Actionable, Safe and Resilient AI: The Adarga Research Agenda

Frontier AI you can trust to get the mission done. We are developing technology that is actionable where the cost of being wrong is highest. We prove it by building, deploying and defending it in those environments.

We have passed an inflection point, Frontier AI is being applied to live operational use cases in Defence, Security and Resilience organisations around the world. The opportunities are varied and consequential, in intelligence analysis, planning and wargaming, command decision support, cyber defence, logistics and crisis response, AI provides the ability to increase the speed and quality of an outcome. This provides marginal, then complete, advantage over an adversary and safeguards our way of life.

However, the risks remain substantial in parallel. Risks within systems such as hallucinations from foundation models; agent violations of imposed constraints; or data poisoning are primary areas of concern for AI labs. This goes hand in hand with discussions around optimal human-AI teaming. Subsequential concerns her include dependencies on AI systems, the ability to operate in contested environments, and the ethical and legal frameworks necessary to hold others -and us-to account.

Against this backdrop, attitudes remain mixed. There are those who remain algorithm-adverse, distrusting of automation of any work that touches their role. There are those who are AI advocates but have a risk appetite that extends beyond AI’s current capabilities, and there are those who recognise the power of AI and agentic workflows but believe the financial costs and infrastructure required will prohibit deployment in a contested environment.

Adarga believes these risks are real and recognises there are many challenges to an AI enabled Defence, Security and Resilience sector. Nonetheless, we believe the progress is inevitable. It is our responsibility to ensure the organisations we support can take advantage of the opportunities on the table by deploying frontier AI in an actionable, safe and resilient way.

This applied research agenda presents the focus areas Adarga know are required to seize these opportunities and deliver AI enabled outcomes with justified confidence. It presents the open questions we seek to solve and invites industry partners, government and academia to collaborate with us to do so.

Actionable, Safe and Resilient

For a commander, analyst or operator to feel comfortable making a decision that may put the lives of others at risk, they require the utmost confidence in the systems that led to that decision. If they do not have this confidence then the systems will be circumvented, checked or there will be hesitation, in so doing removing the velocity advantage AI systems are supposed to bring.

Still, we must recognise it is not enough to strive for confidence alone. Confidence that is misplaced will lead to mistakes more damaging than lost time. These are the mistakes that capture the mind of society: AI that leads to outcomes which do not conform to the intent of their human teammates, information that gets lost, a picture that becomes incomprehensible to the human mind and people die. It is our role at Adarga therefore to ensure the confidence put in frontier AI is justified.

Justified confidence in actionable, safe and resilient AI systems will only be achieved if technology is developed in the right way, with the right questions in mind. Adarga focuses on two core pillars that sit at the heart of this balance and aims to answer the open questions below in order to address them.

Pillar 1: AI Assurance

AI Assurance should be framed not as a governance wrapper or compliance phase in development, but as the science of justified confidence in AI behaviour. It is the mechanism that allows the user to feel comfortable using a system, and others to hold them accountable. For advanced AI systems, this cannot come from benchmark scores alone. It requires a deeper solution to how models reason, how uncertainty should be expressed, how robustness degrades under changing conditions, and what forms of guarantee are genuinely supportable.

This includes mechanistic understanding of the internal processes that drive outputs at every level from circuit to system, the presentation methods to a human, principled estimation and propagation of uncertainty, robust detection of operation outside validated conditions, and rigorous evaluation methods that remain meaningful even when systems can adapt to oversight or benchmark structure. On top of this, we must evaluate novel architectures, methods and the processes for application in DSR. This comes together to allow humans to understand how an outcome was achieved and confidence in the decisions it supports.

Pillar 2: Adaptive Knowledge Representations

The world our customers operate in is dynamic, contested, misinformation ridden, and temporally uneven. Facts change, sources disagree observations arrive asynchronously and confidence in any given claim depends not only on the content, but on when its metadata and context: how it was captured, where it came from, and how it relates to other evidence. Ensuring the integrity of this knowledge remains critical to ensuring outcomes happen because of the right information. Providing transparency across this is the art of allowing a human to understand why an outcome was achieved.

Many current AI architectures remain weak on this point. Parametric knowledge becomes stale, vector retrieval often obscures provenance and temporality, whilst many agentic systems weight context incorrectly for a given task. Static knowledge stores struggle to update consistently under streaming inputs and within contested environments keeping track of which information is available and how this affects decision making capabilities is a challenge. It is true that context engineering, agentic integration across data architectures and the resurgence of semantic layers shows this problem is recognised, but it is far from solved.

This is not just a knowledge-engineering problem, it is a problem that dictates the success or failure of a solution. The quality, currency, and structure of available knowledge place hard limits on the reliability of downstream reasoning. If an AI system cannot reason over the right information at the right moment, then no amount of fluent generation or assurance makes it actionable.

These pillars must be addressed together: Assurance without adaptive knowledge produces guarantees detached from the evidential quality of the information being reasoned over and is inactionable. Adaptive knowledge without assurance produces systems that look informed but cannot justify confidence. The frontier is their intersection: systems that know what they know, know where that knowledge came from, know when it is no longer sufficient. Our solutions must produce calibrated, testable, decision-relevant confidence in their outputs.

Technical progress, on its own, is not sufficient. Assured AI only becomes real when the organisational and social structures around it (governance, workforce design, procurement norms, user education) are co-developed and can integrate AI. We aim to address this wherever we can, trough open discussion, advisory and market leading product design.

Collaborate with us

Adarga has spent a decade building AI for Defence, Security and Resilience organisations, nevertheless this agenda is too big for any single organisation, and deliberately so. We are actively seeking partnerships with academia, government laboratories, think tanks, standards bodies and industry peers. Whether this includes joint publications, shared benchmarks, visiting researcher arrangements, and co-supervised PhDs and postdocs. We publish this agenda to invite conversation, challenge and extension, not to close the space.

We are also hiring. If you want to do frontier science on AI assurance and knowledge representation problems, with the rare advantage of real users, real data and problems that genuinely matter, Adarga is built for you. If this is of interest to you send us an email. hello@adarga.ai