How a small, governed set of analytical building blocks can deliver the answers quality and manufacturing teams need—and return the same answer every time.

When "Good Enough" Isn't Good Enough

Picture a question that should be almost boring:

"How many deviations involving temperature excursions were there last week?"

If your data says 17, you expect the answer to be 17. Not 16. Not 19. Not "roughly 17." And you expect it to be 17 the next time someone asks, and the time after that as long as the underlying data and the definition of "temperature excursion" haven't changed.

This sounds obvious. But it's exactly where today's artificial intelligence (AI) analytics tools tend to fall apart in regulated environments.

Here's why: most AI analytics systems are built on generative agents — large language models (LLMs) that decide at runtime how to approach a question. Ask one "how many temperature excursions last week?" and it might query structured deviation codes. Ask it the same question five minutes later and it might take a completely different path such as performing a semantic search, expanding "temperature excursion" into related phrases, or improvising a new approach. Each path might be plausible. Each might even return 17.

But plausible is not the same as reproducible.¹ And in regulated manufacturing, reproducible is the whole point.

The Problem With Freedom

For exploratory research, AI flexibility is a feature. You want an assistant that can brainstorm, adapt, and delightfully surprise you. But for quality and manufacturing operations governed by the U.S. Food and Drug Administration (FDA), EU Good Manufacturing Practice (GMP), and other global regulations, that same flexibility becomes a liability.

Consider what happens when an auditor asks how you arrived at a number. If your AI took a different analytical path each time it ran, such as querying different fields, applying different definitions, or expanding terms differently, can you reproduce the result reliably? Can you show the exact steps from question to evidence? Can you prove the answer would be the same tomorrow?

In most agentic AI systems, the answer is no. The route from question to answer is invisible and unreproducible. The model chose its path at runtime, and there's no guarantee it would choose the same one again.

This is the challenge MasterControl set out to address with its research on deterministic analytics to deliver an approach that asks a fundamentally different question: What's the smallest set of governed analytical building blocks that can answer the questions quality and manufacturing teams actually need to answer?

Eight Building Blocks, Infinite Questions

Here's the core insight: most enterprise analytical questions, no matter how complex they seem, can be broken down into a small number of recurring operations. MasterControl's research identifies eight semantic primitives using building blocks that, when combined, can cover the vast majority of governed analytical tasks:

The power isn't in any single block. It's in composition the way these blocks chain together to answer increasingly sophisticated questions.

Think of it like a well-stocked kitchen. You don't need infinite ingredients to cook an enormous variety of meals. You need a finite set of high-quality ingredients and the recipes to combine them. MasterControl's eight primitives are the ingredients. The compositions are the recipes. And because each ingredient is governed—with defined input types, output types, and execution rules—every dish comes out the same way every time.

From Simple to Sophisticated: One Running Example

The easiest way to see this work is to keep asking harder questions about the same data. The questions get more sophisticated, but the execution model stays the same. What changes is the composition, not the foundation.

Notice what's happening: the vocabulary stays small—just eight operations—while the number of meaningful analytical programs grows. You can ask deeper questions without ever expanding into uncontrolled, unvalidated territory.

And critically, if a question can't be answered by any valid composition of these blocks, the system says so explicitly, rather than silently improvising an answer using a method no one validated.

Why Smaller Is Better in Regulated Environments

There's a common assumption that deterministic systems are just a safety-constrained approximation of more capable AI agents—that you trade power for control. MasterControl's research suggests this framing may be backwards for bounded enterprise analytics.

The argument is straightforward: a larger action space doesn't necessarily mean larger analytical coverage. Many of the different paths an agentic AI might take are just alternate ways to compute the same result. If your governed data says 17 temperature excursions, an agent that returns 17 after using five different tools isn't analytically more capable than a deterministic system that returns 17 through one governed path. The extra tools are implementation freedom, not additional analytical information.

What you get by keeping the analytical language small and governed are four advantages that matter enormously in regulated environments:

• Reproducibility. Same question + same governed meaning + same data snapshot + same procedure version = same result. Every time. The number is not permanently fixed—new deviations can arrive and definitions can be revised—but the guarantee is narrower and more useful: the answer is stable as long as the inputs are stable.

• Traceability. The analytical path itself becomes part of the governed product. You can show an auditor exactly how the system got from question to evidence.

• Testability. A procedure can be versioned. Inputs and outputs can be typed. Regression suites can assert that a known question over a known snapshot still returns the expected result.

• Change control. When a new analytical capability is needed, it becomes a controlled design decision that’s validated and registered, rather than an invisible runtime event where the model decides to try something new.

For regulated quality and manufacturing operations, the objective isn't merely to produce a persuasive answer. It's to make the route from question to evidence inspectable enough that the answer can be reproduced and challenged.

Knowing Where the Boundary Is

An honest system knows what it can't do. MasterControl's “Seventeen Every Time” deterministic approach doesn't pretend to cover tasks outside its contract.

An open-ended agent can browse the web, invent a new analysis, write code, call an unfamiliar service, or take an operational action. Those capabilities are outside the completeness claim and that boundary is a feature, not a limitation. When a question requires a method not represented by a governed composition, the system identifies the gap explicitly. The organization can then validate and register a new capability if it belongs in the analytical domain.

Expansion becomes a deliberate, documented design decision, not something that happens invisibly when a model decides to improvise.

What Could Go Wrong (and How to Know)

MasterControl's research is also clear-eyed about completeness and its failure modes. There are three distinct ways a system can fall short, and conflating them leads to wrong conclusions:

1. Semantic mapping completeness: Did the system correctly map the user's language to the governed analytical meaning? "Temperature excursion" must resolve to the approved definition, not an improvised synonym list.

2. Expressive completeness: Given the correct meaning, can the primitive language represent the required analysis?

3. Data completeness: Does the governed data actually contain the records and fields required to answer the question?

These failures are independent. A perfect composition can fail because required data is missing. A classifier can misinterpret a question even when the primitive language is complete. Neither disproves expressive completeness, and understanding the difference matters when you're diagnosing why an answer came out wrong.

How to Put the Thesis to the Test

This isn't just theory. MasterControl outlines a concrete way to test whether deterministic analytics can match agentic approaches in practice:

1. Build a representative corpus of real analytical questions from quality and manufacturing users.

2. For each, establish a governed reference interpretation and expected result.

3. Compare an agent allowed to select and compose tools dynamically against a deterministic compiler restricted to the eight-primitive basis.

4. Measure coverage, correctness, reproducibility across repeated runs, trace equivalence, failure explicitness, and governance cost.

If deterministic coverage approaches agentic coverage while reproducibility and governance improve materially, the thesis is supported: you can get the analytical power you need with a much smaller, more controllable behavioral surface.

The Bottom Line

The usual framing presents deterministic systems as a safety-constrained approximation of more capable agents. For bounded enterprise analytics, that framing may be backwards.

If a small, closed language spans the useful analytical question space, then much of an agent's additional freedom isn't additional analytical power. It's freedom to choose among alternate execution paths that produce the same answer but can't be reproduced, audited, or trusted in a regulated context.

For the user asking how many temperature excursions occurred last week, the best system isn't the one capable of inventing the most ways to count them. It's the one that understands the governed meaning, executes the appropriate analysis, produces the supporting evidence and, given the same inputs, if the answer should be 17, the system returns 17 every time .

That's the deterministic analytics advantage: not less intelligence, but intelligence you can trust, reproduce, and defend.

To learn more about MasterControl's “Seventeen Every Time” research report  on governed, deterministic AI for life sciences manufacturing, explore our work at MasterControl AI/ML on Hugging Face [JJ3] [RV4] or contact us to discuss how structured analytical approaches can support your digital transformation journey.

What You Should Do Next

• Subscribe to the GxP Lifeline newsletter to access other informative articles like this.

• Contact us to get answers for your software, quality, and compliance questions.

Sources

1. “From Question to Evidence: A Small Analytical Algebra for Governed Data Analysis,” MasterControl white paper by Matt Bray, Logan Green, Hemant Jomraj, Shardul Pande, and Viktoria Rojkova, Sept. 2026.