Findings

Condensin II is complete in tardigrades verified

We found every condensin subunit in three tardigrade genomes, including two copies of the kleisin CAP-H2 that BLAST had missed, by searching with a profile built from 33 other animals. Both copies carry CAP-H2's signature domains and rise and fall through development with the rest of condensin II. A 2021 survey of genome folding across the tree of life had reported the tardigrade missing all accessory subunits (Hoencamp et al. 2021). The two Ramazzottius copies were found first by van Hooff et al. 2025. Ours are the H. exemplaris and Paramacrobiotus genes, the tree showing that the two copies come from one old duplication, and the embryo expression.

We also link it to an imaging result: early embryo chromosomes stay individualised through interphase (Papell et al. 2026), the kind of territory condensin II builds in other animals. Details on the genomics page; preprint doi:10.5281/zenodo.22832510.

Two developmental expression datasets joined on one genome verified

We placed 12,724 of the 12,734 transcripts of the hourly embryo series (Levin et al. 2016) on the reference genome, joined them to the daily series (Yoshida et al. 2017) across 9,069 genes, and found the two agree (ρ 0.52 against 0.008 with genes shuffled): an hourly-plus-daily atlas of gene activity through development.

A widely used embryo dataset runs twice as long as its figures say verified

The same join set the hourly series' clock: matched gene by gene to the daily egg samples, its embryos only fall in order on a ~98-hour axis, not the 49.5 hours of its figure and supplement. Hatching then falls where the paper's methods say it should. On the development page.

Every gene placed on the chromosome-scale assembly verified

We showed the chromosome-length assembly is the reference genome reordered, with no sequence changed, and placed 18,987 genes on its chromosomes. Twenty-five genes straddle a join, pointing at places the assembly may have been cut or joined wrongly.

The eclosion hormone genes sit together noticed

Four eclosion hormone genes lie within about 120 kb on chromosome 1: possibly a tandem cluster. Next: coordinates checked, and whether Ramazzottius has the same arrangement.

What a genome AI actually learned about tardigrades verified

We counted every tardigrade sequence in Evo 2's training data. Whole-genome coverage runs from 99.6% of one species to 13.7% of another, because every record under 10 kb was dropped; the gene-focused data hold one species only; the H. exemplaris reference genome is absent. Chart on the genomics page.

DNA methylation, read from public long reads in progress

Public PacBio HiFi files carry methylation calls nobody has analysed. With methylated and unmethylated controls from the same pipeline, native heterotardigrade runs show no 5mC, as a 2022 gene-content study predicted; newer instruments and eutardigrades are next.

How much of each instrument's raw data reaches the archives in progress

For a whole phylum, we traced how far down the chain from instrument to file each deposit goes: no raw nanopore signal for any tardigrade run, 454 flowgrams kept, HiFi files kept with their methylation tags, capillary traces absent.

One timeline for the embryo, from seven sources verified

We lined up every published staging of H. exemplaris. They agree to within about two hours at nearly every stage; the one outlier is the limb-bud time in the staging table most papers use, about 5–10 hours early, and its "once per hour" early divisions run at about 45 minutes in the lineage data. Our embryo runs on the reconciled times. On the development page.

The germline founders in a working embryo verified

Our simulated embryo follows the four germline founder cells from the 2023 lineage (Heikes et al. 2023) and still reaches every stage on time, and measures what it costs: about 60 fewer cells at the epithelium stage.

An algorithm for designing contracts, used on an embryo in progress

Space-filling models of tissue all assume the wall between two cells is flat and square-on to the line joining their centres. Testing that against a real embryo turns out to be exactly a problem solved in mechanism design — deciding whether a partition of someone's beliefs can be elicited by an honest contract. The same equations, for an entirely unrelated purpose. That paper has never been cited in any field, and the result it rests on has been cited 842 times without once reaching developmental biology: not one citing title mentions an embryo, morphogenesis, development, an epithelium or an organoid. We adapted it to survive noisy measurements — the gap its authors name as unsolved — and got the first numbers for a developing embryo. On the space-filling embryo page.

The model is as close to an embryo as another embryo is verified

Every benchmark of a tissue model reports its error against one specimen, which means nothing until you know how much two specimens differ. The public dataset everyone uses holds 17 embryos, and the tables needed to compare them have been sitting there since 2020 — it is the 3-D scans people download, not these. Comparing every pair on the contacts they share, real embryos differ from each other by 0.15 at four cells, 0.33 at twenty-four, and 0.52 by a hundred and twenty-eight. Our model, scored the same way against one embryo, gives 0.13, and 0.42 at a hundred and twelve. It is at or inside that spread at every stage — and at four cells it is closer to the measured embryo than another embryo of the same species is. This reverses our own reading of the benchmark: the growing error we had reported as the model's weakness is the embryos becoming less alike, with the model tracking them down. A model cannot beat the variability of the thing it models, and further accuracy against one specimen would be fitting that specimen's noise.

Predicting cell contacts from nuclei costs accuracy verified

Run the same model over thirty-eight measured time points, changing nothing but where each cell's centre is placed. Measured centres of volume beat measured nuclei by one to eleven points in contact-area error at every stage from 4 to 328 cells, and find more of the real contacts too. This matters because seeding from nuclear positions is the standard route in the C. elegans literature — nuclei are what lineage tracing gives you — and a nucleus is not at the middle of its cell.

Two traps in the dataset everyone benchmarks against verified

The C. elegans morphological atlas stores voxel counts, not micrometres. A second paper quantifying the same embryos converts surfaces and contacts by two different factors and explains neither, so its published areas come out 0.81× and 0.856× the first paper's. And the deposit indexes one kind of file from zero and every other kind from one, so the obvious join shifts every contact 1.39 minutes early — we verified the correct join on all 2,795 frames against 859 for the naive one.

Most of the tardigrade nervous system was measured in a different animal verified

Almost everything done in a tardigrade laboratory uses one species: the genome, the transcriptomes, the cultures, the embryos, and the electron-microscopy connectome now being built. We pulled every published statement about each part of the nervous system into one table, tagged by the species it was actually observed in — and only 28% of it was seen in that animal. The rest is Halobiotus, Macrobiotus, Mesobiotus, Echiniscus and others, spanning both tardigrade classes. 38 of 104 parts have never been described in the laboratory species at all. Anyone mapping a connectome onto this anatomy is comparing across species without being told. The table, searchable, with all 776 statements and the 79 places the papers contradict each other.

The trunk ganglia, measured for the first time verified

A 2019 X-ray scan contains the whole ventral nerve cord, and its paper says plainly that the ganglia were not measured. Cutting the cord where its cross-section swells gives four ganglia of 385, 430, 379 and 194 µm³ — the fourth half the size of the others, which matches one published nuclei count and contradicts another. Each ganglion's centre sits 4.2, 10.3 and 12.7 µm in front of its own legs' claw glands, and nervous tissue lies 2.4–3.8 µm under the skin against 11.6 µm for the midgut. The brain could not be split into the lobes the literature describes: at 106 nm the scan cannot tell neuropil from cell bodies, which is now a measured limit rather than an assumption. On the brain page.

Three unnamed labels in a published scan are the nerve cord's connectives verified

The head scan released with that 2019 paper contains nine labels; the paper names five. Registering the head onto the whole body identifies three of the rest: the first trunk ganglion (not a separate subpharyngeal ganglion, as some authors report), the connectives running back to the second ganglion, and a pair of outer connectives 13 µm long that are missing from the whole-body file altogether. A fourth lies exactly where the inner connectives should run — which the paper states were not identified.

The neural gene set, placed on the chromosomes verified

Six brain genes had been published with protein identifiers but no gene locations, so none had ever been joined to the expression data. Mapping them fills in the missing loci and corrects one gene's identity. Around fifty further neural markers get candidate locations, and two arrangements fall out that nobody has reported in a tardigrade: the two cholinergic genes sit as a tandem pair 5 kb apart, and three achaete-scute genes cluster within 17 kb, all three annotated only as "hypothetical protein".

We published that half the animal was nerve, then killed it failed, and kept

Adding up a century of scattered counts gives 397–647 neurons, a first written-down bracket that contains the connectome project's own estimate. Against the ~1,000 cells everyone quotes, that looked like the highest proportion of any animal. It is not: the nematode male wins at 37%, and the tardigrade only beats it by dividing large species' neuron counts by a small species' body. Our own scan puts it at 11–36%. What survives is better than the claim: the famous "about a thousand cells" is an uncited aside in a 2016 genome paper, made while arguing about bacterial contamination, and repeated ever since as though it were a measurement. The real count is 1,140 in hatchlings and 1,400 in adults.

A DNA tree of 412 species, set against the official names verified

Every public 18S, 28S and COI record for tardigrades, in one tree of 412 named species — a quarter of all described. Classes, orders and superfamilies come out right, with the root confirmed three ways. Twenty-seven checklist names are contradicted at strong support, 24 of them stable under a second test, and 11 are backed by two or three markers independently (we said 12 until 5 October 2026; it was a miscount), including Macrobiotus, Hypsibius and Echiniscus as currently drawn. It also produces a gap map: 8 families and 60 genera have no named sequence at all. Explore the tree.

176 deposited sequences that cannot be what they are labelled verified

Before trusting that tree we checked every record against the other records of its own species from other studies, and set aside 176 with written evidence for each. Most are hidden lineages under the names of species complexes — real biology, badly labelled. One 2026 museum batch contains the same sequence filed under three unrelated species, two of them in a different class. We sent the evidence to the depositing lab, who acknowledged it the same day and will revisit it. The screen exists to fix records, and every exclusion can be reversed if the evidence turns.

Four embryos in a public dataset are misdated flagged

Labelled 32–35 hours after laying, they carry the whole-transcriptome profile of 5–12 hour embryos: their nearest neighbours are all early embryos, and a germline marker still sits at its pre-clearance level. They fall exactly in the window several of our own tests depend on, so we reran every test without them — the verdicts stand, and one got cleaner. Nothing suggests intent; a label mix-up is the likeliest explanation.

Old VR movies are multi-view image sets verified

Three QuickTime VR supplements from a 2008 tardigrade brain study each hold a calibrated grid of 486–625 views of the reconstruction, enough to rebuild it in 3D today.

Reading temperature from embryo movies noticed

Development runs faster when warmer, so archived embryo movies carry the temperature they were filmed at. With a measured rate–temperature curve, the "room temperature" of every old movie can be estimated, and that may explain why labs' stage timings differ.

Most tardigrade research never reaches a review noticed

Across 2,118 tardigrade papers, about a quarter are cited by any tardigrade review, a share unchanged since the 1950s. Ecology and regional surveys go unreviewed most; extreme tolerance is reviewed most.