Why the lab exists

In an AI system, what should live where?

An agent will be wrong. It will be replaced by a newer model. It will be interrupted mid-task. Whether the permissions it was granted, the record of what it did and the work it still owed survive those events depends on where they live. Kept in the model’s context, they last as long as its attention. Kept outside it, they hold whether or not it remembered.

That is the question the lab was set up to answer, and every result we publish is an attempt to answer some part of it experimentally.

What we hold

Rules of the systems we build

These are design commitments, not findings about every AI system. They were made before the experiments; the experiments ask whether they have observable consequences.

  • The model proposes. It never writes authoritative state directly.
  • Belief alone cannot grant permission. Confidence, however often it has been right, does not open a gate.
  • What the organisation has decided lives apart from what any agent believes, and is changed by people.
  • What actually happened is recorded outside the model, and returns to it only as evidence.

Each placement, and what runs there today →

How we test it

Hold one side still. Attack the other.

We build worlds where the placement is an experimental question rather than an assumption: a settlement whose history lives in an authoritative ledger, inhabited by minds with their own fallible beliefs. Then we intervene. Holding the minds fixed, we degrade the institution’s mechanisms and watch which failures appear. Holding the institution fixed, we ablate, reset, replace and corrupt the minds and watch which properties move.

Predictions are registered before the run, contaminated collections are voided rather than rescued, and the misses are published at the same volume as the wins. The settlement runs in public →

What the evidence permits

So far

Reliability was separable from the model.

Five pre-declared properties — one accepted reality, typed refusal of invalid acts, duty recovery from institutional history, no duplicate accepted work, no false completions — held under every cognition-layer intervention we ran, each enforced outside the model. One world, one constitution, one comparator model: separable, not subordinate.

The result →

Trust decided admission; the ground decided consequence.

The same false claim, delivered identically, was admitted by a trusting actor and refused by a distrusting one. The world refused every act that followed from it. Positive trust is a threshold, not a truth filter.

The result →

Authority governed interpretation whether or not it was true.

A correspondence marked authoritative restored uninformed receivers to the informed baseline when it was correct, and redirected nearly every receiver, the informed partner included, when it was wrong. One carrier, one corpus, one frozen instruction pair.

The result →

Every question, its verdict and its state →

Not shown

What the evidence does not say

  • That the gates resist a determined adversary.

    The published interventions degrade, reset or replace cognition. None of them try to beat it.

  • That several grants of authority compose safely.

    Two individually sound grants can together satisfy a control that was meant to need two independent parties. That is a property of the set of grants, and it has not been tested in the organisational configuration we propose.

  • That placement makes a weak model good.

    Where a property lives decides what survives the model being wrong. It does not decide whether the model’s judgement, interpretation and work are useful. That remains a research concern, not a settled one.

  • That a persistent agent on our own engine earns its place.

    Whether it adds measurable value over the configurations that ship today is what the experiments have to establish. No experiment is registered.

Taniwha is an applied AI research lab. We investigate how artificial intelligence can remain useful inside systems where truth, authority and consequence exist independently of the model.