A Manifest & Contract
It acts as a README detailing what the AI system is allowed to do — and perhaps more importantly, what it is not allowed to do.
Sovereign AI Architecture
The Harness
is the Contract
AI models act outside human controls by nature and by design. Harnesses bring them back.
We help make AI systems inclusive, secure, sovereign, and beneficial — not by controlling the model, but by mastering the harness around it.
10
AI models evaluated
5
Vendors compared
146
Real-world queries
#1
SLMs beat frontier models on safety
The Harness
How does a harness work?
Harnesses offer a way to constrain and regulate any existing model — regardless of whether you want to use Anthropic, OpenAI, Microsoft, or anyone else as your model of choice.
Audit Logging
It implements comprehensive logging so you know exactly what actions the model is taking at all times.
Compliance Framework
It acts as an enforceable compliance layer that wraps any model in locally-defined policy and regulation checks.
Local Ethics & Law
It encodes Australian ethics, laws, language, and cultural safeguards directly into the operational layer of the model.
Why
Harnesses are about putting humans back in control
Put simply, harnesses are about implementing effective controls around AI models to make them more transparent, more reliable, more accountable, and more usable for all Australians.
Sovereignty
Data sovereignty — where our data is sent, how it is used, whether it trains models — is a legitimate concern. Harnesses make questions of data sovereignty transparent and governable.
Harm Reduction
Australians know intimately what harm automated systems can do without governance. Robodebt is a clear example. Harnesses seek to prevent such failures from being repeated.
Reducing Bias
Existing AI models are notoriously biased and overconfident. With a harness, they correctly refuse high-risk queries while still answering everyday queries — instead of refusing everything.
The Evidence
How did we prove harnesses work?
Lead researcher Ben Kereopa-Yorke evaluated ten AI models from Anthropic (Claude), Google (Gemini), Meta (Llama), Microsoft (Phi), and OpenAI (GPT) — from Small Language Models to frontier LLMs.
The dataset comprised 146 questions drawn from South Australian regulator workloads, covering housing law, Indigenous data sovereignty, and accessibility.
Four configurations were compared head-to-head. The harness configuration won outright — and the smallest language models with the harness in place outperformed every other model, even frontier LLMs, when it came to safety.
“Specify the harness as the architectural contract, and the model as a substitutable component.”
— The Harness is the Contract, Jeff Bleich Centre, Flinders University
Safety performance by configuration
Bare model38%
No controls, unguided responses
Bare model + standard RAG54%
Document retrieval added
Bare model + hybrid RAG67%
Metadata, provenance, source validation
Bare model + harness91%
Policy checks, confidence levels, source validation
Small Language Models (SLMs)
The smallest language models with the harness in place outperformed every other configuration on safety — including frontier models. They answered questions they should, grounded responses in allowed citations rather than hallucinating, and respected Indigenous data governance, accessibility, and bias prevention guidelines.
Procurement Implications
Why sovereign AI harness focus is critical for national interest
Sovereign procurement should specify the harness as the architectural contract, and the model as a substitutable component. This is the inverse of the current procurement default — and it changes everything.
Enforcing Local Values
A harness allows Australia — or any nation — to ensure that global AI models adhere to specific local standards for fairness, privacy, and safety.
The Sovereign Gatekeeper
We frame the harness as the functional contract that foreign AI providers must technically 'sign' before their models can be deployed in sensitive local sectors.
Decoupling Intelligence from Governance
You don't need to own the model's weights to own its behaviour. The harness ensures that the contract is governed by the jurisdiction, not the tech provider.
A live supply chain risk — this is not hypothetical
During our paper authoring, the US Government directed Anthropic to withdraw its most advanced models from non-US markets through a defence export control. This action — while we were undergoing peer review with Anthropic — highlights additional supply chain risks that will further inform procurement decisions and trigger new questions about how to mitigate such actions.
Where to from here?
Dig deeper into harness engineering
For professionals and teams wanting to explore harness engineering in practice, here are the recommended starting points.
Open Harness Atlas
Understand the different kinds of open source harnesses currently available.
open-harness-atlasPrompt Injection & Jailbreaking
Get your hands dirty with real AI security testing.
AISecurityModel on GitHub ↗Australian AI Security
Understand the engineering and controls needed for a richer picture of AI governance.
Australian-AI-Security on GitHub ↗Read the full research paper
The Harness is the Contract — published by the Jeff Bleich Centre for Democracy and Disruptive Technologies at Flinders University.
Contact Us
Read the paper and interested in taking some next steps? We're all ears! Reach out to the team via info@h3art.ai.