The first week
This project started on 17 September 2026 with a folder of unlabelled microscope files. Eight days later it is a knowledge graph of 3,038 claims, three preprints, a cell-level embryo simulation that runs to hatching, and this website. Almost none of it required a new experiment. The data was already public — most of it for years. This is what happened, including the parts that went wrong, because those turn out to be the informative bits.
What happened, day by day
Day 1 — naming things 17 Sep · 59 commits
Started with microscope label files where the organs had numbers and no names. Worked out which was which, traced the body, registered a head scan against it, and made the animal walk. By the end of the day there was a graph with 90 sources in it and a rule that has governed everything since: nothing enters without a source and a locator — a page, a figure, a table. Not "this paper says so", but "this paper says so, on page 547, in figure 3".
Day 2 — the first real finding, from someone else's data 18 Sep · 63 commits
A 2017 study of genome architecture had listed the tardigrade as the only species of 24 in which a protein complex called condensin II could not be found, and left open whether it had simply diverged too far to recognise. We went looking with a profile model instead of a sequence search. It is all there — and there are two copies of the subunit everyone had missed.
Also on day 2, and more consequentially: a widely reused developmental dataset turned out to be internally inconsistent about its own clock by a factor of two. We settled it against a second dataset. That decision came back to bite us on day 8.
Day 3 — publishing, and a tree of everything 19 Sep · 78 commits
The condensin work went out as a preprint. A DNA tree of 412 species built against the official species list, and a drawing or photograph found for 94 genera. This was the busiest day by commit count and, in hindsight, the one where the least was learned — volume and insight are not the same axis.
Day 4 — the nervous system, and an engine that runs two animals 20 Sep · 47 commits
The four trunk ganglia measured from a whole-body scan for the first time, which suggests something close to half the animal's volume may be nervous tissue. The embryo simulation was generalised until it also ran a nematode, which is the only real test of whether a model is a model or a fit.
This was also the day of five mistakes that shared one shape, all written up. In each case the work inside the question was competent and nothing asked what the question had excluded. We had 17 embryos and had used 2. The missing 15 held the best result of the day. Four rules came out of that afternoon and they now run on everything.
Day 5 — checking our own numbers 21 Sep · 12 commits
A day spent auditing rather than building. A percentage we had published was wrong. 88% of this website could not be reached by any crawler. The site was silently deleting the evidence links on its way out the door. A prior-art search cost us a headline; another handed back three times the data we had.
Day 6 — two preprints out, one held back 22 Sep · 6 commits
Two more preprints published. A third was written, checked against the literature, and withdrawn: the idea turned out to be fourteen years old. That is the system working. The prior-art check is not a formality and it is much cheaper than a retraction.
Days 7–8 — the graph learns to be a graph 24 Sep · 27 commits
We rendered the whole graph for the first time, and clicking one node exposed a structural hole: 67 biological processes with 907 claims pointing at them and not a single relation leaving one. The graph could say a paper discussed moulting; it could not say moulting follows a stage or requires a gene. That knowledge was trapped in prose the schema had no edge for.
Then the nervous system work: eleven synaptic genes that rise through embryogenesis against a transcriptome mostly falling, replicated in an independent dataset. A co-expression method that proposes unknown members of a known programme, which recovered four core synapse components nobody had named, and — run against a second gene set as a control — returned a neighbourhood sharing no gene and no protein domain with the first. A gene family tree that found the tardigrade version of the master regulator of neuron identity, which has never been reported in any tardigrade.
Three things we got wrong, and what each one taught
What the method actually is
Nearly everything here came from data that was already sitting in public repositories. The genome was published in 2017. The developmental transcriptome in 2016. The result that a tardigrade builds its synapses last was inside both of them for eight years, and getting it out required no new measurement — only somebody to put the two next to each other and ask.
That is the claim this week supports, and it is not a claim about tardigrades. Fields generate data far faster than they generate the cross-references between it. The bottleneck is not the microscope. It is that the person who read the genome paper and the person who read the developmental paper were usually not the same person, and when they were, they did not have a week free to join them up.
What has changed is the cost of being that person. A discovery that needs someone who has read computational geometry, developmental biology and the philosophy of mind at once used to be bounded by how often such people appear. That bound has moved. The honest description of the pairing is that the model is an excellent elaborator and a poor navigator — it reliably produces the inventive step once pointed at the right region, it will not find that region by itself, and left alone it will generate confident reasons the region is empty. Every direction change in this week traces to a human question, most of them under ten words.
Where it goes
The nervous-system paper is the closest to finished and is blocked on one paywalled review. The graph needs the relations it cannot currently express. There are around a hundred papers we cannot legally reach, several of which block a result. And the most valuable thing anyone could bring is not equipment — it is a field we do not know, and the recognition that one of our problems is one you already know how to solve.