Most AI systems have one move: assert. Ask a question, get an answer — confident whether or not the evidence was there. They rarely say "no," rarely show the grounds they acted on, and rarely remember, in any way you can audit, what they committed to yesterday. UDI is our attempt to change the default.
UDI stands for Universal Developmental Interface. We also read the last letter as Intelligence, and the ambiguity is on purpose: the interface is not a wrapper around an intelligence, it is the shape the intelligence takes as it develops. The interface is the thing growing up.
Three parts carry it, and we say them the same way everywhere:
Mind — it decides, or it refuses.
At the center is a verified gate. Evidence flows in; the gate resolves it into one of three things: commit, refuse, or NULL — nothing yet. Refusing is not failure. It is a first-class answer: the system declining to act when the evidence does not support acting. Most architectures cannot say no without being told to. This one holds "no" as a real output, with its reasons attached.
A system that cannot refuse cannot really be trusted to agree.
Body — agents that hold the work.
Around the gate is a substrate where independent agents do the work — each with least privilege, loop guards, and human approval where it matters. They can be built and deployed by different parties. The open version of this substrate is TinyHive: a hive that runs on your own hardware and holds pieces of a life so a person is not the switchboard for all of it.
Spine — the ledger that remembers.
Every commitment and every refusal is written to a hash-chained, content-addressed record. Change one byte and the address changes — so tampering is not hidden, it is a different thing entirely. That record lives at ÃDO. It is not a token and not a currency; it is provenance, an address that is its own proof.
Why "developmental."
The word matters. A model is usually trained once and then frozen. UDI is meant to develop: to accumulate memory that carries forward, to let pressure from unresolved evidence return to the same place that created it, and to hold coherence just below the point of collapse rather than forcing a premature answer. Whether that developmental process becomes anything worth calling intelligence is an open question — one we would rather test than assert.
So we test it. In the experiments, a soft-body robot wheel learned to move more robustly once we let it refuse — stop, let momentum bleed, re-center — instead of forcing motion through a bad state. In the benchmark, we report a memory result with the hashes needed to check it. When something fails, we keep the failure. That is the whole discipline.
What this is not.
This is development-stage research. We do not claim a verified public benchmark win; we label what we show as observation, hypothesis, or demonstrated result. The protected internals of the gate stay private — they are background IP, and you will never find them on these pages. Nothing here is a claim of consciousness or general intelligence, and nothing here is financial advice.
What we do claim is narrow and, we think, worth building: that accountability can be a property of the design rather than a policy bolted on afterward — that a system can be made to show its grounds, keep its record, and say no. If we are wrong, the fastest way to show it is to try to break it.