The day we checked our own numbers

No new organism, no new dataset, and one of the more useful days of the project. We went to close an open question about our own work and found the number it rested on was wrong. Then we measured whether this site can actually be read and found that 88% of it could not. Then we found the site was deleting the evidence behind the glass tardigrade on its way out the door. Two prior-art searches cost us a headline each — and one of them handed back three times the data we had.

A published percentage, and why it was wrong

We publish a table of every statement anyone has made about each part of the tardigrade nervous system, with the species it was actually observed in. The headline was that only 28% of it came from Hypsibius exemplaris, the animal every laboratory, every genome and the connectome effort all use. We were about to send that to the group building the connectome.

It is 42%.

The error was not arithmetic. The table is built by pivoting our graph by structure, so a claim attached to eleven structures appears eleven times, and the species share was summing those per-structure counters. Every claim was silently weighted by how many structures it touches — and that weight is not neutral. Comparative-anatomy claims about other species attach to more structures each (2.58 against 2.20), because a paper surveying a whole head touches more parts than a paper on one ganglion. So the repetition pushed the H. exemplaris share down.

Two things survived untouched, and they are the ones that matter most: 38 of 104 structures have no H. exemplaris statement at all, and there are 79 pairs of directly contradicting claims. Both are properties of a structure or an edge rather than of a count, so the multiplication could not reach them. The headline sentence also survives — 58% of what is known was seen in another animal — it is just less dramatic than what we published.

The correction is dated on the page rather than edited away, and the cause is written into the code so the next person to touch it can see the trap. We then went looking for the same shape of bug everywhere else: a static pass over every tool found fourteen places that aggregate this way, and checked each for the two conditions that make one harmful. Only the one. Two clears worth naming: our citation-flow analysis takes one primary field per work, so it is clean; and the expression figure in our published condensin preprint sums transcript values into genes, which would have spuriously correlated any two genes sharing a transcript — 0 of 20,816 are ambiguous, so it is unaffected.

The corrected table · the counting trap is now finding P029, because it is a general property of aggregating any knowledge graph, not a quirk of ours.

Reading order, measured instead of confessed

Every write-up of this sort carries a sentence like "our search may have been biased", and it is usually unfalsifiable. It is not, because version control records the order in which we read things. Each source's first appearance in our source list is a timestamp, independent of anything we later concluded.

Correlating a source's reading rank against the share of its claims that concern the model organism, across the whole graph: 1,384 claims, 150 sources, Spearman −0.568, p < 0.0001 against a null that shuffles reading order and holds the corpus fixed. Sources we read later carry less H. exemplaris. So our corpus over-represents it, and every species share we publish is an upper bound. That is now measured rather than flagged, and it is finding P028 — the method applies to any literature-derived database, any systematic review, and any corpus an agent assembles.

Eighty-eight per cent of this site could not be read

Before redesigning anything we measured whether a reader can actually move through it. The result changed the question. Of 1,588 pages, 83 were reachable from the front page by following links. The species section — 1,551 pages, 884,673 words, 88% of everything written here — built every one of its links in JavaScript, so nothing that does not run a browser could reach any of it. That excludes archival crawlers, most first-pass search indexing, and any agent that fetches a page rather than rendering it.

The fix was a list. Every species, A–Z, as plain links, at the foot of the species page. Reachability went from 5% to 100% and orphan pages from 335 to 0. The radial tree is still the way to explore; this is the way to find, which is what someone who already knows the name wanted anyway.

The site was deleting its own evidence

Worse, and older. Click any organ in the glass tardigrade and you get its measured volume and how it was identified. Behind it sits 691 KB of claims — every published statement about that organ, with its source and page. The public build was shipping {structures: {}}: all of it deleted.

That was correct when the graph was private. It has been wrong since 19 September, when we made the graph public, and nobody noticed because the local copy looks complete. So every visitor has been seeing measured volumes and nothing to follow.

Now 1,122 claims and 1,165 citations that link straight to the paper ship with the model, through the same screen every public graph export passes. The build's own leak check caught two attempts that would have shipped raw graph identifiers — including identifiers embedded inside prose — and those are fixed rather than excepted. Each structure also carries a Further in block, placed above the evidence rather than below it, because a link after 192 claims is a link nobody reaches. And it works in reverse: /glass/#s=brain opens a structure directly, so the nervous-system table now offers "see it measured in the glass tardigrade" on the eight of its 104 parts that are also in the model.

Two searches that cost us, and one that paid

We had a result we liked: that every benchmark of a tissue-geometry model reports its error against one specimen, and nobody knows how much two specimens differ, so the numbers are uninterpretable. We checked whether anyone had done it. They had. Azuma, Okada and Onami (2023) quantified inter-individual variability across 52 C. elegans embryos; Guignard et al. (2020) did it for ascidians, down to the area of individual cell contacts. Our framing is withdrawn.

What survives is narrower and is written that way: neither reports the quantity we measure — the magnitude of disagreement in contact area between embryos that both have that contact — neither frames variability as a ceiling for interpreting a model benchmark, and the number for the deposit everyone benchmarks against is ours.

Correction, 5 October 2026. Half of that was still wrong. We had read Guignard et al. from the abstract only, and the full paper does report the quantity: in ten ascidian embryos, the relative areas of contacts shared by homologous cells varied by less than 20% (their Fig. 2E). What remains ours is the use of that variability as the ceiling for a model benchmark, and the number for the C. elegans deposit. The preprint carries the same correction in its second version.

And the search paid for itself. Azuma's 52 embryos are public, CC BY, with their membrane segmentations: three times the 17 we had, acquired independently, segmented by a different pipeline. That is exactly what decides whether our ceiling is a property of the animal or an artefact of the dataset — and either answer is worth publishing. The second search, on whether anyone has tabulated species coverage in tardigrade neuroanatomy, found that measuring taxonomic bias in a literature is a well-established genre. Our difference is the unit: statements about a named structure per species with locators, rather than papers per taxon. That is a smaller claim than we started with, and it is the true one.

A question closed, thirteen years on

When we read a paper, any sentence where the authors say something is unresolved becomes an object in the graph carrying the quote, and later reading is checked against it. One of those was recorded on 17 September from a 2017 review: how are tardigrade eyes innervated, and do they keep a connection to the skin? It had been open since 2013.

It closed. A 2026 reconstruction counted the cells in one eye and found four — a pigment cell, a rhabdomeric cell, a ciliary cell and a support cell — and not one of them has an axon. Five separate neurons beside the eye carry the signal out. The old argument about whether the eye is part of the brain or part of the skin turns out to be the wrong shape of question: it sits inside the brain, and appears to be a piece of skin that sank in and stayed epithelial.

The full story is now its own page, and it ends where the evidence ends: nobody has traced where those five axons terminate, so "how the eye is wired" is answered at one end only. The diagram draws that as dashed, on purpose.

What this day was actually about

Nothing here required new data. A published number was corrected by re-reading our own code, a site was made readable by adding a list, evidence was un-deleted by removing a line, and two results were made smaller and truer by searching properly before writing them up. The most valuable thing we did was check things we had already decided were finished.

Both of the mistakes are now findings in their own right, because neither is ours alone: the counting trap applies to any knowledge graph, and the reading-order measurement applies to any corpus anyone assembles, including the ones assembled by machines.