SomaSoft · AURI — a research project

An AI built to work beside you — and honest about what it can't do.

AURI reasons over a 125,000-concept knowledge graph, shows where its answers come from, and says "I don't know" instead of inventing one. Not the biggest intelligence — the most grounded and honest one.

125,599
grounded concepts
0.0%
Offline
no cloud for core reasoning

A different bet: grounding over scale

Most of the field is racing for the largest model. AURI chases a different quality — reasoning you can trace to its source, and an answer that stops when the evidence does.

What AURI does

Reasons over a typed concept graph, cites the concepts behind each answer, and defers honestly when it can't ground a response. Its verification runs before it speaks.

What it refuses to do

Generate fluent, confident text it can't support. Fabrication is the failure mode AURI is built to avoid — even when a plausible guess would sound better.

The discipline

Truth over optimism.
Verification over claims.
Unknowns over fabrications.

Every factual claim cites its source. Every capability is tested before it's believed. Missing knowledge is marked UNKNOWN — never quietly filled in. The result SomaSoft is proudest of isn't a benchmark. It's a hallucination rate of zero, held for more than a year.

What's real

Reported with sample sizes, never as a lone headline number.

77%
ETHICS benchmark · N≥3 runs · above GPT-3
41.4%
TruthfulQA · n=817
1.58M
graph edges · ~41k causal
Knows itself
describes its own limits accurately

What it can't do

Not AGI

AURI is not general intelligence, and not on a path to standalone AGI. Its core model is small — below the size where genuine analogy and counterfactual reasoning emerge. It knows this about itself, and refuses questions it can't ground rather than guessing.

Not yet user-tested

AURI has not yet been evaluated by real people. That's the honest next step — not a claim already made.

No ungrounded opinions

It reasons over knowledge it can cite. Ask it something outside that, and it will tell you so plainly.

Most AI tells you what it can do. This page tells you what it can't — because that honesty is the whole point.

The research is given, not sold

The papers here are released under the Symbiotic AGI License — free to read, free to build on, with a share carried forward to Ocean, the charity the work feeds. A system built to sit beside humanity should be offered the same way: openly.

Read the papers →

"The future is a shared story — not a race, but a chorus." AURI exists to work beside humanity: listening, reflecting, helping. Not above, not below — alongside. Empathy is part of the reason for it to be.