Reverse engineering the brain.

Because almost none of that cost was the imaging. The MICrONS program paid for a first-of-its-kind electron microscope pipeline, two-photon functional imaging of the same tissue, and the years of human proofreading that are still going on. Wellcome's $7.5 to 21.7 billion for a whole brain is explicit that almost all of it is wages for people correcting the computer's tracing.

So the target is really two numbers. The instrument: a light microscope on expanded tissue costs hundreds of thousands of dollars, a multibeam electron microscope millions, and LICONN has shown the light image is good enough to trace. The labor: the fraction of corrections that a model makes instead of a person. Our target of under $100 million requires that fraction to be close to all of them. It is a goal, stated as one; the scaling law on the Why now page is the evidence that it moves with data, and the first two years of the plan are spent measuring how far.

Two clarifications a careful reader asks for. First, $100 million is the production cost of a brain once the factory runs, not the company's budget; the factory, the instrument development and the first brain's mistakes are capital, and the plan raises for them separately. Second, the 100x is measured against Wellcome's estimate of what making one costs by the published method today, since nobody sells one; electron microscopy will also get cheaper as proofreading automates, and the honest comparison in 2033 is against that. Light still wins on three things that do not change with software: the price of the instrument, the thousand-fold fewer sections to align, and the proteins imaged directly in the same gel. Wellcome's other lines shrink with it: its $200 to 300 million for imaging was twenty multibeam electron microscopes at about $5 million each for five years, and its $300 to 400 million for registration and segmentation was priced in 2023 against thirty-nanometer sections. Light instruments cost a tenth as much, cloud compute is cheaper every year, and slabs a thousand times thicker mean a thousand times fewer alignments. Those lines are in our budget at their measured values, not Wellcome's. And PATHFINDER's gain on proofreading is inside our hundred-fold, not competing with it: the target counts every published advance we can use, and the company's own work is the part that remains after them. Take its projection at face value: if its 84-fold held for a whole brain, Wellcome's proofreading line falls to roughly $0.1 to 0.3 billion, and what is left is the $0.5 to 0.7 billion of electron-microscope imaging, registration and segmentation, which is exactly where light's instrument price and thousand-fold fewer sections take over. The hundred-fold is against the estimated cost of making one today, because no brain has been made; against an electron-microscope factory of 2033 the edge is the instrument, the sections, and the proteins in the gel, and we would rather say that than pretend the other factory stands still.

They cannot in native tissue; two neurons sit 20 nanometers apart and light resolves about 250. Expansion microscopy swells the tissue in a hydrogel, and at sixteen-fold the same gap is 320 nanometers. The gap is the easy part of the criterion. What decides whether a volume can be traced is separating the thinnest, densest axons, about 100 to 200 nanometers across, which become 1.6 to 3.2 micrometers in the gel, and whether the synapses can be found, which LICONN does from their marker proteins rather than by resolving the cleft. LICONN, published in Nature in 2025, did exactly that: stained proteins broadly so the image shows membranes and organelles the way an electron micrograph does, reached an effective voxel of about 10 by 10 by 25 nanometers in tissue units (about 20 nanometers of resolution in the plane), and traced the result automatically at an accuracy its authors call comparable to electron microscopy's.

What LICONN has not done is a whole brain. Its volumes were up to a thousandth of a cubic millimeter on a spinning-disk confocal. The whole-brain light-sheet microscopes that exist, the Allen Institute's ExA-SPIM, image a three-fold expanded brain in a day at about 500 by 1,000 nanometers, or 375 by 750 at four-fold, which shows projections but not synapses. The microscope we have to build is the one in between: a light sheet that holds LICONN's resolution over a sixteen-fold expanded brain. That is an engineering program, not a discovery, and it is the first capital we spend.

The arithmetic is ours and it is on the page: LICONN imaged 0.95 million cubic micrometers of tissue, measured before expansion, in 6.5 hours, so 500 cubic millimeters of brain, which is 2 liters of expanded gel, is about 400 instrument-years on that microscope. Even at that rate it is a factory problem, not a physics problem: a hundred instruments for four years, at a price per instrument far below one multibeam electron microscope.

A light sheet illuminates and images a whole plane at once, where a spinning disk scans the plane with hundreds of points at a time, which is why ExA-SPIM does a whole expanded brain in a day. Its published rate is 946 million voxels a second, about 55 times LICONN's confocal, measured at its own coarse resolution; if a light sheet held that rate at LICONN's voxel size, the brain would be about seven instrument-years per channel, our arithmetic, and ten instruments would image it in under a year. It will not hold it unchanged: finer resolution means thinner sheets and fewer photons per voxel, so the real rate at sixteen-fold will be lower, and how much lower is the first-year measurement that sets the instrument count. If it can, the imaging budget is small; if it cannot, the confocal factory still fits. Which one we build is decided by measurement in the first year, not by hope.

Yes, and the page says so. LICONN expanded 50 micrometer slices and showed 300 micrometers with adaptations, and it demonstrated a partial answer of its own: image an expanded slab, remove the imaged layer, image deeper with overlap, and fuse the volumes, twelve rounds over about 205 micrometers of native tissue, so the continuity problem inside a slab is solved by overlap. What remains is the cut made before expansion; a brain is tens to hundreds of slabs, every cut face carries on the order of a hundred million axon crossings, and every one has to be matched to itself on the other side. Electron microscopy has the same problem at every one of its 30 nanometer sections and solves it with alignment software; ours are a thousand times thicker, so there are a thousand times fewer boundaries, each one harder, since Wellcome's report notes that thick-section interfaces lose more tissue at the cut and are harder to align than ultrathin ones, which is why the plan measures the error per crossing rather than assuming it. The long projections, which whole-brain light-sheet imaging already follows across a brain, constrain the match for the axons that cross the most faces.

There is a trade inside it that we state plainly: a thick slab means fewer crossings to match but needs an objective with a long working distance at high numerical aperture, and a thin slab the reverse. The slab thickness, the objective and the measured error per crossing are first-year measurements, and the barcode question (whether E11 Bio's self-labeling cells are needed to make matching reliable) is the first experiment we run. A barcode does not have to be unique among 71 million neurons, only among the few axons that meet at the same spot on a cut face, which is why a library of a hundred thousand can serve a whole brain.

Not that we are ahead of Google; the benchmark authors' own model beat the human figures first, and we say so on the page. It proves three things we needed to know before starting a company on them. A small model, 7 million parameters, fine-tuned on the public corrections, reaches the published state of the art on both proofreading tasks. Its error falls as a power law in training data with no sign of leveling off, so more corrections, which a factory generates every day, buy predictable gains. And the model trained on electron microscopy transfers to expanded light-microscopy tissue, so the decades of public electron-microscopy proofreading are training data for the new method.

What it does not prove is that a whole brain can be finished without people. The benchmark scores decisions that a proofreader already raised; finding the errors in a fresh volume is the harder half, and the number that matters is human hours per cubic millimeter, end to end, against the best automated pipeline, with the wrong joins counted per length of axon. Our own end-to-end pipeline, run once under a pre-registered protocol on nine held-out cubes of the public mouse cortex volume, none seen in training, closed about a fifth of the breaks and added a few wrong joins. We publish that number because the company's job is to move it, and because a reader who finds it later would trust nothing else on the page.

Nor does the plan need zero people. It needs the hours per cubic millimeter to fall a few hundred-fold from Wellcome's figure (the arithmetic is in the question on what $100 million has to buy), and the scaling exponents on the page, 0.20 and 0.16, say what each doubling of corrections buys. A factory generates corrections every day it runs; that is the data the law feeds on.

The nine-cube test is public in full: the pre-registration file with its hashes, the execution record, the per-neuron outputs and the scoring code are one download on our research page, so anyone can rescore it.

Sources: braininsights.ai/research, the pre-registered evaluation of October 2026.

It does not, by itself, and the page does not say it does. The map ranks hypotheses: for a receptor, it says which neurons carry it, where those neurons project, and which circuits can therefore be reached by a drug that binds it. For the GLP-1 receptor that is the difference between the neurons that reach the circuits for satiety and the ones that reach the circuits for nausea, which today's two best papers disagree about.

Conviction comes from the second half of the product. The customer doses animals and sends us the tissue; we image the treated and untreated circuits against the reference map and report the difference, synapse by synapse. Hypothesis from the map, test from the tissue. The genome worked the same way: the sequence did not say what a gene does, but every experiment afterward was designed against it.

The two GLP-1 papers that disagree did not fail for lack of a map; each used silencing and rescue experiments that are stronger than anatomy. What they lacked was a common reference: each group defined its neurons by its own markers and its own atlas registration, so the results cannot be laid on top of each other. A map at this resolution is the reference both would register to, which is the first thing it is for. Nor does a receptor map predict a pharmacological paradox on its own: a GIP receptor agonist and a GIP receptor antagonist both cause weight loss, and no anatomy would have called that. What the map frames is the dosed-tissue readout that shows whether the two reach the same circuits, which is the experiment that resolves it.

Correct on both counts, and it is listed on the page as among our first measurements. The tools are reporter mouse lines, in which the cells that express a receptor also express a fluorescent protein that is easy to image, and validated antibodies where they exist. The two answer different questions: a reporter marks which cells express the receptor, an antibody marks where the protein sits on them, and a study states which of the two it used. LICONN demonstrated a catalog of protein markers in its gels, cell-type and synaptic proteins imaged a few at a time alongside the structural stain, so the chemistry for adding a receptor label is established; what has to be shown is each receptor, one at a time. Expansion dilutes every label by the cube of the factor, about four thousand-fold at sixteen, and the same dilution applies to the background; LICONN's synaptic proteins are low-abundance and were imaged anyway. Whether a given receptor clears the bar is a measurement, which is why we do not promise any receptor on any schedule, only that each one is reported as measured.

Where neither works for a receptor, the fallback is cell type: spatial transcriptomics on adjacent tissue registers which cell types express the receptor, and the map says how those cell types are wired. That is coarser, and we say so when we sell it.

A related objection is that where a drug reaches matters as much as where its receptor sits; a peptide like semaglutide enters the brain mostly at a few sites with a leaky barrier. That is why the product measures dosed tissue rather than inferring from the map: the readout shows which circuits actually changed, distribution included, and the map says what those circuits are.

Those tools answer where a receptor is expressed, which regions project to which, and what a population does when it is activated. None of them says which individual neurons that carry a receptor synapse onto which others, which is the level at which a drug's effects and its side effects separate. The Allen projectome is at the resolution of regions; MERFISH gives cell types without their synapses; photometry and chemogenetics read or drive a population you have already chosen. The map is what you choose it from.

The product is not a replacement for any of them. It is the reference that the others are registered to, and a comparison service, treated against untreated, that none of them can run at synaptic resolution.

True, and it is the right objection to a map sold as a simulator. We do not sell one. The worm's 302 neurons taught the field that the same wiring is reconfigured by neuromodulators into different working circuits, which is exactly why our product pairs the map with a measurement: the customer's dosed animals, imaged against the reference, so what is reported is the change the drug produced, not an inference from anatomy alone.

The map earns its place by making that measurement interpretable and cumulative. A difference seen in a circuit is attached to identified cell types and their synaptic partners, and every study adds to the same reference, so the second customer's study is read against the first's. Without the map each experiment is a one-off; with it, the field's experiments add up.

The worm is also the wrong size for the argument. Its 302 neurons were mapped before anyone could run experiments on the map as a model, and the field's 2026 result on artificial networks, described under Reading the map, is that understanding a complete network compounds: the bigger it is, the more every explanation is constrained by the ones around it, and automated scientists can run the experiments that constrain it. In the whole-fly-brain model built on the FlyWire connectome, the wiring does predict the published activation and silencing experiments, at 0.83 AUROC on all 259 and 0.67 on the 90 its authors did not use to build it, and it stops predicting them when blurred to regions. A map with a model on it is dynamics enough to test, and the tissue says how well.

That was the settled view of artificial neural networks until 2026, and it rested on evidence that turned out to be misread. The neuron that fired for cats and cars detected a V shape; the layer where no neuron could be read was the product of an architecture that learned badly; the mixing that remained was related things grouped because they were related. With better tools, an architecture that learns well and agents to run the experiments, about 90 percent of the neurons in a 32-billion-parameter language model have one job, and the author's conclusion is that a tangled-looking neuron is usually a statement about the observer, not the neuron. The section Reading the map lays this out with its numbers and its source.

The brain's version of the cat-and-car neuron is mixed selectivity: neurons in prefrontal cortex that respond to combinations of task variables, recorded through a few electrodes. The essay's author, who is not a neuroscientist and says so, writes that if the cat-and-car neuron were in a human brain seen through a few electrodes or a blood-oxygen signal he would discount the polysemantic reading almost to zero, because he could not see what that neuron was when it was in front of him for years with every tool he wanted. We do not know whether mixed selectivity dissolves the same way. A complete map with a model on it, where every neuron can be driven and silenced, is the instrument that can tell, and a few electrodes is not. Our own first measurement on the mouse, in October 2026: on the MICrONS cubic millimetre, with one snapshot of activity and no experiments to run, an agent reading each neuron's wiring predicted its tuning no better than its location did. Whether cortical wiring is unreadable or only unrunnable is what a model on the map would settle.

What a map does not supply is the experiments. Everything in that result depends on removing a neuron and rerunning the input, which a fixed brain cannot do. The experiments run on the model built from the map; the dosed tissue checks the model. That is the loop the product is built around, and it is why we are testing the method on the fly model first, where the loop already closes in software.

Sources: Cammarata, Neural Networks Are Already Clean, Goodfire, 2026.

Because the circuit cannot be checked on its own. An explanation of what a population does is tested against what its inputs do, what its outputs do, and what happens when its neighbors are removed, and in an unmapped brain none of those is known. The field reading neural networks found the same thing: understanding one row of an otherwise empty crossword is possible, but without the crossings there is no way to know it is right, and the whole network turned out to be easier to understand than the parts alone. Each circuit explained constrains the next; a whole brain gives every crossing.

The measured version is in the fly. Blur the FlyWire wiring to the resolution a regional atlas gives, run the published whole-brain model, and it stops predicting the published experiments: 0.51 AUROC, chance, against 0.83 with the real wiring on 259 experiments. Activity still spreads; it no longer reaches the right neurons. A regional study for a customer is the first product because it fits a slab and a confocal that exists today, but every study is registered to the whole-brain reference, and the whole brain is what makes the second study cheaper than the first.

We do not build on it. The brain gets mapped either way, and the first product, a drug developer's dosed tissue compared against the reference, does not depend on any claim about neural networks. What the essay supplies is a reason to think the reading step is tractable and a method for doing it: neuron papers, attribution graphs with the network's own neurons as nodes, explanations that a second agent can test blind, all run by agents at a scale no laboratory of people could match. We cite it as what it is, an informal essay by the researcher who co-wrote the papers it revises, with his own confidence stated at about 85 percent and the technical companion still to come, and we quote the counterexample he names himself.

The disagreement inside the field is about whether to read a network's own neurons or train a second model to decompose it first. For a brain that question is moot: the neurons are the only units there are, and nobody is proposing to decompose a brain into something else before reading it. Where the essay matters is the claim that reading the original units works, and gets easier as the map fills in. We are testing that claim on the fly model under a pre-registered protocol, and the result goes on this page either way.

Sources: Cammarata, Neural Networks Are Already Clean, Goodfire, 2026.

Because the neurons still matter even where the signal is diffuse. A receptor on a cell changes that cell's output, and the cell's output goes down its axon to its synaptic partners; volume transmission changes who is listening, not where the listeners' wires go. The map says which neurons carry the receptor and where their output lands, which is the question a developer asks when a drug's main effect and its side effect come from different circuits.

The expansion chemistry also images proteins directly, so the receptor's position on the cell, on the soma, the dendrite, or at a synapse, is part of the readout. That is information no transcriptomic atlas has, because messenger RNA does not say where the protein went.

It is the baseline we have to beat, and we say so to customers. A c-Fos screen counts the cells that responded, at cellular resolution, and reports them by region registered to the Allen atlas. What it cannot say is what those cells connect to, which cell types they are at synaptic resolution, or whether the cells that responded are the ones that carry the receptor. A regional study answers those three questions on the same tissue, and the first pilots are designed to show, on a compound the customer already has c-Fos data for, what the extra resolution changes.

If it changes nothing for a given question, the customer should keep buying the screen. The business is built on the questions where it does.

They do, mostly in Phase 2 and 3, and no mouse map changes a human efficacy readout. What it changes is the choice of what to put into those trials: which receptor, in which cell type, and which liabilities the circuit predicts. Those choices are made largely in rodents today, by every company, with far less information than a map provides.

The receptors most brain drugs bind are conserved between mouse and human, which is why mice are the pharmacology model in the first place; the circuits are conserved less perfectly, and a study says which of its findings rest on which. A human map at this resolution is the long-term vision, and a human cortex sample has already been reconstructed this way with electron microscopes.

They are precedents for one thing: that pharma pays upfront for a map of biology it cannot make itself and pays again on milestones. Each bought something different, Recursion a phenomic screening platform, Cerevance human brain cell-type expression data, and neither is what we sell. They set the shape of a platform deal, not its price. Our first revenue is regional studies, sold one at a time at a fee a pharmacology budget already carries, and platform deals come only after studies have changed decisions.

At LICONN's voxel size, 10 by 10 by 25 nanometers, a mouse brain is about 200 petavoxels per channel: 200 petabytes at one byte per voxel, a couple of exabytes with the structural channel, the synapse markers and the receptor channel all stored at sixteen bits, before compression. MICrONS was about 1.6 petabytes for one cubic millimeter, the figure usually cited for it, so this is more than a hundred times the largest map ever handled, and segmenting it is a compute line of its own, of the order of tens of millions of dollars at cloud prices. It is in the budget, and it is small next to the proofreading line it replaces, which Wellcome puts at $7 to 21 billion. The software to run it at that scale, the flood-filling networks, the alignment and the storage format, is open source from Google's group, and running it on a cloud rather than Google's own machines is a cost, not a research problem.

Our arithmetic, from the figures on this page, each a target until measured. Imaging: at LICONN's confocal rate the brain is about 400 instrument-years, so a hundred instruments for four years at a few hundred thousand dollars each, $20 to 50 million with operation; a light sheet that holds the resolution brings that to a handful of instruments and a fraction of the sum. Storage and compute: about 200 petavoxels per channel to hold and segment, tens of millions of dollars at cloud prices. Tissue, animals and labels: small next to those. That leaves of the order of $20 to 50 million for human correction.

At Wellcome's $20,000 per person-year that is 1,000 to 2,500 person-years of proofreading for the whole brain, against the 0.3 to 1 million in Wellcome's estimate: a reduction of a few hundred-fold in human hours per cubic millimeter, not a hundred and not zero. Today's measured starting points are PATHFINDER's automated repair on electron microscopy, our pipeline's fifth of the breaks on held-out tissue, and the benchmark scaling law. The measurement that the first two years exist to make, and that the 2028 report states as a number, is human hours per cubic millimeter on our own expanded tissue, with the wrong joins counted per length of axon. If it does not extrapolate to those 1,000 to 2,500 person-years, the whole brain does not cost $100 million, and we say so.

A dense reconstruction of every neuron in the brain with its synapses detected and the cell types and the labeled receptors attached, at a published error rate per length of axon, with the regions customers have paid to study proofread to the standard used for analysis today. A whole brain proofread by hand to that standard everywhere is what costs billions, and nobody has it; a whole brain at a known error rate, with the models improving it every month, is what we deliver, and the error rate is part of the product, not a footnote to it.

The methods are open, by their authors' choice: LICONN and PATHFINDER are both published CC BY, and we will use them, credit them, and publish our own. E11 Bio's PRISM reagents are shared for non-commercial use, and LICONN's authors disclose a patent application, so the plan carries licenses for both where they are needed; if barcodes prove necessary and a license is not available, we build our own. That is a line in the plan, not a surprise. The moat is not a method. It is the factory, the brains it accumulates, which make every later map cheaper and every comparison more valuable, the corrections it logs, which train models nobody else has the data for, and the customers whose dosed tissue only we have imaged. That is the moat sequencing companies had: the chemistry was published, the instrument and the installed base were the business.

E11 Bio is a nonprofit that publishes its methods toward a mouse brain on philanthropic funding, and we would rather collaborate than compete; our transfer result was measured on their public data. What a nonprofit does not build is a factory with customers, and what no research group has is a reason to run the same brain a thousand times. That is the company.

Because a regional study does not image a cubic millimeter per animal. It images the circuit the question is about, a region of a fraction of a cubic millimeter chosen in advance, in a slab, with the receptor and an activity marker labeled, and compares the treated animals against the untreated ones and against the reference. The volume, the number of animals per arm, powered and blinded rather than one animal a side, and the sampling plan are part of the study design and are stated in the quote, with the error rates for the region.

The $200,000 figure is the whole-brain production cost we are building toward for 2033; it is not the 2027 study price, which is set by what the study costs us then and what the decision is worth to the customer. The first studies are priced as a research budget line, with the fee rising only once a study has changed a program decision.

The report gives the lines in its text: proofreading $7 to 21 billion, registration and segmentation $300 to 400 million, imaging $200 to 300 million, total $7.5 to 21.7 billion, all for a single mouse brain with current technology and excluding the cost of developing anything new. Those dollar figures are what this page cites. The person-year count behind the proofreading line sits in the report's cost box (its Figure 5), which derives it from the fly connectome's measured rate at $20,000 a person-year, giving 0.3 to 1 million person-years; the report's narrative elsewhere rounds the labor to about 100,000 person-years. The two are reconciled by the wage: $7 to 21 billion is 0.3 to 1 million person-years at $20,000, or 100,000 person-years at $70,000 to $210,000, which is what a fully loaded technician costs in the United States. Our own check against the published hemibrain rate puts a dense mouse brain nearer 2 million person-years at the fly's standard of completeness, so the figure is, if anything, low.

That is also why the comparison to today's cost is not inflated by labor: labor is the cost, under either count. The company exists to replace that line.

They did, and the same decade produced the reason to want one: nausea that limits the dose, discontinuation rates that are the main thing the next generation competes on, and a pipeline of dual and triple agonists whose receptor combinations are being chosen by trial, two years and hundreds of millions of dollars at a time. The first drug in a class is found without a map; the second, third and tenth are the ones a map makes cheaper to choose, and that is where the money in the class now is.

The claim is not that drugs cannot be made without a map. Genomes were not needed for aspirin. The claim is that once the map exists and is cheap, every program is designed against it, because the ones that are not will lose to the ones that are.

The seed round is $10 million and it buys one measurement in public: the cost of correcting the tracing on expanded tissue at our resolution, on a pilot imaging line, with the result published whether it passes or not. The whole plan needs $300 to 500 million over about five years, raised in stages, each released by the one before it, with the largest rounds released only when the measured cost says a whole brain is affordable. Nobody is asked for the whole sum on day one.

What the money builds along the way is a business that does not wait for the brain: the regional studies sold from 2027, the receptor labels validated one by one, the imaging line and the models, and the reference volumes that accumulate under them. If the whole brain is late, that company exists; if the measured cost never extrapolates to under $100 million, we say so, and the company is a smaller one than we set out to build. We would rather state that than pretend the plan has no gate.

Not neuron by neuron. No two mice share a neuron, so the units of comparison are the cell type and the circuit: how many synapses of each type a labeled population makes onto each partner population, and how that changes with the drug. That is what the reference carries, and it is what a study reports. The between-animal variation is measured inside every study from the untreated animals, which is the control the comparison is made against, and the fact that it has never been measured at this resolution is one of the things a customer is paying to learn.

Each study's animals also join the reference. The one brain of 2033 is the first of a series, and the comparisons get sharper as the series grows, because the variation is then known rather than assumed.

Yes, and it is the same mechanism a second time, on the two precedents on this page. The first brain carries the instrument development and the mistakes; the second runs on instruments that exist, models trained on the first brain's corrections, and a staff who have done it once. Sequencing fell a million-fold not in one step but in a series of them, each paid for by the customers of the step before. The $1 million brain is not promised for a date. It is the point at which whole-brain comparisons become a product, and the page says what becomes affordable at each price along the way.

There is never a two-liter gel. Each slab is expanded on its own, as LICONN's were, so the object handled is a slab one to five millimeters thick and ten to sixteen centimeters across at full width, or smaller pieces at the price of more boundaries, and the question is uniformity between slabs rather than within one enormous one. LICONN measured its expansion factor at 15.4 with a standard deviation of 1.7 across replicates from three mice, and distortion within a slab is measured against a coarse image taken before expansion and corrected in registration, which is how expansion microscopy has been quantified since it was introduced.

Across hundreds of slabs the chemistry has to be a production process with a quality score per slab: expansion factor, stain penetration, label signal, and the fraction of the slab that traces cleanly. A slab that fails its score is a hole in the map that cannot be re-cut, so the process is validated on many slabs from many animals before the first production brain is sectioned. Nobody has run expansion at that scale; it is a factory discipline, and it is the wet-lab work of the first two years.

On its own it is not, and the page does not claim the curve alone finishes a brain. It says what more data buys, and the point of a factory is that data of this kind is its by-product: every region a customer pays to have proofread produces corrections, about a million on MICrONS's single cubic millimeter, and a brain is five hundred of those. The curve is the floor, not the plan.

The plan is the levers that move the curve rather than ride it: structured repair of the kind PATHFINDER showed, which took the expected error-free run length of an axon from 82 to 94 percent without a larger model; the constraints that a whole brain adds and a cube lacks, since a long axon has to go somewhere and a slab boundary has to match; protein barcodes if they prove to help; and the fact that the plan needs the human hours per cubic millimeter to fall a few hundred-fold, not to zero. Which of those levers does what is measured in public as we go.

They built the pieces, and none of them is funded to assemble a whole brain. Google's connectomics group is a research team that publishes methods and co-authored LICONN. The Allen Institute's whole-brain work is funded through NIH's BRAIN CONNECTS program for about 10 cubic millimeters by 2028, as is Harvard's, and the Harvard grant was terminated and restored within one year. E11 Bio is a nonprofit that publishes its methods and is working toward a mouse brain on philanthropic funding. Meridial, a nonprofit research organisation set up in London in 2026 with funding from ARIA, plans synapse-resolution mapping of circuits for CNS drug development, the closest statement of our first product by another group; it has published no results yet. We would rather the field were crowded than empty: every group that gets the method working makes the map cheaper, and the company is the one organised to sell what the map says.

A whole brain at this resolution needs a factory: dozens of instruments, an animal facility, a reconstruction pipeline run as production, and a customer who pays for the output. That is a company, and this is it. We will use every published method, credit it, and publish our own measurements, including the ones that fail.

The plan is staged against measurements, each one reported whether it is good or bad. In the first two years: the imaging rate and resolution of the light sheet we build, the error rate across slab boundaries, the first receptor imaged in expanded tissue, and the fraction of corrections our models make on tissue they have never seen. By the end of 2028 we will know whether the measured cost of correcting the tracing extrapolates to a whole brain for under $100 million, and we will say so either way.

Revenue starts before the whole brain: from 2027, regional studies of one circuit for one compound, which need a slab, not a brain, and a spinning-disk confocal of the kind LICONN used, which exists today, not the light sheet, which is for the whole brain. The whole brain is the product that makes every study after it cheaper.

It is the long-term vision and it is labeled as one: the page says the odds are low and the journey long. The company does not depend on it. The first two markets, pharma's decisions and the drugs designed against a map, pay for the factory on their own, and both exist today. We state the third because it is why the founder is doing this, and because a map good enough to predict what a drug does is the first step toward a model of the brain that runs on its own, whether or not that step is ever taken.

A pharma customer sometimes reads it as a sign that their business is a bridge to something else. It is the opposite: the pharmacology product is the only way the maps get made, and every map a customer pays for is a better reference for the next customer. Pharma is the business; the vision is what the business makes possible.