Introduction

In June 2026, the IDEaS program issued a challenge call that should make every defence AI company pay attention: "Reliable AI sensor fusion for real world missions" — seeking solutions that embed *compliance-by-design* into multi-sensor, multi-domain fusion workflows.

The message is clear: the Canadian Armed Forces don't just want AI that works. They want AI that knows when it doesn't know.

This is the problem of uncertainty quantification — and it's harder than most people realise.

The Confidence Problem

Every sensor is imperfect. Radar returns clutter. EO/IR cameras degrade in poor weather. ADS-B signals can be spoofed. When you fuse data from multiple imperfect sources, the errors don't cancel out — they compound.

Most AI fusion systems produce a single output: "This is a target at position X, moving at velocity Y." What they don't tell you is how confident they are in that assessment. A track with 95% confidence demands a different response than one with 55%.

In commercial applications, low confidence is an inconvenience. In defence, it's a safety-critical failure mode.

What Is Uncertainty Quantification?

Uncertainty quantification (UQ) is the science of measuring what your model doesn't know. In the context of multi-modal sensor fusion, there are two distinct types:

Aleatoric Uncertainty — Noise in the Data

Aleatoric uncertainty is inherent to the sensors themselves. A radar operating in heavy clutter has higher aleatoric uncertainty than one in clear conditions. A camera at night has higher aleatoric uncertainty than one at noon.

This type of uncertainty *cannot be reduced* by collecting more data — it's a property of the sensing environment. The best you can do is measure it and propagate it through your fusion pipeline.

Epistemic Uncertainty — Gaps in the Model

Epistemic uncertainty comes from the model's lack of knowledge. When a neural network encounters a scenario it wasn't trained on — a new vehicle type, an unusual flight profile, a novel jamming technique — its epistemic uncertainty spikes.

This type of uncertainty *can be reduced* with more training data, better architectures, or domain adaptation. But in defence, you're always operating at the edge of the known.

Why Compliance-by-Design Matters

The IDEaS challenge call specifically asks for "compliance-by-design" — not compliance as an afterthought. This reflects a broader shift in defence AI procurement:

1. **Auditability:** Every AI decision must be traceable. "Why did the system classify this as a threat?" needs a concrete answer.

2. **Safety cases:** Defence regulators require evidence that AI systems are safe to operate. Uncertainty bounds are a core part of any safety case.

3. **Human-on-the-loop:** CAF operators need to know when to trust the AI and when to override it. Confidence scores enable this decision.

4. **Allied interoperability:** When sharing intelligence with Five Eyes partners, confidence levels travel with the data. A track with 40% confidence means something different in a coalition context.

The NovaFuse Approach

Our fusion engine implements uncertainty quantification at every stage:

**Stage 1 — Sensor-Level Confidence:** Each sensor stream is assigned a real-time confidence score based on signal quality, environmental conditions, and historical reliability. A radar in clear weather gets a higher weight than one in clutter.

**Stage 2 — Fusion-Level Uncertainty:** When multiple sensors contribute to a single track, we propagate their individual uncertainties through the fusion process using Bayesian inference. The result is a combined track with a mathematically rigorous confidence bound.

**Stage 3 — Decision-Level Explainability:** Every output includes not just a classification and position, but a full uncertainty breakdown: "This track is 87% likely to be a fixed-wing aircraft, with a position uncertainty of ±12 metres. Confidence is limited by degraded EO/IR coverage in sector 4."

This three-stage approach means operators always know what the system knows — and what it doesn't.

The Force Multiplier Effect

Uncertainty quantification isn't just a safety feature. It's a force multiplier:

• **Fewer false alarms:** High-uncertainty tracks are flagged for human review rather than triggering automated responses. This reduces operator fatigue and prevents fratricide.
• **Better resource allocation:** When the system knows which sectors have low confidence, it can redirect sensors to fill the gaps. This is especially valuable in Arctic surveillance where sensor coverage is sparse.
• **Faster decision cycles:** Operators spend less time second-guessing the AI and more time acting on reliable information.
• **Coalition trust:** When sharing tracks with allies, confidence levels enable appropriate weighting. A 95% confidence track from a partner nation is treated differently than a 60% one.

Connecting the Threads

This post connects to several themes we've explored in this blog series:

• **Explainable AI (Blog #6):** Uncertainty quantification is a prerequisite for explainability. You can't explain a decision without understanding its confidence.
• **Edge AI (Blog #4):** Running UQ at the edge requires efficient Bayesian inference — a core NovaFuse capability.
• **Federated Learning (Blog #8):** When training models across allied networks, uncertainty quantification ensures that low-quality contributions don't degrade the shared model.
• **Digital Twins (Blog #11):** A digital twin that doesn't model uncertainty is just a pretty picture. UQ turns simulation into prediction.

Conclusion

The IDEaS challenge call for "compliance-by-design" is a signal that defence AI procurement is maturing. The era of black-box AI — "trust us, it works" — is ending. The next generation of defence AI will be built on transparency, auditability, and rigorous uncertainty management.

NovaFuse's approach to uncertainty quantification isn't a feature we added to meet a requirement. It's foundational to how we build fusion systems. Because in defence, the cost of being wrong isn't a bad recommendation — it's a life.

About NovaFuse: NovaFuse is an Ontario-based AI company specialising in multi-modal sensor fusion, federated learning, and edge AI for defence applications. We are an IDEaS CFP-006 applicant and active participant in the Canadian defence innovation ecosystem.

Related Reading: