A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution

Science 384(6696):eadk4858. Published 10 May 2024. DOI: 10.1126/science.adk4858. PMID 38723085; PMC11718559.

• 31 authors. Shapson-Coe, Januszewski and Berger contributed equally. Corresponding authors: Jain (Google Research) and Lichtman (Harvard).
• Peer-reviewed primary resource paper introducing the dataset "H01".
• Bibliographic record verified against Crossref, Europe PMC/PubMed and the manuscript's own front matter. No correction or erratum is registered with Crossref (verified, 29 September 2026).
• Read 29 September 2026 by a subagent (claude-opus-5-5, reasoning effort max), from the NIH author manuscript, not the version of record (see below). The main session spot-checked six claims against the source text and tables: patient age, cell total, synapse total, miss rates, "engramics" and the 0.092% wording. All match; the cell total is 57,216 in the tables and 57,180 in the text, as the summary notes.

Version read: the NIH author manuscript (accepted manuscript), not the Science version of record. PMC labels it "Author manuscript; available in PMC 2025 January 10 … Published in final edited form as: Science 2024 May 10." Any copy-editing or figure corrections made only in the version of record are unknown to this reading.

What was read

• Main text. The author-manuscript PDF (32 pages) was read in full:

  • structured abstract, one-sentence summary and graphical abstract;
  • introduction, Methods, Results and Discussion;
  • acknowledgements, funding, data and code availability, author contributions and references 1–92;
  • all seven figure captions.

  The text extraction (1,883 lines) was read line by line. All 32 pages were rendered and inspected. The graphical abstract and Figs. 1–7 were inspected at their embedded 300-dpi resolution, with zooms on Figs. 4G and 5F.

• Other renderings of the main text. PMC's HTML, JATS XML and plain text are the same manuscript. A word-level comparison with the PDF found differences only in front matter, reference placement and the supplementary-material note (verified). They were not read separately.

• Supplementary Materials (author-manuscript DOCX; its own contents list is "Materials and Methods, Figs. S1 to S11, Tables S1 to S13"). All 1,473 text lines were read. All 16 figure images (Figs. S1–S11) were inspected at native resolution. A second, LibreOffice rendering (58 pages) confirmed that no text was missed.

• Tables S1–S13. The DOCX prints only S4–S7 and S10–S12. It says the rest are "only presented via the link" to the authors' Google Sheets, so all 13 linked sheets were exported.

  • Read in full: S2–S7 and S9–S13.
  • Read for structure and checked programmatically, not cell by cell: S1 (5,019 section rows) and S8 (15,289 neuron rows).
  • The sheets are live documents. Where the DOCX also prints a table (S4–S7, S10–S12), the values agree (checked by eye). The S7 sheet adds an F-test row.

• Our checks. The numeric checks cited below as "verified" are scripted in papers/consciousness_connectomics/shapsoncoe2024_petavoxel_checks.py. Every assert passes.

Not obtained or not read.

• Version of record. The science.org article, supplement PDF, supplementary-tables ZIP, Movie S1 ZIP and MDAR reproducibility checklist were not obtained. Every science.org request returned a Cloudflare challenge, which was not circumvented. Internet Archive captures of those URLs are redirects only.
• PMC's own supplement files (tables ZIP, Table S13 XLSX, video). These sit behind a proof-of-work download check, which was not circumvented. The supplement DOCX and the manuscript PDF were instead obtained from Internet Archive captures (16 January 2025) of PMC's files.
• Movie S1. Not obtained and not watched. No movie legend exists in the obtained materials; the main text says "see video" for the two triangular-cell groups.
• bioRxiv preprint. Not fetched: a different version (2021, four versions, different title). Only its metadata was retrieved.
• Not audited: the H01 data release, CAVE, CREST, VAST, the code (GitHub, Zenodo, MATLAB archive), Neuroglancer states, and the extra linked sheets (synapse ground-truth lists and inspected subsets). No reanalysis was run.
• Ethics and consent. No ethics, IRB or consent statement appears anywhere in the manuscript or supplement text read (verified by search). It may appear in the version-of-record or MDAR material.

Reading status: complete for the accepted-manuscript text, supplement text and all 13 tables. Incomplete relative to the version of record: its text, the MDAR checklist and Movie S1 are unread. Sources, hashes and every access attempt are in papers/consciousness_connectomics/shapsoncoe2024_petavoxel.provenance.json.

Question and contribution

Can a millimetre-scale piece of human cortex, spanning every layer, be imaged at synaptic resolution, reconstructed and shared? What does a first pass through it show?

The contribution is a resource: H01, a ~1 mm³, 1.4 PB serial-section EM volume running from layer 1 to white matter. It comes with:

• an automated cell and neurite segmentation;
• about 150 million detected synapses with excitatory/inhibitory (E/I) labels;
• a manual census of every cell body;
• community proofreading and exploration tools.

It also offers two "vignettes": orientation classes of layer 5/6 triangular cells, and rare connections of many synapses between one axon and one target.

It is not a complete wiring diagram of any neuron population. The slab is 170 µm thick, so almost every neuron, and nearly every long-range axon, is truncated. The automated reconstruction is largely unproofread, and nothing functional was recorded. The authors present it as a starting point: "this petascale dataset is a start".

Specimen and its clinical origin

• Patient. A 45-year-old woman with drug-resistant epilepsy, whose epileptic focus in the left medial temporal lobe was resected. The structured abstract calls it an underlying "hippocampal lesion".
• Sample. "Excess" left anterior temporal cortex (anterior middle temporal gyrus) removed to reach the focus; the resected piece was about 2.5 × 0.8 cm. Its orientation and myelin pattern fit Hopf's "tmag" area. Neuropathology found no significant changes in the adjacent anterior temporal lobe by light microscopy. It also lacked the outer layer-2 band seen when hippocampal sclerosis extends into temporal cortex.
• Fixation. Immersion in cold aldehyde "immediately after excision"; the delay in minutes is not given. The tissue was not perfused, so blood is present (46 circulating white cells).
• Not reported: epilepsy duration, medication history, seizure burden, cognition, handedness, in-vivo imaging of this tissue, or any behavioural or functional data from the person.
• Authors' own caveats.
  • Long-term epilepsy or its treatment may have had subtle effects.
  • Some oddities have no known cause: very large spines, varicosities filled with unusual material, axon "whorls", and "dark" pyramidal-shaped cells in layers 4–6.
  • Only comparisons across patients with different disorders could show whether this sample is normal.

Reconstruction pipeline

1. Preparation. Vibratome slices 300 µm thick were stained with reduced osmium–thiocarbohydrazide–osmium and embedded in resin. The second slice was used, and its block was trimmed to 4,584 × 1,975 µm.
2. Sectioning.
   • 5,019 ATUM sections were cut, 30–80 nm thick (92.6% at 30–33 nm; mean 33.9 nm; 0.170 mm in total).
   • Cutting became unstable after section 1,639, so the knife was replaced after section 1,695.
   • The replacement knife was not parallel to the block face, giving a first run of partial ("re-entry") sections. Renewed instability led to rotating the block 180°, giving a second re-entry run of 62 sections.
   • 48 sections broke into several pieces.
   • Double-thickness sections are duplicated in the digital volume, giving about 5,293 layers.
3. Imaging.
   • A 61-beam Zeiss multibeam SEM imaged at 4 × 4 nm. Correcting for 28% cutting compression makes the pixels 5.55 × 4 nm. Shrinkage between living and embedded tissue is not corrected.
   • Dwell time was 200, 400 or 800 ns (4,439, 516 and 64 sections; verified from Table S1).
   • Raw data were 1.83 PB, of which 1.43 PB are tissue. The text reports 326 days of imaging. Table S1's per-section durations sum to 133 days (verified); the difference is not explained.
   • Per-section quality control (Table S1, verified): 57 sections have missing tiles and 1,557 sections have jitter. The microscope configuration changed mid-series ("new system", 2,693 sections).
4. Alignment. Elastic-mesh stitching and alignment used a proprietary Google keypoint detector, followed by optical-flow realignment (SOFIMA). 383 sections were marked invalid for the flow optimisation (287 single, 96 in runs of 2–9).
5. Segmentation.
   • Flood-filling networks ran at three resolutions (32, 16 and 8 nm), trained on 375, 1,285 and 6,142 million voxels of manual ground truth.
   • Vessels, five tissue "fissures" and badly distorted or missing-section regions were masked. The fissures are regions where neurites cannot be told apart; two run through the whole stack, attributed to damage during resection.
   • 36.1% of tissue voxels lie outside the segmentation, either masked or extracellular.
6. Agglomeration and error correction.
   • An 11-stage agglomeration produced 1.7 B graph edges.
   • 6.0 M distinct merge cuts were applied. 2.9 M came from a skeleton subcompartment classifier (overall 0.995 correct, but axon initial segment (AIS) precision/recall only 0.77/0.71). The other 3.7 M, partly redundant, came from glia/neurite type separation. Enforcing soma separation removed a further 0.6 M edges.
   • Two releases: c2 (fewer breaks, more merges) and c3 (more conservative). c3 derives from the earlier "20200916" graph, and all analyses use it.
7. Synapses.
   • A 3D U-Net was trained on 3,217 hand-labelled synapses from six regions, one per layer.
   • 1.9 B candidate sites were filtered, paired within 800 nm and re-oriented by skeleton voting (34.1 M flipped).
   • Duplicates were merged: pairs closer than 750 nm always, and pairs at 750–1,050 nm by a logistic model (AUC 0.875). This left 166.2 M; restricting to synapses with an axonal presynaptic side left 149,871,669.
8. E/I labels.
   • A ResNet-50 classifier was trained on 5,312 synapses, labelled by the type of the neuron they come from.
   • Its output was then overridden in two cases. A synapse was set to excitatory if its postsynaptic segment had a detected spine within 1 µm. A synapse from a known pyramidal cell or interneuron was set to that cell's type.
   • Result: 111,272,315 E and 38,599,354 I synapses.
9. Cells.
   • Every soma was painted by hand every 128th section and classified manually (49,080 non-vascular cells). Vascular nuclei were painted separately (8,136 cells in Table S5).
   • Microglia vs OPCs were separated by a linear classifier on SegCLR embeddings, trained on 36 microglia and 20 OPCs.
10. Layers. A density clustering (HDBSCAN) on soma position and size gave three clusters, whose edges were fitted with six concentric circular arcs. These define L1–L6 and white matter. They are algorithmic; no independent validation is reported.
11. Tools. Neuroglancer (viewing), CAVE (community proofreading of c3), VAST (painting and export) and CREST (path finding and proofreading).

Completeness and accuracy: distinct quantities

The paper reports detection recall and precision, label accuracy and proofreading burden. It does not report neuron-level accuracy after proofreading (compare Dorkenwald's 99.2% volumetric F1) or edge-level recall of the neuron-to-neuron graph. Low false discovery does not imply completeness. Inhibition is the class most under-detected.

Results, with numbers

Cells and tissue

• Cell census (Table S5, verified sums).
  • 16,087 neurons:
    • 10,531 excitatory/spiny: 8,803 pyramidal, 1,535 other spiny and 193 spiny stellate cells;
    • 4,688 interneurons;
    • 868 unclassified.
  • 32,315 glia: 20,139 oligodendrocytes, 5,474 astrocytes and 6,702 microglia/OPCs.
  • 678 unclassified cells and 8,136 vascular cells, for 57,216 in total.
  • Glia outnumber neurons 2:1, and oligodendrocytes are the commonest cell. Among classified neurons, E:I is 69:31.
• Neuron density. About 16,000/mm³: 15,741 using 1.022 mm³, or 15,321 using 1.05. The authors put this roughly one third below light-microscopy estimates for human temporal cortex and nearly 10× below mouse association cortex.
• Glia by layer.
  • Astrocytes tile layers 2–6. In superficial layer 1 they are denser, smaller and intermingled (12,210/mm³).
  • Microglia/OPC density is even across layers (~6,000/mm³), and they sit close to vessels.
  • Oligodendrocytes rise toward white matter (54,034/mm³ there) and line radial vessels.
• Neuropil composition (share of segmented volume, excluding the 36% of voxels outside the segmentation): unmyelinated axons 40.2%, dendrites 25.8%, glia 15.5%, somata 9.4%, myelinated axons 7.5%, collapsed vessels 1.5%, AIS 0.07%, cilia 0.03%.
• Myelin. Mostly perpendicular to the section in white matter, tangential in layer 1, with radial bundles. The authors note these orientations affect diffusion-MRI signals.
• Vessels.
  • About 230 mm of vessels, lined by 4,604 endothelial cells (~20 per mm) and 3,549 pericytes and other perivascular cells (~15 per mm).
  • 74 "bloodless bridges" connect capillaries.
  • Other perivascular cells: 574 smooth-muscle cells, 396 perivascular macrophages, 128 perivascular lymphocytes and 78 fibroblast-like pericytes.
• Odd objects. Twenty kinds of unusual or unidentified objects (Fig. S2); 34 binucleate cells, mostly astrocytes; one neuron with two axons (Fig. S3).

Synapses and E/I balance

• Density and targets. 149.9 M synapses were detected (0.147/µm³). Of those with an identified target, 99.4% are on dendrites, 0.394% on somata and 0.197% on the AIS.
• By layer. Excitatory density is highest in layers 1 and 3, and inhibitory density peaks in layer 1. The E fraction is slightly lower in layer 1.
• Pyramidal vs interneuron input. In each of layers 2–6, pyramidal neurons receive a larger excitatory share of inputs than interneurons. Mean E/(E+I) is 0.66–0.75 against 0.55–0.63, with Welch p < 10⁻⁸ in every layer (Table S7). Whole cells show this pattern too: pyramidal somata, AIS and proximal dendrites are mostly inhibitory-innervated, while spiny dendrites are mostly excitatory (Fig. 5D–E). See the inconsistencies and our limitations below: the direction survives, but the table and legend have errors, and the spine rule shapes the comparison.

Layer 5/6 triangular ("compass") cells

• Population. 876 bipolar spiny cells deeper than 1,750 µm, about one third of spiny neurons in layers 5–6.

• Basal-dendrite angle. Mean ~126° from the local radial axis.

• Direction around the radial axis. The distribution is bimodal:

  • 347 point "forward", toward section 5292;
  • 339 point "reverse", toward section 0;
  • 186 point sideways, and 4 are excluded.

  The two groups are mirror-symmetric in angle and branching. Their axis matches the main axis of the underlying white-matter axons, which the authors infer is the brain's anterior–posterior axis.

• Clustering. In the middle half of the stack, 245 of 431 cells have a nearest neighbour pointing the same way (116/97/89/129 by group). Fisher's exact p = 0.00517 (verified); after shuffling labels, p = 0.773. The function is unknown.

Axons making many synapses on one target

• Connection sizes (per input connection, Table S8): 96.49% have one synapse, 2.99% two and 0.35% three (verified).
  • Connections of four or more synapses are 0.164% (verified). The text's 0.092% is the share with exactly four.
  • 39% of the ~2,743 neurons with at least 3,000 dendritic synapses have at least one input of ≥7 synapses. Table S8 gives 39.7% of 2,729 using all input synapses.
  • All of this is on unproofread c3, which the authors expect to underestimate connection strength because of split errors.
• Two proofread layer-3 pyramidal axons (Table S9, verified).
  • "Pyr1": 397 partners and 706 synapses. 97.5% of partners get 1–4 synapses, but five inhibitory partners get 10–54: 53 synapses at clustered sites onto one interneuron (Fig. 7E) and 54 onto another, which the text does not mention.
  • "Pyr2": 251 partners; its strongest are 30 synapses onto an inhibitory cell, 13 onto an excitatory cell and 9 onto an inhibitory cell.
• Morphology. Many strong connections are formed by terminal branches that grow onto the target on both sides of an axon–dendrite crossing. The authors call this suggestive of "intentionality".
• Null model.
  • Each sampled axon is displaced 15 µm in XY and its synapses are re-placed around its shaft, using fitted en-passant and terminal-bouton distance distributions.
  • 28,639 of 52,674 sampled axons could be simulated. Ten simulated axons with ≥6-synapse artefacts were excluded.
  • Discrete Kolmogorov–Smirnov tests against 79.8 M observed axon fragments give p < 10⁻¹⁰: D = 0.047 overall, 0.045 for excitatory and 0.051 for inhibitory axons.
  • Observed counts exceed the null from 5 synapses upward: 6,792 vs 2,688 at 5, 679 vs 28 at 8, and 432 vs 0 at ≥11 (Table S13, verified).
• Which axons are strong. Of the 12,674 axon fragments whose strongest connection has ≥5 synapses, 10,839 (86%) are inhibitory (our count from Table S13). The dataset-wide tail is mainly an inhibitory phenomenon, whereas the illustrated examples are mostly excitatory-to-inhibitory.

Internal inconsistencies we verified

Ordered by consequence. Each is checked against the paper's own tables or figures; the scripted ones are in the checks file.

1. Fig. 5F and Table S7 (E/I input to pyramidal cells vs interneurons).
   • The legend's colour key is reversed. The legend says interneurons are blue and pyramidal cells orange, but the blue violins sit at Table S7's pyramidal means (0.705, 0.749, 0.686, 0.663, 0.709; read from the figure).
   • Table S7's N rows are exchanged. Its printed Welch t values reproduce only with the pyramidal and interneuron Ns swapped (4 of 5 layers; layer 5 is ambiguous). With Ns as printed, the interneuron total (8,306) exceeds all 4,688 interneurons in the census. The swapped Ns match Table S6's density ratios. The sheet's F-test row, by contrast, reproduces with the printed (mismatched) Ns.
   • Table S7's layer-1 entry (N 133, mean 0.572, SD 0.146) repeats the layer-6 entries exactly. It disagrees with the layer-1 violin, which reads about 0.45 by eye.
   • Inferred consequence: the text's claim (pyramidal > interneuron, layers 2–6) survives, but the layer-1 value and the printed sample sizes are unreliable.
2. The null-model KS statistics test the wrong end of the distribution. Every D is reached at strongest-connection strength 1: observed cumulative 0.98428 vs null 0.93730 (D = 0.047; verified). The tests therefore detect a shortfall of 2–3-synapse strongest connections (1.39% observed vs 5.87% expected), not the claimed excess of strong ones. No test of the tail is reported. At 4 synapses the observed share is below the null (0.0249% vs 0.0280%), contrary to the Fig. 7F legend ("all connection strengths greater than 3").
3. "Four or more synapses (0.092%)" is actually exactly four synapses. Four or more is 0.164% (verified from Table S8). The printed shares otherwise sum to 99.92%.
4. Imaged volume. The main text says 1.05 mm³. The supplement says 1.022 and prints the unit "mm²".
5. Triangular cells. The text's "~77.5%" forward/reverse does not match 686/876 = 78.3%. Fig. S8 carries panels D–F that its legend does not describe: an analysis of 864 cells using angles around the apical, not radial, direction. It is only a guess that 77.5% comes from that earlier run (670/864 = 77.5%).
6. Cell totals.
   • The text gives 8,100 vascular cells and 57,180 cells in total. Table S5 and Fig. 4B give 8,136 and 57,216. Endothelial plus pericyte counts alone make 8,153.
   • The supplement's "2,339 pericytes" does not reconcile with 3,549.
   • The text's "49,080 neurons and glia" includes 678 unclassified cells.
   • The 457 microglia/OPC cell bodies never passed to the classifier are not mentioned.
7. Table S6. The layer-3 spiny-neuron density equals the layer-3 oligodendrocyte density exactly (8,927.3/mm³). This is a suspected copy error; Fig. S5E plots the same values, so it cannot be checked independently.
8. Captions and cross-references.
   • Fig. 3 cites a "red rectangle in G", but there is no panel G.
   • Fig. 6C says "apical dendrites pointing forward/reverse"; it is the basal dendrites.
   • The supplement cites "fig. 7I", but Fig. 7 has no panel I.
   • The supplement says the Welch tests covered layers 1–5, while the text and Table S7 say layers 2–6.
   • Fig. 7E labels nine synapse groups where the text says "eight sites".
9. Bookkeeping. None of these changes a conclusion.
   • The postsynaptic exclusions leave 133,711,099 synapses, not 133,704,881.
   • The sectioning narrative sums to 5,046 sections (or 5,016), not 5,019.
   • 276 duplicated double-thickness sections imply 5,295 layers rather than 5,293.
   • "On average 270 nm and 1,500 nm" from the axon shaft does not match the fitted distributions (means ≈402 and ≈2,256 nm; modes ≈245 and ≈1,748 nm).

Inferred overall: none of these overturns the resource or the qualitative findings. Items 1–3, however, bear directly on the two quantitative claims about connectivity.

Limitations

Stated by the authors:

• One surgical sample from a patient with epilepsy. Subtle effects of disease or treatment cannot be excluded, some anomalies may be pathological, and normal human tissue is unlikely ever to be available by this route.
• Variability between individuals at nanometre scale is unknown. Association cortex is partly shaped by experience, which may make human circuits hard to compare across people.
• Residual split and merge errors remain; c3 analyses likely underestimate connection strength. Paths found with CREST need manual checking because of merge errors and synapse false positives.
• One lab cannot proofread the volume. The dataset is "incompletely scrutinized", and methods for finding meaning in connectivity data are "in their infancy".

Ours:

• Inhibition is systematically under-detected. About 35% of inhibitory synapses are missed against 11% of excitatory ones. The error rates come from 42 axons with per-axon rates from 0 to 50%. Any graph or model built from H01 will be biased toward excitation unless corrected.
• The E/I comparison between cell types is partly built in by the labelling rule. "Spine within 1 µm on the postsynaptic segment → excitatory" favours spiny (pyramidal) targets. On densely spiny dendrites it may also relabel shaft synapses (inferred). The paper does not report how many labels the overrides changed, or a comparison without them.
• The evidence for non-random strong connections is weaker than presented.
  • The reported test does not target the tail (item 2 above).
  • Artefact screening was applied to simulated, not observed, strong connections.
  • Tail null values rest on single simulated axons; for example, the null count at 10 synapses comes from one axon (Table S13).
  • The 15 µm displacement also destroys genuine geometric relations such as co-fasciculation. An excess over this null shows non-random placement, not "intentionality".
  • It is unclear whether the observed distribution excludes chandelier (AIS-targeting) axons. The simulation sample did exclude them, and they are a known multi-synapse motif (uncertain).
• Triangular-cell angles were measured without compression correction. The 28% compression is along the section's long axis. Nearest-neighbour pairs are not independent observations. Low-resolution centroid vectors (every 128th section) are used, and the anterior–posterior assignment is inferred from myelin, not from surgical orientation records.
• Missing modalities: electrical synapses (not assessed), transmitter and receptor identity beyond E/I, neuromodulators, synaptic strength beyond counts (spine and bouton size not analysed here), plasticity state and any activity.
• Reproducibility is limited in places: a proprietary keypoint detector, internal Google compute and task systems, and some BigQuery databases that need credentials (the data are downloadable). The tables are live Google Sheets rather than a frozen archive.

Relevance to our research (our inference)

• What a human cortical reconstruction captures about an individual. A static, fixed snapshot of about one-millionth of one person's brain (1 mm³ of ~1.2 × 10⁶ mm³; a guessed round figure), taken at the moment of surgery, from a diseased brain:

  • cell identities, positions and shapes;
  • synapse positions with E/I labels and known, class-dependent errors;
  • glia, vasculature and myelin;
  • a few individual oddities of unknown cause.

  Naively scaling 1.4 PB and 326 microscope-days per mm³ to a whole brain gives zettabytes of data and on the order of a million microscope-years (inferred). Acquisition also destroys the living tissue: it is possible only for tissue already removed from the person, here surgical waste. The cut sections themselves survive on wafers and can be re-imaged.

• What it does not capture. It records no memory, preference, history or behaviour of the person, and no activity. It has no molecular identity; E/I is inferred from ultrastructure. It contains almost no complete neuron or long-range pathway. The authors themselves place reading memories from wiring ("engramics") in the future. Their cited C. elegans result sharpens the point: 40% of connections differ between genetically identical worms. Individual wiring is large and specific, but no one yet knows how to read personal content from it.

• Consequence for a replica that carries on as a person. The state of the art in human reconstruction is a millionth-scale surgical fragment with no link to the person's behaviour (from the paper's scope). Treating whole-person anatomy as unavailable for the foreseeable future is our inference, and the timescale is a guess. Kurisutina's person "slot" therefore has to come from behavioural and interview evidence, and identity or continuity claims cannot be propped up by appeal to connectomics.

  • This extends the Dorkenwald conclusion: structural fidelity is a separate property to measure. In human tissue we cannot even measure it for whole cells.
  • It extends the Shiu conclusion. A connectome-only model there made coarse sensorimotor predictions because the wiring was near-complete and proofread, with transmitter predictions. H01 offers neither for any circuit.
  • The "rare powerful connections" are the kind of person-specific structure that could store learned information. That they are non-random rests on an untested tail excess in unproofread data, and that they carry anything personal is untested.

• For the validation design:

  1. Structural resolution row. Record per-class recall and precision (here inhibitory recall is about 65% vs excitatory about 89%), the proofreading state, and the share of edges attached to identified cells (17% here). An "individual anatomy" arm built on unproofread data would inherit a bias toward excitation that a population prior might not.
  2. Null models. Require the test statistic to target the claimed effect: a tail test for a tail claim. Report Monte Carlo uncertainty for rare-event null values, and screen observed and simulated data for artefacts identically.
  3. Label-rule circularity. When a labelling rule uses a feature that also defines the comparison groups (spines → excitatory; pyramidal = spiny), report a rule-ablated comparison. The same applies to any consciousness marker whose computation shares inputs with its label.
  4. Reconcile reported numbers by script before relying on them. This paper yields nine verified inconsistencies across text, legends and tables; two affect a headline statistic's legend and sample sizes.
  5. Biological unit row. Today a human "individual" reconstruction means one surgical fragment of uncertain normality with no same-person function (unlike MICrONS). No experience label exists for it, so it can inform structure-to-function modelling only through population-level inference.