AI Explainer: Why We Start with Retrieval

People love to open up their AI models and compare the size of their parameters. It sounds impressive—until you ask what it actually does.

In nuclear, accuracy isn’t optional, and explainability isn’t a luxury. When your AI can’t tell you where it got an answer, it’s not helping you. It’s gambling with your credibility.

The Parameter Arms Race Misses the Point

First, let’s clarify what we even mean by “parameters” and the rest of this AI jargon.

Training a model means feeding it vast amounts of text to mimic how humans write and reason. Parameters are the model’s internal settings that help it recognize patterns and predict what comes next. More parameters usually mean more capacity, but not necessarily more understanding.

The resulting trained model is like long-term memory—useful, but prone to mixing fact with fiction. A foundation model can mimic the shape of a licensing argument. It can’t tell you whether it built its template from the right version of Regulatory Guide 1.206, or if it misused an AP1000 DC precedent for a TerraPower OL application. Scale alone doesn’t make a system nuclear-ready.

Retrieval, on the other hand, keeps the library open beside you. Instead of relying on what the model remembers, it pulls the answer from verified sources in real time. That’s what we mean by being grounded, where every is output traceable to where the fact came from.

Retrieval Comes First

That’s why Everstar’s approach starts with retrieval, not retraining. When Gordian answers a question, it pulls directly from curated regulatory requirements, guidance, and precedent. It’s not just using a frozen copy of them buried in model weights.

Every output shows its work. You can click any citation and see the exact sentence in context, the way nuclear reviewers expect to see it. That traceability is what turns “AI output” into something you can actually send to a regulator or stand behind in a public meeting.

It also makes Gordian faster to deploy. There’s no six-month training cycle or cybersecurity review for every new data ingest. You connect to your validated sources, and retrieval takes it from there.

Curated Knowledge, Not Crawled Data

Retrieval only works as well as the sources behind it—which is why we built a fact database worthy of a regulatory audit.

Everstar’s system is built on a curated, version-controlled fact database engineered by nuclear professionals. Each document is organized by schema, mapped to specially designed metadata, and tracked through revision history. Engineers can see not just what changed, but why.

That means when Gordian cites SRP 3.6.2, you know it’s the current section, and when it references a license amendment, you can see exactly which docket it comes from. The result is speed and confidence. The system retrieves the right resources, not just plausible ones.

A concrete example: in our Last Energy project, Gordian assembled a research package alongside the environmental assessment itself. It retrieved every applicable requirement, guidance, and precedent in context. Eight weeks of manual digging collapsed into one. No black box, no shortcuts, just faster clarity.

Train When It Adds Value

We’re not anti-training. We’re just disciplined about when it’s worth doing.

Retraining or fine-tuning a model on nuclear data can lock in obsolete guidance, leak proprietary or export-controlled information, and strip away the ability to see what’s happening inside. It produces static knowledge when nuclear work depends on constant change.

So our philosophy is simple: retrieval first, selective training later — only where it measurably improves accuracy or efficiency. Think of it as surgical, not systemic.

Scaling Securely

By separating knowledge from the model, we can scale without retraining. When the world’s large models get better at reasoning, summarizing, or language fluency, Gordian gets better too—instantly. No new risk reviews or reapproval cycles.

Security isn’t bolted on; it’s baked in. The data stays in your controlled environment. The model reads from it but never absorbs it. That separation is the difference between “AI-enabled” and “AI you can trust.”

What It Means in Practice

For utilities and vendors, the impact is straightforward:

• You get faster research and drafting cycles.
• You cut down on rework from uncited or outdated material.
• You can verify every statement before it leaves your team.

Retrieval isn’t the flashy part of AI, but it’s the reliable part. It’s what turns a language model into a licensing tool. When you pair that retrieval with the fluency and judgment of modern large models, you get something new: an assistant that actually understands both the words and the work.

The Proof Is in the Pilot

You don’t have to take our word for it. Run a side-by-side pilot and see how retrieval-first performance compares to anything that claims to be trained for nuclear. (Learn more about how to design a good pilot in this article, and reach out to hello@everstar.ai when you're ready.)

Gordian produces source-cited answers that stand up to expert review. That’s not the “right way” to do nuclear AI. It’s simply the credible way.

Believable, traceable, and built to stand up in front of a regulator—that’s what makes it worth trusting.