How Tardigraph is built

We rebuild one animal, the water bear Hypsibius exemplaris, from what science has already published about it. Every fact goes into one knowledge graph with the page it came from, and the models, pictures and simulations on this site are built from that graph and tested against real measurements.

The loop

Each part of the animal goes round the same loop:

1. Gather. Papers, genomes, scans, videos and supplementary tables, recording what each one's licence allows.
2. Record. Claims go into the graph, each with its source and a locator: a page, figure, table, file or database record.
3. Build. Simulation rules point at the claims they encode.
4. Test. We compare the model's output with measured data and write the verdict down in words.
5. Predict. We write down what the data imply that nobody has checked yet.
6. Confirm or break. When new data arrive, the predictions are resolved and the model revised.

Five rules

The graph stays trustworthy because a few rules never bend:

• Nothing without a source and a locator. If only an abstract was read, the source says so and the claim is marked medium confidence. No paper is cited for anything it wasn't read to say.
• Guesses are labelled. When a model needs a number no source gives, the number becomes an assumption in the graph, and one command lists every assumption the model rests on.
• Disagreements are kept. When two sources disagree, both claims stay, joined by a "contradicts" link. A newer result that supersedes an older one "refines" it; nothing is deleted.
• Failures stay. Every comparison with data is a test with a date and a verdict: pass, close, partial or fail. The failures are in the graph beside the passes.
• Species are marked. Anything learned from another tardigrade, a fly or a worm is marked with its species and only carried over to our water bear through a stated inference.

A validator checks the graph's structure after every edit, and the site build refuses to publish if anything private (a local file path, a flag not yet investigated) has slipped into a page.

Look before building

Before working anything out ourselves, we look at what the field has already measured. Most fields have gathered far more than anyone has joined together: expression data deposited for one question can answer another, and a genome assembled years ago can hold a gene nobody searched for. For each kind of data we list what can routinely be derived from it, then look for the outputs nobody has reported. That is how our condensin II preprint came about: public genomes, public expression data, public structure-prediction tools, and a question nobody had put to them.

Papers also say what they couldn't answer. Each "it remains unknown whether…" becomes an open question in the graph, and each new batch of reading is checked against the list, so an old question answered by a later paper gets noticed.

Who does the work

Meow-Ludo Meow-Meow sets the direction and makes the calls. The reading, coding and checking are done with AI research agents working inside the rules above. Several agents can read in parallel, each writing to its own staging file, and their work merges into the graph only after validation. Standing decisions, such as what may be published and how the site should sound, live in written instructions and a memory that every session reads, so a correction made once holds from then on.

While reading, the agents also watch for data that aren't quite right: duplicated panels, impossible numbers, sequences filed under the wrong species. These go into a private register and are checked one by one; a flag is not an accusation. One of them, a batch of public DNA records in which the same sequence was filed under three unrelated species, has been reported to the people responsible.

What counts as new?

This project is run by a person and a language model together, and on 20 September 2026 we disagreed sharply about whether a result was worth publishing. It turned out we were applying two different tests without either of us noticing: does this statement exist somewhere? against does the audience who needs it know? The second is the right one, the first gets worse the better you are at searching, and the difference nearly cost us a paper. The disagreement, the measurement that settled it, and the rule we now share.

What went wrong, and what changed

• We had seventeen embryos and used two. A public dataset we benchmark against holds 17 measured embryos; we had downloaded the large 3-D scans for two and stopped seeing the rest, while writing "we need more specimens" as a blocker. The other fifteen held the most important result of that week. Now: before asking for more data, list what is already on disk and say what each part could answer.
• We published a trend that was our own setting. A parameter that exists only to fix a coordinate choice was left at a value that bit harder as the data grew, manufacturing a decline we then explained — and cleared of being a dataset artefact with a test that could not see a fault in the method. Now: sweep every free parameter before interpreting anything that varies with sample size.
• A failed build was published and deleted about 2,950 files from this site for a few minutes. Now the build leaves a marker only when it finishes cleanly, and publishing refuses to run without it.
• The embryo outgrew the browser. Real cell membranes for hundreds of cells don't run in a web page, so the physics moved to a graphics card as a local program. Its first test against measured worm embryos failed badly, because it misjudged which parts of neighbouring cells touch; the fix brought it within reach of our browser engine.
• Our DNA tree disagreed with the official classification in 27 places. Checking every one against the literature, 8 match published findings, 8 are debated, 5 aren't discussed anywhere, and 6 were our own artefacts. The artefacts stay listed on the page.
• Source names came out as initials ("V et al. 2017") for 54 papers when the graph went public. We refilled them from each paper's DOI record.

What is open

The graph, our models, measurements, tools and findings are public under CC BY 4.0, with the tools on GitHub. The papers themselves are not: we cite them and link to them. Figures and videos appear here only when their licence allows it, and every source is listed on the credits page.