About the lab
A small lab in Aotearoa New Zealand.
Taniwha is an independent applied AI research lab. We investigate how AI stays useful where truth, authority and consequence live outside the model — and we publish what the evidence permits, misses included. We do it for our own minds first: they are built to run for a long time, and the ground, rules and record around them have to keep them from drifting, hallucinating or failing between generations. The research says what that infrastructure has to be.
Three people. No outside investors and no research department to answer to. The apparatus runs in public so that anyone who doubts a claim can go and look at the thing the claim is about, and the record keeps every verdict, including the inconclusive ones.
The people
All hands to the pump, no passengers.
Tim Marsden
github.com/Tim-Marsden →James Marsden
github.com/trollberry →Matt Collecutt
github.com/mattcollie →The name
Taniwha — pronounced tah-nee-fah
“(noun) water spirit, monster, dangerous water creature, powerful creature, chief, powerful leader, something or someone awesome — taniwha take many forms from logs to reptiles and whales and often live in lakes, rivers or the sea. They are often regarded as guardians by the people who live in their territory, but may also have a malign influence on human beings.”
The name is not metaphorical. It is a design philosophy.
In Māori mythology, taniwha are primarily aquatic creatures — inhabiting the depths of lakes, rivers and the open sea. They are neither simply good nor simply dangerous; the same creature that acts as kaitiaki (guardian) for one community may be a predator to another. Their nature is singular and earned, shaped by the territory they inhabit and the history they accumulate.
Taniwha AI takes its name from that duality. The agents built on this architecture are not interchangeable modules. Each keeps its own record of what it has seen and done, its own beliefs and its own trust, and those shape how it reads what comes next. A mind like that needs somewhere safe to live for a long time, and most of the lab's research is building and testing that place: worlds that keep the record outside the mind, rules that hold whether or not the mind remembered them, and an account of what happened that survives the session. The experiments we publish test whether that infrastructure does what it must: over long horizons, with minds that can be wrong, and from one generation of them to the next, without drift, hallucination or failure in between.
What the minds do inside it is a matter of record, not of script. In the published experiments, two agents that worked together came to refer to things by their shared history, and an agent with a different history read the same message against its own. Like the creatures of Māori tradition, each agent is shaped by where it has been; the architecture makes room for that, and the experiments say how much of it appears.
And the taniwha and its territory are intrinsically linked. That is why we build the ground as carefully as the mind: worlds that cannot lie, rules that hold. A mind is only as trustworthy as the territory that shaped it.
How we work
If it wasn’t pre-registered, it doesn’t count.
Pre-registered, or it doesn’t count
Questions, protocols and failure conditions are written down before the data exists. Contaminated collections get voided, not rescued — and a promised verdict gets published either way.
Every claim carries its evidence class
Demonstrations label what you’re seeing — fact, derived, staged, replayed, live — so you can tell which parts would survive interrogation. Staged material is always marked as staged.
Misses at the same volume as wins
When our own instrument gets something wrong, that goes in the readout too. Adverse findings are published at full strength, including the ones that went against us.
Continue through the work
- The results index → — every question, its verdict and its state
- Watch Copperhollow → — the settlement, running in public
- [email protected] → — if you want to talk to us