The architecture

What we build — and how far along each part is.

One thesis across three planes: a mind that persists, a ground that can’t lie, and rules that hold outside the model. Every part carries an honest status — shipped, in closed alpha, in active development, demonstrated, or researched.

Long-horizon agents

Mind

Persistent cognition

Beliefs, memory, trust, and decision state for agents that develop over time.

Taniwha Engine

Active development

Current generation in active development; a first-generation research line still powers Briarwatch.

  • Per-agent belief graphs with provenance
  • Memory, consolidation, and forgetting
  • Trust that builds slowly and collapses fast

Ground

Checkable world-state

Deterministic worlds an agent can observe, question, and be wrong about.

Kano

Active development

Reproducible per engine version — no unseeded randomness, no mutable world database.

  • Whole worlds grown from a single seed
  • Causal — each layer conditioned on the one beneath
  • Queryable by AI over a tool protocol

Assurance

Rules outside the model

Enforce policy on agent actions and keep an audit trail — independent of any one model.

Ārai

Shipped

Open source (Apache-2.0 / MIT), runs locally, no hosted service required.

  • Pre-flight gate on agent tool calls
  • Enforced, advisory, or observed per integration
  • Hash-chained local audit log

built on ↑

Kete

Closed alpha

The organisation platform built around the Ārai core.

  • Org policy across agents, classed per tool
  • Audit retention, query, and export
  • Structural map of a codebase

Demonstrations

Public proof, labelled by engine generation

Briarwatch

Research

First-generation research line

Long-run social-cognition simulation and the validation environment for the first-generation engine.

Copperhollow

Demonstration

Current generation

Current-generation minds on deterministic ground — a mining camp with a reason to exist.

Research

Result within scope; value still under test

Codeworld

Research

A second implementation of the deterministic-ground thesis — instrument validated within scope.

  • Label everything by what it is — shipped, alpha, demonstrated, or researched.
  • Every public value carries its evidence class — fact, derived, staged, replayed.
  • Minds, ground, and rules are separate planes, not one dependency stack.

About this view

A curated overview

This is a public-facing view of our development surface — what we’re building and roughly how far along each part is. It’s a snapshot we curate and refresh as things mature, rather than a live feed.

It’s meant to show the shape of the work at a glance, not the granular state of any individual component. For the detail behind a given part, follow it through to its own page.