Falsify us
Tell us where we are wrong.
Evidence-first only counts if you can check it. So this page stays open all the time. Pick any claim we make, anywhere on this site or in the work underneath it, bring your source, and we will test it with you. If it does not hold, we change it.
Every claim is meant to be tested
We answer in plain words
We publish the corrections
A claim you cannot break is just a slogan. Try to break ours.
Where to aim
The places we most want you to test.
The strongest claims should take the hardest look. Start here, then send what you find.
The UNI working paper
Our work in active inference, written to be testable. It is a preprint on Zenodo with expert review still pending, not peer-reviewed, and we present it that way. Read it and try to break the math.
Open UNIThe numbers we publish
Where value goes, what is measured, and what is only claimed. Check the receipts against what you can see for yourself.
Open the Evidence ExplorerThe live labs
Five interactive POMDP labs you can move the dials in, plus a public MCP server any LLM can call. Watch the math move and tell us if it does not behave the way we say it does.
Open the labsThe benchmark world
The Stratified Palimpsest. A pure active-inference agent on a world made hard on purpose. We claim no reward is paid, no semantic state leaks, and same-seed runs are identical. A separate verifier you run yourself proves it or names the frame where the claim broke.
See the benchmarkFile a correction
Show us the part that does not hold.
Name the claim, point us to where you found it, and bring your evidence. The more specific you are, the faster we can check it and the sooner we can correct it in the open.