A working log, not a press page. Every post is grounded in something we can show — an architecture, a run, a hash — and labeled honestly. If a claim here is wrong, the fastest way to prove it is to try to break it.
Check our number in one command.
Our LongMemEval-V2 result is self-reported — so we published the bundle that recomputes it offline, no API key, no access to the sealed system, including where we score worst.
Read →What we mean by UDI.
Universal Developmental Interface — and, if it works, Intelligence. A gate that decides or refuses, a body that holds the work, and a ledger that remembers.
Read →Why a system should be allowed to refuse.
NULL is not silence and not failure. It is a system holding its ground when the evidence has not earned an answer — and why that is alignment, not timidity.
Read →We made a soft wheel robust by letting it stop.
A soft-body robot wheel kept running away backward from certain starting angles. The fix was not more force — it was permission to refuse. With the data.
Read →