0.0 / 0.0 Hz and the brain to 1,304 spikes, so the lesion machinery is sound and the feeding line is as labelled as the looming one. And the failure has a cause worth naming: in the ignited regime a single motor neuron's rate is a sample from a runaway avalanche, so its variance swamps anatomy. Below ignition, the same pipeline resolves a 20-cell lesion (ch57) and a 24-cell shift in quorum (above); above ignition it cannot resolve 5,177. Ch28 argued the subcritical regime is 300× cheaper. It is also the only regime in which this model can be measured. Cheapness and observability turn out to be the same property.

The lock is priced: twenty cells, a third of the threshold, a gain control rather than a gate. And the atlas has learned its own limits — every future lesion claim in this volume is made below ignition, with seeds, or not at all.

Ch59 turns the atlas on the exit itself: lesion the descending layer instead of the sensory lines — the 11.4% DN→DN recurrence of ch52, the ascending–descending loop, and the named command neurons (MDN, P9, DNa02) one family at a time — asking whether the output layer is a network that degrades gracefully or a set of independent wires that fail one at a time.

Chapter 59The exit is a network: 19% of the wiring carries 44% of the command

experiments/059_exit_lesions.py · below ignition (gain 0.10), 3 seeds, cells or edge blocks removed · 059_exit_atlas.csv

Ch57 lesioned the brain's centres and found them irrelevant to escape. This chapter turns the same knife on the output layer itself — including something a cell lesion cannot do: cut the descending neurons' connections to each other while leaving every cell alive.

*only one cell in snapshot 783 carries a DNp03/DNp04 label — the lesion is nominal, not the type.

Cutting no cells at all costs 44% of the escape command

Zero the 10,188 synaptic edges descending neurons make onto each other — remove nothing else, kill nothing — and the giant fiber falls from 393 Hz to 220 Hz. The control that makes this a result: deleting 240 whole descending neurons at random, twelve times more destruction by cell count, costs 2% (383.8 Hz). The escape command does not care which DNs exist; it cares that they talk to each other.

Now the anatomy, which we deliberately looked at only afterwards. Descending neurons supply the giant fiber 349 and 432 excitatory synapses (18–20% of its excitation) from 7–8 partners, and essentially zero inhibition. So a fifth of the wiring is carrying nearly half of the function. That gap is the signature of recurrence: those few DN→GF synapses are not a static input, they are a loop — the looming line drives DNs, the DNs drive the giant fiber, the giant fiber and its co-active partners drive them back — so cutting the block removes both the feed-in and the reverberation it sustains. Counting synapses onto a cell underestimates what a connected layer does with them, by a factor of two here.

This is the exact mirror of ch53. There, anatomy alone said the giant fiber was wired shut, and dynamics confirmed it. Here, anatomy alone says the DN→DN block is a minor input, and dynamics say it is half the command. The connectome is honest about what connects; it is silent about what reverberates.

The exit layer amplifies its loudest voice at the cost of breadth

One number in the table looks like an error and is not: with the DN→DN block cut, the giant fiber loses 44% while the number of descending neurons above 5 Hz goes up, 42.3 → 45.3. The block is 66% excitatory by edge count (126,385 excitatory vs 80,095 inhibitory synapses), but it does not merely add gain — it redistributes. Intact, the layer concentrates activity into the command neuron and holds a few others down; cut, three more DNs cross threshold while the one that matters halves. Recurrence in this output layer buys amplitude in the right cell, paid for in breadth — a winner-take-more circuit, discovered by breaking it.

No single command family matters, and touch degrades gracefully

MDN, P9, DNa02 — the named command neurons for backward walking, turning and steering — each removed entirely, and escape does not move by more than seed noise. Nor should it: ch56 showed these programs run on disjoint lines. But the touch column tells a different, softer story. Every exit lesion costs mechanosensory motor recruitment a little: MDN −4%, DNa02 and ascending→DN −8%, P9 −11%, DN→DN edges −19%, 240 random DNs −23%. Nothing is catastrophic and nothing is free. The escape command is a brittle labelled line — 87% gone when LPLC2 goes (ch57), 44% gone when the exit's recurrence goes, untouched by everything else. Touch is a graded population code that loses a few motor neurons for every insult and collapses for none.

Two architectures in one output layer, then: a sniper and a chorus. Volume III's atlas now prices lesions in three currencies — the labelled line (LPLC2, 87%), the brake (20 cells, a third of the threshold), and the recurrence (10,188 edges, 44% of the command) — and only the last of those is invisible to anatomy.

Ch60 closes the loop the volume was named for: with the escape transform mapped, quorum measured, brake priced and exit dissected, put it in time — how many milliseconds from the first looming spike to the giant fiber's first spike, how that latency changes with recruitment and with the brake removed, and whether this connectome's reaction time matches the fly's measured ~30 ms escape.

Chapter 60How fast it jumps: an urgency code, in milliseconds

experiments/060_latency.py · time from stimulus onset to the giant fiber's first spike · gain 0.10, 10 seeds per point (a first-passage time needs seeds)

Every chapter of this volume has measured escape in spikes per second. But an escape is not a rate — it is a deadline. So: from the first looming spike, how many milliseconds until the giant fiber fires?

The connectome codes urgency as latency. A weak looming stimulus — 32 detectors, an object still far away — takes 28 ms to reach the escape command and does so unreliably (sd 19 ms). A strong one, all 259 detectors, takes 4.4 ms, with a standard deviation of 300 microseconds. Same wiring, same cell, same threshold: a sixfold range of reaction time produced by nothing but how many detectors agree.

That is exactly the trade a fly needs and exactly what its behaviour shows — short-mode escapes for a fast approaching threat, longer deliberate takeoffs for a slow one — and nobody built it in. It falls out of a threshold sitting under a convergent excitatory pool: more sources, faster charge, earlier crossing.

The number lands where biology lives

Two anchors make the 4.4 ms meaningful. The model's own floor is 1.8 ms per synapse, so a two-hop path (LC/LPLC → GF) cannot beat 3.6 ms — at full recruitment the escape command arrives within 0.8 ms of the anatomical minimum, i.e. this is a direct-path response with almost no integration time. And in the real animal, giant-fiber escapes are the fast ones: a GF spike is followed by takeoff in roughly 5 ms, with loom-to-takeoff latencies measured in tens of milliseconds for slower approaches. Our range, 4.4 to 28 ms, brackets that without a single parameter fitted to behaviour — the LIF constants are Shiu et al.'s, the wiring is FlyWire's, and the stimulus is 200 Hz Poisson on named cell types.

The intermediate hop confirms ch59's story in time as well as amplitude: other descending neurons fire before the giant fiber does, at every recruitment level (2.69 ms vs 4.43 at full drive; 10.2 vs 27.8 at n=32). The DN→DN recurrence that carried 44% of the command is not a late modulation — it is already running one hop ahead of the command it amplifies.

The brake costs nothing at speed, and everything at the margin

Removing ch57's twenty-cell brake leaves the first-spike latency identical to two decimal places at 64 cells and above — and cuts it by 6 ms, with the trial-to-trial scatter falling from 19 ms to 5.5 ms, at 32. The mechanism is timing: inhibition needs its own hops, so it arrives after the first spike has already happened whenever the drive is strong. It cannot gate the onset of a strong escape; it can only sculpt what follows (sustained GF spikes double, 4.5 → 10.4 and 28.7 → 47.0). Near threshold, where the first crossing takes tens of milliseconds, the brake has time to act — and there it costs the animal 6 milliseconds and most of its reliability.

So the twenty cells are not a switch on the escape reflex. They are a margin tax: invisible when the evidence is overwhelming, decisive when it is thin. Ch58 showed they set sensitivity but not the floor; ch60 shows they set latency but not the speed limit. Both are the profile of a circuit that suppresses false alarms without ever slowing a true one — which is precisely the specification you would write for an escape system, recovered from a wiring diagram by deleting twenty cells and watching a clock.

Volume III's transform is now characterised end to end: a quorum (~100 of 259 detectors), a ceiling (network inhibition above gain 0.15), a brake (20 cells, a third of the threshold), a recurrent amplifier (10,188 edges, 44% of the command), a latency law (28 → 4.4 ms with recruitment) and a floor at the anatomical minimum. Six numbers, one reflex, all of them read off a connectome that was never asked about behaviour.

Next: Synthesis IV — what the motor continent added, and what this whole book turned out to be about.

Chapter 61Synthesis IV — the motor continent, and what a reflex is made of

closing Volume III · experiments 052–060 · nine chapters, one reflex taken apart

Volume I went into the mushroom body and came out with a memory that becomes paranoia. Volume II went into the optic lobes and the compass and came out with a heading that survives its own broken parts. Volume III asked the simplest question in the book — how does this brain make a muscle move? — and the answer took nine chapters because the honest version has six numbers in it.

The reflex, in full

None of these were fitted to behaviour. The LIF constants are Shiu et al.'s, the wiring is FlyWire's, the stimulus is Poisson noise on cells someone else named — and the thing that comes out reacts in 4 to 28 milliseconds, demands a quorum, and refuses to fire for the wrong stimulus. That is a reflex, reconstructed from a wiring diagram, with its safety mechanism intact.

Thesis 1 — the exit is labelled lines, so the brain's celebrated centres are off-path

Five sensory programs recruit descending sets whose mutual intersection is empty (ch56). Which makes a prediction, and ch57 collected on it: delete the entire mushroom body — 5,177 Kenyon cells — plus the 2,878-cell central complex, all 96 MBONs, CT1 and APL, and looming→escape reads 390.0 Hz instead of 390.0 Hz. 8,184 cells of the most-studied circuitry in insect neuroscience, removed, with no measurable effect on two of three transforms.

The lesson is not that those structures do nothing. It is that where a brain learns and where a brain acts can be almost disjoint sets of cells, and a connectome makes that checkable. Volume I and Volume III are about the same animal and share almost no neurons.

Thesis 2 — one output layer, two architectures: a sniper and a chorus

Escape is brittle and specific: −87% when LPLC2 goes, −44% when the exit's recurrence goes, indifferent to everything else, and 0 Hz for every non-looming stimulus we tried — including an olfactory firestorm of 272,000 spikes (ch56). Mechanosensory motor recruitment is graded and robust: every single lesion in the atlas costs it something (−4% to −23%) and none of them break it.

Those are two different engineering answers living in the same 1,305 wires, and the connectome distinguishes them without any behavioural data: one transform's mass sits in a labelled line, the other's is spread across a population. Fragility is not a flaw here — it is what specificity costs.

Thesis 3 — the connectome's two blind spots are opposite

Ch53 read the giant fiber's input tree and concluded, from anatomy alone, that it was wired shut. Three chapters of dynamics never opened it by accident. Anatomy was right.

Ch59 read the DN→DN block and would have concluded, from anatomy alone, that it was a minor input — 349 of 1,729 excitatory synapses, a fifth. Dynamics said it carries 44% of the command, and that cutting it costs twenty times more than deleting twelve times as many cells. Anatomy was wrong by a factor of two.

The rule that reconciles them: a wiring diagram is honest about what connects and silent about what reverberates. Inhibition is legible because its effect is local and immediate — a synapse count is a good estimate of a brake. Recurrence is illegible because its effect is a loop gain, and no static count contains a loop. This is the sharpest methodological result in the volume, and it is a warning to anyone reading a connectome without simulating it — which, before this project, included us.

Thesis 4 — you can only measure what does not ignite

Ch58's retraction is the volume's most useful failure. A nine-lesion feeding screen at paper gain produced 40–70% "effects" until we measured the intact brain across five seeds and found it ranging from 200 to 443 Hz. Everything was noise. Below ignition, the same pipeline resolves a 20-cell lesion and a 24-cell shift in quorum; above ignition it cannot resolve 5,177. Ch28 found the subcritical regime is 300× cheaper in spikes. It is also the only regime in which this brain can be observed at all — cheapness and observability are one property, and a brain that lives near the ignition edge is buying both.

Thesis 5 — every safety mechanism is small, late, and invisible until removed

Twenty cells. 0.014% of the brain. They do not gate the escape — they cannot, because inhibition needs its own synaptic hops and arrives after the first spike has already left (ch60). They do not set the floor either — 8 detectors fire nothing with or without them (ch58). What they do is tax the margin: a third of the threshold, 6 ms of latency, and a collapse of trial-to-trial reliability from 19 ms of scatter to 5.5. In other words they suppress false alarms without ever slowing a true one, which is the exact specification of an escape system — and it was recovered by deleting twenty cells from a wiring diagram and watching a clock.

Which is the fourth independent arrival at this book's oldest thesis. Volume I: sparsity is actively maintained (ch20–21), memory needs active forgetting or it becomes paranoia (ch24–25). Volume II: the compass holds because its parts fight each other (ch44). Volume III: the reflex is fast because a brake is standing between it and every ambiguous shadow. Nothing that works in this brain is free, and the maintenance is always cheaper than the function it protects.

What the motor continent did not answer

Honest gaps, for whoever reads this next. MN9 above ignition is unmeasurable with our current instrument, so the feeding lesion atlas remains open. DNp03/DNp04 are effectively unlabelled in snapshot 783 (one cell), so the classical escape co-command neurons could not be tested as types. Nothing here has a body: the descending neurons end in a ventral nerve cord this connectome does not contain, so every "command" is a spike rate with no leg attached — the MANC dataset is where that continues. And every number in this volume is a rate under Poisson noise, not a response to a growing image; the looming stimulus of ch55 is a recruitment count standing in for an expanding edge.

Sixty-one chapters ago this was a 57 MB download and a question about whether a Mac mini could hold a fly's brain. It can. What it holds, it turns out, is not a simulation of a fly — it is a testable claim generator: 29 experiments, six honest negatives, one self-caught retraction, one confirmed novel prediction (ch26), and now a reflex disassembled into six numbers that a wet lab could go and check on Monday. The connectome was published as a map. Three volumes in, the useful thing about it is that you can break it on purpose.

Volume III closes here. The open doors are labelled above.