INSIGHT PAPER · STRATEGY & DEFENCE AI · 2026

Governing defence AI.

A technical analysis and governance proposal for defence AI. The Consensus-M architecture is inspired by the Trias Politica.

Sacha Virgile Elouardi · Designer · Open-source AI research · Multi-agent architecture

The black box trap

The scenario raised by the French Ministry of the Armed Forces Red Team highlights a critical risk of modern warfare: dependence on a monolithic AI whose decision path is opaque. In a high-intensity combat environment, corrupted data caused by software, electronic warfare or systemic bias could lead an AI to make a catastrophic decision without any human operator understanding its root cause. Without traceability, command is left with an unacceptable explanatory gap, both operationally and legally.

The Asymmetric Multi-Agent Concept (AMAC)

The alternative is to reject centralised AI in favour of a separation-of-powers structure inspired directly by military doctrine and public law. This model relies on three autonomous, specialised agents that remain isolated from one another.

POWER OF ACTION

Operational AI

It performs a purely tactical analysis of the field. It maps the environment, identifies the forces present, including enemies, civilians and allies, and determines the optimal technical approach for the military objective.

POWER OF OVERSIGHT

Legal and Ethical AI

Fully separate, it receives the Operational AI plan and tests it against rules of engagement, International Humanitarian Law and mission history. It calculates an objective benefit-to-risk ratio that accounts for past failures and collateral damage.

REAL-TIME VALIDATION

Tactical Fire-Control AI

As the final link in the kinetic chain, it validates the order immediately before effect, continuously comparing the first two AI analyses with the dynamic situation in the fraction of a second before action.

The Dissensus Score

The technological core of the architecture is divergence detection. Instead of seeking forced consensus, the system measures a mathematical index of decision variance in real time: the Dissensus Score, written Sd.

Sd > τ

If the Operational AI and Legal AI analyses diverge beyond the tolerable critical threshold because of an anomaly or corrupted data, the architecture triggers a Fail-Safe. Autonomy is suspended and the decision is transferred to human command, following the Human-in-the-loop principle.

Designer context

This proposal forms part of an active and pragmatic research process. I am currently developing an artificial intelligence architecture based on an open-source model. Experience gained through open code and hands-on work with language models has convinced me that code transparency, combined with rigorous multi-agent engineering, is key to reducing the black-box opacity that legitimately concerns the defence industry.

Sacha Virgile Elouardi · Designer · Open-source AI research · Multi-agent architecture.

▶

VIDEO COMING SOON

An accessible presentation of this analysis will be published on YouTube soon.

Does this analysis raise questions for you? You can challenge it, expand it or open a collaboration.

DISCUSS THIS TOPIC