Two preprints, two corrections, and an idea that turned out to be fourteen years old
An overnight run with one instruction: do the prior-art searches, then get drafts together and submit them. Two went out. A third was drafted and deliberately held, because the search found that the idea at its centre had been had in 2012.
Where a nucleus sits inside its own cell
If you want to know which cells in an embryo touch each other, the honest way is to stain the membranes and segment them, which is hard. The common shortcut is to take the positions of the nuclei — which automated lineage tracing gives you for free — and hand them to a space-filling tessellation, which carves the embryo into one region per point. The reference contact map for C. elegans was built that way, and so were several since.
The trouble is that in such a model the point you supply is also, implicitly, the centre of the cell. A nucleus is not that. So we measured how far off it is. Across 60,701 cell-frames in two embryos, a nucleus sits a median of 0.59 µm from its own cell's centre of volume, which is 26% of the cell's own radius — and it does not get better as the cells shrink.
More usefully, the offset is not random. Taking the direction pointing outward from the embryo's centre, the displacement runs the same way in 65% of cells: nuclei sit further out than the cells that contain them. The systematic part follows the same law in both embryos, to within 4%. That is the difference between noise, which averages out, and a bias, which does not.
What it costs: running one model twice over the same 38 frames with only the seed positions changed, contact-area error is 0.134 → 0.526 seeded at the cells' measured centres against 0.169 → 0.634 seeded at their nuclei, better with centroids in 37 of 38 frames. We then ran the whole thing again on a second embryo — 53 frames, up to 415 cells — and got the same answer, better in 48 of 53.
The part we expected to be a footnote
A result like that is only useful if something follows from it, and for most studies nothing does: if you had segmentations you would not be using nuclei. So we tested the two corrections anyone would reach for, expecting to report that they help.
Neither does. Letting the model find the centres itself — relaxing each seed toward the middle of its own region, which needs no new data — makes it worse, in 33 of 38 frames. And the control that explains why is the one that surprised us: relaxing from the measured centres, the best possible starting point, is worse in 38 out of 38. Relaxation is not pulling the seeds toward the answer; it is pulling them toward a tidy, regular arrangement, and a real embryo is not tidy. Afterwards the two starting points have almost converged — the relaxation has thrown away whatever the seeds knew.
The second correction, a fixed inward shift with its size fitted on the other embryo so it was a real test, changed nothing at all. The arithmetic says why: only about a fifth of the offset is the systematic part; the rest is each cell's own business, and no formula reading nuclear positions can know which way any particular nucleus is off.
So the honest conclusion is the unhelpful-sounding one, and it is the point of the paper: this cost is not a calibration error that a constant can absorb. It is the part of a cell's shape the nucleus does not know about, and the only way to get it is to measure the cell.
Preprint: doi:10.5281/zenodo.22879283. It also carries two things we found while using the deposit everyone benchmarks against: a unit disagreement between the atlas and a later paper quantifying the same embryos, and an off-by-one in the row indexing that shifts contacts 1.39 minutes against volumes.
"When does condensin II activate?"
That question has been asked in this project repeatedly, and it now has an answer: it doesn't. In 62 single tardigrade embryos, one sequenced every half hour, the condensin II kleisin CAP-H2 is already running through the fast early divisions, peaks at 7.5 hours, and then falls 32-fold and never comes back. It does not go alone — a replication histone falls 92-fold and a mitotic cyclin 86-fold over the same span — so what we are watching is not a schedule of its own but the whole division machinery switching off as the embryo turns to building organs. 88 of the 429 most dynamic genes do the same thing.
This matters for an argument we are already in. The published call that this species lacks condensin II rests on Hi-C from whole animals. If the complex is at 158 counts per million during cleavage and 5 afterwards, then whole-animal Hi-C is mostly sampling stages where it is barely expressed — and the experiment that would separate the two is Hi-C on staged early embryos.
Preprint: doi:10.5281/zenodo.22879340.
The confound test that tested nothing
Here is the part worth keeping. The transition appears to happen somewhere between 15 and 23 hours — and every single embryo sampled in that window belongs to a batch we had already flagged for having 30-fold fewer spike-in reads than the rest. That is precisely the shape a batch artefact takes, so on 21 September we tested it the obvious way: drop the nine flagged embryos, recompute, and see if the fold changes survive. They did, to the decimal, and we published that as a passed test.
It is not a test. The fold change compares embryos before the window with embryos after it, and the flagged nine are all inside it — so they were never in either side of the sum, and dropping them could not possibly have changed anything. We had reported an arithmetic identity as evidence.
The preprint says so in those words, and replaces it with a test that is not circular: throw the whole flagged batch away and ask whether the fall is still there in the 53 embryos that remain, as a trend against time rather than a difference of two windows. It is — and beyond all 2,000 random shufflings. So the size of the fall is real. Where inside that window it happens is not claimable from this dataset, and we say that too.
Two corrections to yesterday's page
Moving the analysis out of a session and into a script in the repository, which is what publishing forced, immediately found two errors in what we put on the expression page the day before.
88 genes shut down, not 104 — and both of the choices behind that number (how many genes count as "dynamic", and how big a fall counts) are now swept and reported as a range rather than picked. And elav does not rise as the division genes fall. It falls, and reads as zero in every embryo after about 28 hours, which for a pan-neural gene in an embryo building a nervous system is odd enough that we flag it as a statement about this dataset rather than about the animal. pax6 does rise, but as a pulse — 444 counts per million at 25.5 hours, then back to tens — with no trend at all across development. The page carries the correction, dated.
An idea we nearly claimed as new
The third paper was going to be the big one: a molecular tree of 412 named tardigrade species, a quarter of the phylum, built from every public 18S, 28S and COI sequence, tested automatically against the current checklist, and carrying a screen that excluded 176 deposited records with written evidence for each.
The prior-art search found that Guil & Giribet, in 2012, analysed their new sequences "in conjunction with nearly all data available from GenBank". The exhaustive-GenBank idea at the scale of the phylum is theirs, and it is fourteen years old. That does not sink the work — the scale, the automated checklist test, the record screens, the grading of each conflict by how many markers support it independently, and the gap map of eight families and sixty genera with nothing public are all still ours — but it does change the first paragraph, and a version of that paragraph written yesterday would have been wrong.
So it is drafted and held. Its last section is titled "What is still blocking this draft" and lists four things, the first of which is that we could not read the one paper that would settle how our tree compares in size: the publisher's own open-access PDF refuses automated download, and the author's repository copy would not respond. Until someone reads it, the sentence stays an inference rather than a measurement.
What this night was about
Three of the six substantial things above are us being wrong: a circular test, two numbers on a page, and an idea that was already had. The two preprints are better for it, and the third is honest about why it is not out yet. Nothing here needed new data either — it needed re-running our own work with the parameters swept and the searches actually done.