1Summary
The Neural Continuity Research Program (NCRP) depends on one capability above all others: the ability to record the structure of nervous tissue at a resolution fine enough to identify every neuron, every neurite and every synapse within a volume. This report states what that capability would have to deliver, phase by phase, and where present methods fall short. It is a requirements document. It does not describe a reconstruction of any human brain, and it does not propose one.
Principal findings
- Synaptic-resolution mapping requires sampling on the order of 10 nanometers or finer in at least two axes and no coarser than about 30–40 nanometers in the third. At the reference voxel adopted here (about 500 nm³), one cubic millimeter of tissue produces about 2 × 10¹⁵ voxels, or roughly 2 PB of raw 8-bit image data.
- Raw data volumes scale with tissue volume: about 0.9 EB for a mouse brain, about 16 EB for a marmoset brain, and about 2.4 ZB for a human brain of 1.2 million mm³. The mouse figure alone would occupy more than half of the 1.6 EB tape archive at Facility Alpha; the marmoset and human figures exceed it by factors of about 10 and 1,500 respectively.
- For Phase II, a whole mouse brain acquired over 36 months requires a sustained net acquisition rate of about 10¹⁰ voxels per second, roughly ten times the net rate of the fastest single instruments in current use.
- Automated segmentation accuracy, not imaging speed or compute, is the binding constraint. Manual proofreading at present effort levels is infeasible above mouse scale; a human-scale effort would require error rates several orders of magnitude lower than those now achieved.
- A structural connectome supports at most a neuron-level model (EFL-3). Synapse-level and molecular-state models (EFL-4 and EFL-5) require molecular annotation for which no method yet operates at scale.
- Twelve requirements for the acquisition-and-processing pipeline at Facility Alpha are stated in this report. Most are met or can be met within Phase I; three depend on archival capacity, network capacity or annotation staffing not yet in place.
2Introduction and scope
The NCRP was chartered on 9 June 2026 to investigate whether human cognition could eventually be represented, reconstructed and sustained on computational systems. Its four Program Phases proceed from foundations (Phase I, 2026–2028) through mammalian-scale reconstruction (Phase II, 2028–2031) to human-scale feasibility (Phase III, 2031–2034). Each phase assumes that structural data of sufficient quality can be acquired, processed and preserved at the scale the phase requires. This report tests that assumption against the physics of imaging, the arithmetic of data volume, and the present state of automated reconstruction.
The term connectome is used here in its narrow sense: a map of neurons and the synaptic connections among them, with enough geometric detail to locate each connection on the cells it joins. The first complete connectome of any animal, the nervous system of the nematode Caenorhabditis elegans, was reconstructed by hand from serial electron micrographs and published in 1986. Four decades later, automated methods have produced a wiring diagram for the whole brain of the adult fruit fly and petavoxel-scale reconstructions of fragments of mouse and human cortex. These are substantial achievements. They are also between six and eight orders of magnitude smaller, by tissue volume, than a human brain.
The report proceeds from resolution requirements to imaging modalities, data volumes, throughput by phase, segmentation and proofreading, and molecular annotation. It then considers the relationship between structural data and functional models, and closes with requirements for the pipeline at Continuity Research Facility Alpha. Compute and energy requirements for running models derived from these data are treated separately in DSC-TR-2026-008.1
What this report does not address
Current synaptic-resolution methods require chemically fixed tissue that is physically sectioned or ablated during imaging. They cannot be applied to a living brain. Within the NCRP, human tissue is obtained only post-mortem, through donation programs reviewed by the Research Ethics Review Board (RERB).2 Questions of identity and consent raised by such work are addressed in DSC-WP-2026-006 and are outside the scope of this report.
3Resolution requirements
The required resolution is set by the smallest structures that must be traced continuously or identified unambiguously. Three dominate. The thinnest unmyelinated axons in mammalian cortex are on the order of 50–100 nanometers in diameter, and must be followed across thousands of consecutive images without a break. Dendritic spine necks can be thinner still. And the synapse itself must be recognized from the apposition of a presynaptic bouton, containing vesicles roughly 40 nanometers across, with a postsynaptic density across a cleft of about 20 nanometers.
| Structure | Typical dimension | Consequence for sampling |
|---|---|---|
| Plasma membrane (lipid bilayer) | 5–10 nm thick | Must be resolved as a continuous boundary; sets lateral sampling at ≤ 10 nm for reliable automated tracing |
| Synaptic cleft | ~20 nm | Detectable at 4–8 nm lateral sampling; synapse identification degrades rapidly above ~10 nm |
| Synaptic vesicle | ~40 nm diameter | Vesicle clouds identify presynaptic boutons; individual vesicles resolved only at ≤ 5 nm |
| Thin unmyelinated axon | ~50–100 nm diameter | Axial sampling must not exceed roughly half the diameter of the thinnest process traced (≤ 30–40 nm) |
| Dendritic spine neck | ~40–200 nm diameter | Spine-to-parent assignment fails if neck is lost between sections |
| Gap junction plaque | 2–4 nm intercellular gap | Not reliably resolved at connectomic throughput sampling; requires higher resolution or molecular labels |
Two sampling regimes satisfy these constraints in practice. Anisotropic acquisition, typical of serial-section methods, samples at about 4 × 4 nanometers in the section plane and 30–40 nanometers between sections. Isotropic acquisition, typical of focused-ion-beam methods, samples at about 8 × 8 × 8 nanometers. The two regimes produce similar voxel volumes, near 500 nm³. This report adopts 500 nm³ as its reference voxel and 1 byte per voxel as the reference encoding.3 Finer sampling improves segmentation accuracy but increases data volume in proportion; coarser sampling reduces volume but raises error rates nonlinearly, because a single missed thin process propagates into merge or split errors across an entire cell.
4Imaging modalities
Four families of methods are relevant. They differ in resolution, speed, whether they destroy the specimen, and what molecular information they can carry. No single method meets every requirement; a production pipeline is likely to combine them.
4.1Serial-section electron microscopy
In serial-section electron microscopy, a fixed, heavy-metal-stained and resin-embedded block is cut into ultrathin sections, typically 30–40 nanometers thick, which are collected on tape, wafers or grids and imaged individually by transmission or scanning electron microscopes. Because sections are imaged after cutting, imaging can be parallelized across many instruments, and sections can be re-imaged if a defect is found. Multi-beam scanning instruments, which raster tens of electron beams simultaneously, have raised single-instrument throughput to the order of 10⁹ pixels per second at the beam, with lower net rates once stage movement, focusing and overheads are included.
The principal weaknesses are mechanical. Sections can fold, tear, compress or be lost; each loss removes a 30–40 nanometer slab from the record, and thin processes crossing that slab may become untraceable. Aligning tens of thousands of independently imaged sections into a coherent volume is itself a substantial computational task. Section loss rates in large campaigns are typically reported in the range of a fraction of one percent to a few percent; for whole-brain work, the tolerable rate is lower than current practice.
4.2Block-face electron microscopy
Block-face methods image the exposed surface of the block, remove a thin layer, and image again. Removal is by an in-chamber diamond knife (serial block-face scanning electron microscopy) or by a focused ion beam, which can remove layers of 4–8 nanometers and so permits isotropic sampling. Because the block is never handled as loose sections, alignment is simpler and section loss does not arise. The cost is serialization: each layer must be imaged before the next is removed, and the volume that a single focused-ion-beam instrument can process is small, typically tens of thousands of cubic micrometers per instrument-month. Approaches that divide a large block into thicker slabs, image each slab by block-face methods in parallel, and stitch the results are under development and are the most likely route to isotropic data at cubic-millimeter scale and above.
4.3Expansion microscopy
Expansion microscopy embeds tissue in a swellable hydrogel, anchors biomolecules or labels to the gel, digests or denatures the native tissue, and expands the gel isotropically in water. Single-round protocols expand tissue roughly fourfold in each dimension; iterative and high-expansion protocols reach factors of ten to twenty. The expanded specimen is imaged with light microscopy, which is fast and, crucially, compatible with multiplexed fluorescent labeling of specific proteins. A fourfold expansion imaged with a diffraction-limited system of about 250 nanometers resolves features near 60 nanometers in the original tissue; twentyfold expansion approaches the 10–20 nanometer range.
For connectomics, the open question is whether expanded tissue can carry a membrane or protein-density label sufficient for reliable automated tracing of the thinnest processes, and whether expansion is uniform enough across large volumes. Distortion of a few percent is tolerable for registration but not for tracing if it is spatially irregular. The attraction of the method is that it may allow structural and molecular information to be acquired in a single pass, which electron microscopy cannot readily do.
4.4X-ray methods
X-ray tomography, particularly at synchrotron sources, can image intact stained tissue non-destructively. Micro-computed tomography at micrometer resolution is already used to map whole small-animal brains for cell-body positions and for registering electron-microscopy volumes to anatomical context. Nanotomography and ptychographic methods have reached resolutions in the tens of nanometers on small specimens, but not at the speed, volume or contrast required to trace thin axons and identify synapses across cubic millimeters. Radiation dose limits and source access are further constraints. In the near term, X-ray methods are best regarded as a quality-control and registration layer: a non-destructive record of the block taken before destructive electron imaging begins, against which section loss and distortion can be measured.
4.5Comparison
| Modality | Typical sampling (x × y × z) | Net throughput class | Destructive | Molecular information | Assessed NCRP role |
|---|---|---|---|---|---|
| Serial-section TEM/SEM (incl. multi-beam) | 4 × 4 × 30–40 nm | 10⁸–10⁹ voxels/s per instrument | Yes (sectioned) | Limited (morphology; post-section labeling) | Primary for Phases I–II |
| Serial block-face SEM | 5–10 × 5–10 × 25–50 nm | 10⁶–10⁷ voxels/s | Yes (ablated) | Limited | Targeted volumes; method development |
| Focused-ion-beam SEM | 4–8 nm isotropic | 10⁵–10⁶ voxels/s | Yes (ablated) | Limited | Ground truth; synapse-level detail |
| Expansion microscopy (light) | ~10–60 nm effective | 10⁸–10⁹ voxels/s (expanded space) | Yes (tissue digested) | High (multiplexed proteins) | Molecular annotation; candidate primary for Phase III |
| Synchrotron X-ray tomography | 0.3–1 µm (whole organ); tens of nm (small samples) | Variable; source-limited | No (at moderate dose) | Minimal | Registration and quality control |
5Data volumes
At the reference voxel of 500 nm³, one cubic millimeter (10¹⁸ nm³) contains 2 × 10¹⁵ voxels. At 1 byte per voxel this is 2 × 10¹⁵ bytes, or 2 PB. The published petavoxel reconstruction of a cubic-millimeter-class fragment of human temporal cortex is consistent with this figure. All other volumes in Table 3 follow by multiplication. The estimates exclude tile overlap (typically 5–10 percent), re-imaging of defective regions, and derived products, all of which add to the totals.
| Specimen | Tissue volume (mm³) | Voxels | Raw data | Neurons (approx.) | Synapses (approx.) |
|---|---|---|---|---|---|
| C. elegans (whole animal) | ~3 × 10⁻³ | ~6 × 10¹² | ~6 TB | 302 | ~10⁴ |
| Drosophila (adult brain) | ~2.5 × 10⁻² | ~5 × 10¹³ | ~50 TB | ~1.4 × 10⁵ | ~5 × 10⁷ |
| Cortical sample | 1 | 2 × 10¹⁵ | ~2 PB | ~5 × 10⁴–10⁵ | ~10⁹ |
| Mouse (whole brain) | ~450 | ~9 × 10¹⁷ | ~0.9 EB | ~7 × 10⁷ | ~10¹¹ |
| Marmoset (whole brain) | ~8,000 | ~1.6 × 10¹⁹ | ~16 EB | ~6 × 10⁸ | ~10¹²–10¹³ |
| Human (whole brain) | ~1.2 × 10⁶ | ~2.4 × 10²¹ | ~2.4 ZB | ~8.6 × 10¹⁰ | ~10¹⁴–10¹⁵ |
The human figure deserves a sense of proportion. Raw imaging of a single human brain at the reference sampling would generate roughly 2.4 ZB — about 1,500 times the 1.6 EB tape archive at Facility Alpha and some 6,000 times its 410 PB parallel storage tier.4 The volume is not a reason to abandon the question, but it does mean that a human-scale campaign cannot be designed on the assumption that raw data are retained in full on conventional media.
Derived products
Most of the scientific value of an acquisition is carried by derived products that are far smaller than the raw images. A segmentation volume assigning a cell identifier to each voxel compresses well, because labels are spatially coherent. A synapse table — the list of connections with their locations, sizes and partner identifiers — is smaller again. The final connectivity graph for a human brain, at 10¹⁴–10¹⁵ synapses and some 50 bytes per synapse record, would occupy 5–50 PB: large, but within the capacity of existing facilities.
| Product | Content | Indicative size | Ratio to raw (approx.) |
|---|---|---|---|
| Raw image volume | Grayscale voxels, 8-bit | ~2.4 ZB | 1 |
| Losslessly compressed raw | As above, entropy-coded | ~0.8–1.2 ZB | ~1:2–1:3 |
| Segmentation (label) volume | Cell identifier per voxel, compressed | ~10–100 EB | ~1:25–1:250 |
| Synapse table | Partners, location, size, predicted type | ~5–50 PB | ~1:5 × 10⁴–1:5 × 10⁵ |
| Skeletonized morphology | Neurite centerlines and radii | ~1–10 PB | ~1:2 × 10⁵ or smaller |
The ratio between raw and derived data, roughly five orders of magnitude, frames a central policy question that this report cannot settle: whether, at scale, raw images may be discarded once derived products have been validated. Discarding raw data forecloses re-segmentation with better algorithms. Retaining it may be physically impractical. The forthcoming Neural Data Preservation Framework is expected to address retention classes for raw acquisition data.5
6Throughput requirements by phase
Table 5 converts each Program Phase’s reference acquisition target into a required sustained net rate, defined as total voxels divided by the elapsed acquisition window, inclusive of maintenance, calibration and downtime. Instrument counts assume a net rate of 10⁹ voxels per second per instrument, which is at the upper end of current multi-beam performance and should be read as an optimistic planning value.
| Phase | Reference acquisition target | Raw data | Acquisition window | Required net rate | Instruments at 10⁹ voxels/s net |
|---|---|---|---|---|---|
| I (2026–2028) | Several 1 mm³ cortical volumes (mouse, primate); replicate C. elegans and Drosophila datasets | ~10 PB | 24 months | ~1.6 × 10⁸ voxels/s (~0.16 GB/s) | 1–2 |
| II (2028–2031) | Whole mouse brain (~450 mm³); primate cortical volumes (~10–50 mm³ total) | ~1 EB | 36 months | ~1.1 × 10¹⁰ voxels/s (~11 GB/s) | ~11 |
| III (2031–2034) | Human tissue at ~1,000 mm³ (1 cm³) scale; demonstration of rate scaling | ~2 EB | 24 months | ~3.2 × 10¹⁰ voxels/s (~32 GB/s) | ~32 |
| Reference only: human whole brain (not a Program objective) | ~1.2 × 10⁶ mm³ | ~2.4 ZB | 10 years | ~7.6 × 10¹² voxels/s (~7.6 TB/s) | ~7,600 |
Phase I is within reach of a single high-throughput instrument and is not throughput-limited. Phase II requires a tenfold scale-up in parallel instruments or a corresponding improvement in single-instrument rate, together with sectioning, staining and handling processes that remain reliable across a block the size of a whole mouse brain. Phase III, as currently scoped, is a feasibility demonstration at roughly twice the Phase II data volume in two-thirds of the time. The final row is included only to fix the scale of the problem. A ten-year whole-human-brain acquisition would require several thousand instruments of the best current class operating continuously, or an improvement in per-instrument net rate of about three orders of magnitude.
Specimen preparation is an independent constraint. Uniform fixation and staining of a block several millimeters across is difficult; for a human brain it is not currently possible without subdividing the tissue, which introduces cut surfaces that must be traced across. Any Phase III design must treat preparation as a first-order engineering problem, not a preliminary step.
7Segmentation and proofreading
Raw images become a connectome through a sequence of computational steps: alignment of images into a volume; detection of cell boundaries; agglomeration of voxels into supervoxels and supervoxels into cells; detection of synapses and assignment of their partners; and classification of cells by type. Each step is now performed predominantly by machine-learning models trained on human-annotated ground truth. Each produces errors, of which two classes matter most. A merge error joins two cells into one. A split error breaks one cell into two or more fragments. Both corrupt the connectivity graph, and a single merge between two large cells can misattribute thousands of synapses.
Proofreading is the correction of these errors, historically by trained human annotators. The whole-brain Drosophila reconstruction of 2024, comprising roughly 140,000 neurons, relied on a large distributed proofreading community working over several years. That model of work does not scale to a mammalian brain. A simple calculation illustrates the point. A mouse brain contains about 7 × 10⁷ neurons. At one person-hour of proofreading per neuron, the effort is 7 × 10⁷ hours, or about 35,000 person-years at 2,000 working hours per year. For a human brain of 8.6 × 10¹⁰ neurons, even one second of human attention per neuron amounts to about 12,000 person-years.6
The requirement that follows is that automated error rates must fall until human review is needed only for flagged regions and statistical audit. Table 6 sets phase targets on that basis.
| Metric | Phase I target | Phase II target | Phase III target |
|---|---|---|---|
| Expected axonal path length between automated errors (merge or split) | ≥ 1 mm | ≥ 1 cm | ≥ 10 cm |
| Synapse detection precision / recall (vs. held-out ground truth) | ≥ 0.90 / 0.90 | ≥ 0.95 / 0.95 | ≥ 0.97 / 0.97 |
| Synaptic partner assignment accuracy | ≥ 90% | ≥ 95% | ≥ 98% |
| Median human proofreading effort per neuron | ≤ 10 min | ≤ 30 s (flagged regions only) | Sample-based audit only |
| Proofreading effort, whole reference target | ~1–2 × 10⁴ person-hours per mm³ | ~300 person-years total | Not to exceed Phase II total |
| Neurons with complete, verified arbor within volume | ≥ 90% | ≥ 80% (whole brain) | Estimated by audit; ≥ 95% confidence bounds reported |
Segmentation compute is not expected to be the binding constraint. Current automated pipelines require on the order of 10⁵–10⁷ floating-point operations per voxel, depending on architecture and the number of passes. For the roughly 10¹⁸ voxels of Phase II, the total is 10²³–10²⁵ operations. At a sustained 10¹⁸ low-precision operations per second — roughly what a 10 percent share of ALPHA-1, rated at about 3.8 × 10¹⁹ operations per second peak, would deliver at 25 percent utilization — this corresponds to between about one day and four months of wall-clock time. Data movement, ground-truth generation and quality control are the larger costs.
8Molecular annotation
Electron micrographs record membranes, organelles and vesicles. They do not, by themselves, record which neurotransmitter a synapse releases, which receptors it expresses, how ion channels are distributed along a dendrite, or how neuromodulatory systems act on a circuit. Some of this can be inferred: machine-learning classifiers have predicted neurotransmitter identity from ultrastructure in insect tissue with useful accuracy. Much of it cannot. Any model that aims beyond a neuron-level wiring diagram needs molecular information co-registered with structure.
| Annotation | Relevance | Candidate methods | Compatibility with EM connectomics | Lowest EFL requiring it |
|---|---|---|---|---|
| Neurotransmitter identity | Sign and class of each synapse | Ultrastructure-based prediction; immunolabeling of adjacent sections; expansion microscopy | Partial | EFL-3 |
| Cell type (transcriptomic or morphological) | Assigns dynamical parameters by class | Morphological classification; spatial transcriptomics on matched tissue | Partial | EFL-2 |
| Receptor subtype and density | Synaptic kinetics and strength | Multiplexed immunolabeling; expansion microscopy | Limited | EFL-4 |
| Ion channel distribution | Intrinsic excitability and dendritic integration | Multiplexed labeling; inference from cell type | Poor | EFL-4 |
| Electrical synapses | Fast coupling between cells | High-resolution EM; connexin labeling | Difficult at throughput sampling | EFL-3 |
| Neuromodulator receptor maps | State-dependent circuit dynamics | Multiplexed labeling | Limited | EFL-4 |
| Synaptic protein inventories and phosphorylation state | Plasticity state, short-term memory traces (hypothesized) | No scalable method | None | EFL-5 |
Two strategies are under consideration for Phase I method development. The first pairs electron microscopy with sparse molecular labeling on interleaved sections, accepting gaps in coverage. The second treats expansion microscopy with multiplexed labels as a candidate primary modality, contingent on demonstrating tracing accuracy equivalent to electron microscopy on the same tissue. A head-to-head comparison on matched 0.1 mm³ volumes is planned for the NCRP benchmark suite.7
9From structural connectome to functional model
A connectome is a static record of a fixed specimen. It is not a model, and it does not by itself predict activity. The C. elegans connectome has been complete since 1986; nearly four decades later, models built from it reproduce some but not all of the animal’s behavior, because the dynamics of each neuron, the strengths and signs of connections, neuromodulation and extrasynaptic signaling are not fully determined by the wiring. This is the most important caveat in this report, and the reason the NCRP uses a graded fidelity scale rather than a single notion of reconstruction.8
| EFL | Model class | Structural data required | Additional data required | Status (August 2026) |
|---|---|---|---|---|
| EFL-0 | Statistical or behavioral model | None | Behavioral and recording data | Established methods |
| EFL-1 | Regional network model | Mesoscale connectivity (tracer studies, diffusion imaging) | Regional activity data | Established methods |
| EFL-2 | Cell-type population model | Cell-type census and projection statistics | Cell-type electrophysiology | Partial; region-specific |
| EFL-3 | Neuron-level connectome model | Complete neuron-level connectome with synapse counts and signs | Cell-type dynamical parameters | Demonstrated for C. elegans and partial Drosophila circuits |
| EFL-4 | Synapse-level model with plasticity | Synapse geometry and size; dendritic morphology | Receptor composition; plasticity rules | Not demonstrated |
| EFL-5 | Molecular-state-informed model | As EFL-4 | Molecular inventories and states at synapse and compartment level | Not demonstrated; no acquisition method at scale |
| EFL-6 | Validated individual-equivalent | Undefined | Undefined, including validation criteria | Hypothetical; no known path |
Three consequences follow for acquisition. First, electron-microscopy connectomics, even if perfected, supports models up to EFL-3. Second, progress beyond EFL-3 depends on molecular annotation and on functional data from the same or matched specimens; the 2025 MICrONS study, which co-registered calcium imaging of activity with electron-microscopy reconstruction in mouse visual cortex, indicates the form such datasets will take. Third, validation requires functional ground truth. A model derived from a fixed specimen can be tested only against activity recorded before fixation, which for human tissue is available only in exceptional clinical circumstances and never at whole-brain scale. The absence of a validation pathway is the principal reason EFL-6 is described in Program documents as hypothetical.
10Requirements for the Facility Alpha pipeline
Continuity Research Facility Alpha reached initial operating capability on 18 August 2026. Its primary system, ALPHA-1, provides the compute allocation for NCRP reconstruction, and its storage tiers — a 410 PB parallel file system and a 1.6 EB tape archive — receive data from imaging instruments at Alpha and at participating research institutions. The requirements below govern the acquisition-and-processing pipeline from instrument output to validated derived products. They are stated for Phase I, with Phase II scaling indicated where it differs.
| Ref. | Requirement | Phase I | Phase II | Status |
|---|---|---|---|---|
| AP-01 | Sustained ingest capacity from instruments to Alpha parallel storage | ≥ 2 GB/s | ≥ 15 GB/s (≥ 1.4× required acquisition rate) | Phase I met; Phase II requires network upgrade |
| AP-02 | Fixity: cryptographic checksum computed at instrument and verified on ingest before instrument-side deletion | Mandatory | Mandatory | Met |
| AP-03 | Section-to-section alignment residual | ≤ 10 nm median; ≤ 30 nm at 99th percentile | As Phase I | Under validation |
| AP-04 | Raw data written to tape archive | ≤ 30 days after acquisition | ≤ 14 days | Met for Phase I |
| AP-05 | Segmentation throughput relative to acquisition | ≥ 1.5× acquisition rate | ≥ 1.5× acquisition rate | Met for Phase I |
| AP-06 | Every segmentation and proofreading edit versioned; any derived product reproducible from raw data, software version and parameters | Mandatory | Mandatory | Partially met |
| AP-07 | Densely annotated, held-out validation volumes | ≥ 0.1% of each acquired volume | ≥ 0.01% of each acquired volume, stratified by region | In progress |
| AP-08 | Concurrent proofreading sessions supported, with interactive latency ≤ 200 ms | ≥ 50 | ≥ 200 | Phase I met |
| AP-09 | Human-derived data processed in a segregated enclave under RERB-approved protocols | Mandatory | Mandatory | Met |
| AP-10 | Provenance record linking each product to specimen, preparation protocol, instrument and software | Mandatory | Mandatory | Met |
| AP-11 | Connectivity graphs exported in open, documented formats with versioned schemas | Mandatory | Mandatory | Schema in draft |
| AP-12 | Archival capacity sufficient to retain raw data for all reference targets of the current phase plus one preservation copy | ~20 PB | ~2 EB | Phase II exceeds current 1.6 EB archive |
Three requirements cannot be met with current infrastructure for Phase II. AP-01 requires an ingest network upgrade between partner instrument sites and Alpha. AP-12 requires approximately 2 EB of archival capacity for raw Phase II data and one copy, against a 1.6 EB archive that will also hold other Departmental workloads; this gap must be closed by archive expansion, by off-site preservation copies at Facility Bravo, or by an approved retention policy that reduces the raw data held. AP-07 depends on annotation staffing, which is a workforce rather than an infrastructure question. The storage-media options relevant to AP-12 are assessed in DSC-TR-2026-005.9
11Limitations and open problems
- Specimen preparation for blocks larger than a few cubic millimeters is not solved. Whole-mouse-brain staining protocols exist but have not been demonstrated at the uniformity required by Table 6 targets.
- Error-rate targets in Table 6 for Phases II and III exceed published automated performance by one to two orders of magnitude. They are requirements, not forecasts.
- The reference voxel of 500 nm³ may be too coarse for reliable identification of electrical synapses and the thinnest spine necks. If finer sampling proves necessary, all data volumes scale accordingly.
- Molecular annotation at scale lacks a demonstrated method compatible with electron-microscopy connectomics.
- Post-mortem interval and fixation quality in human donor tissue introduce artifacts whose effect on reconstruction accuracy is not well characterized.
- No method yet exists for validating a model derived from a fixed human specimen against the functioning brain from which it came. This limits every claim of fidelity above EFL-3.
Each of these is a research problem rather than an engineering scale-up, and the Directorate regards the Program Phase schedule as conditional on progress against them. Phase gates will assess acquisition and segmentation performance against the targets in this report before resources for the following phase are committed.
Appendix A. Derivation of reference figures
Voxel count per cubic millimeter: 1 mm³ = (10⁶ nm)³ = 10¹⁸ nm³; 10¹⁸ ÷ 500 nm³ = 2 × 10¹⁵ voxels. Raw bytes at 8 bits per voxel: 2 × 10¹⁵ bytes = 2 PB. Human whole brain: 1.2 × 10⁶ mm³ × 2 PB = 2.4 × 10⁶ PB = 2.4 ZB. Required rate for a ten-year human acquisition: 2.4 × 10²¹ bytes ÷ (10 × 3.15 × 10⁷ s) ≈ 7.6 × 10¹² bytes per second.
- Net acquisition rate
- Voxels delivered to storage per second averaged over the full acquisition window, including all downtime.
- Merge error
- An automated segmentation error in which voxels belonging to two or more distinct cells are assigned to a single object.
- Split error
- An automated segmentation error in which voxels belonging to one cell are assigned to two or more objects.
- Ground truth
- Densely annotated tissue volumes, traced and verified by multiple independent annotators, used to train and evaluate automated methods.
- Reference target
- The acquisition volume against which a Program Phase’s throughput requirement is computed. A reference target is a sizing device and does not by itself constitute an approved acquisition.
Notes
- 1.DSC-TR-2026-008, Energy and Compute Requirements for Large-Scale Neural Emulation (30 July 2026), estimates compute, memory and power for models at each Emulation Fidelity Level. The present report addresses only the data required to construct such models. ↩
- 2.All NCRP protocols involving human tissue are reviewed by the RERB before approval. No living human subject is or can be imaged by the methods described in this report. ↩
- 3.Reference voxel: 4 × 4 × 30 nm (480 nm³) and 8 × 8 × 8 nm (512 nm³) are both rounded to 500 nm³. Eight-bit grayscale is standard for EM connectomics; some modalities use 16-bit, which doubles all raw volumes. ↩
- 4.Storage figures for Facility Alpha are design capacities at initial operating capability. A share of both tiers is allocated to the Integrated Systems Modeling Program and other Departmental workloads. ↩
- 5.The Neural Data Preservation Framework is in preparation and is expected to be released for public comment in September 2026 as DSC-TR-2026-011. ↩
- 6.Per-neuron effort figures are illustrative and chosen to indicate scale. Actual effort varies by more than an order of magnitude between cell types; long-range projection neurons dominate. ↩
- 7.The NCRP benchmark suite, hosted on ALPHA-1, comprises standard datasets and evaluation tasks against which acquisition, reconstruction and modeling methods are compared across Program Phases. ↩
- 8.The Emulation Fidelity Levels are a Program classification describing the level of biological detail represented in a model. Assignment of a model to a level describes what it represents, not whether it is accurate; validation is assessed separately. ↩
- 9.DSC-TR-2026-005, Archival Media Longevity Assessment: Storage Technologies for Century-Scale Retention (26 May 2026). ↩
Background literature (real-world works)
The following published works are real. They are cited for general scientific and historical context only. They do not support the fictional findings of this document, and their authors are not associated with this project.
- White, J. G., Southgate, E., Thomson, J. N., and Brenner, S. (1986). The structure of the nervous system of the nematode Caenorhabditis elegans. Philosophical Transactions of the Royal Society of London B, 314(1165).
- Dorkenwald, S., et al. (2024). Neuronal wiring diagram of an adult brain. Nature, 634.
- Shapson-Coe, A., et al. (2024). A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution. Science, 384(6696).
- MICrONS Consortium (2025). Functional connectomics spanning multiple areas of mouse visual cortex. Nature, 640.
- Sandberg, A., and Bostrom, N. (2008). Whole Brain Emulation: A Roadmap. Technical Report 2008-3, Future of Humanity Institute, University of Oxford.