Why Hallucination Is a Selection Error, Not an AI Flaw

Most conversations about AI hallucination treat it like a fundamental defect — something inherent to how language models work that you either live with or try to suppress. But what if the framing itself is wrong?

Wiley Jones, CEO and co-founder of Doss — the company building an AI-native operations platform for mid-market physical operations companies — has a different take. His background isn’t in machine learning. It’s in electrical engineering, mechanical engineering, and control systems. And from that vantage point, hallucination looks like something engineers solved decades ago.

The reframe: discrete outputs change everything

The core insight is deceptively simple. When you let a language model produce any output from an enormous continuous space, a wrong answer feels like the model is “making things up.” But if you constrain the output to a finite set of discrete options, the same wrong answer is just a selection error — the model picked the wrong item from a known list.

“Hallucination only matters on the output itself,” Jones explains. “So if you bound the output into discrete outputs, and it’s not a continuous spectrum, then hallucination is actually just a selection error.”

That’s not a philosophical distinction. It’s an architectural one. A selection error has a feedback loop. You know what the correct options were. You can measure the deviation. You can feed the error back into the system and improve.

What 200-year-old math already knew

Jones points out that the concept of creating bounded outputs from unbounded inputs isn’t new — it’s the foundation of control theory, a discipline with roots going back two centuries.

“It’s been really funny watching software engineers re-derive control theory from first principles,” he says. “They’re like, ‘oh, what if we took the errors and fed them back in?’ And I’m like, yes, we discovered this 60 years ago. 200 years ago, the Russian scientists figured out stable control systems and robustness and how you can create bounded outputs on unbounded inputs.”

The practical implication: instead of treating each AI agent as an open-ended generator, you design systems with what Jones calls “the sandwich theory of determinism” — specific concrete inputs, specific concrete outputs, and freedom in the middle. You optimize the middle layer for speed and cost, not correctness. Correctness lives at the boundaries.

Competing heuristics and convergence

Where this gets interesting for production systems is when you layer multiple bounded objectives against each other. Jones describes setting up competing optimizations — one heuristic targeting a 78% accuracy baseline that the system continuously pushes upward, while another demands resolution time under one minute.

“You start setting these competing optimizations against one another. And I think this system will eventually converge towards governed optimal behavior,” he says. “And again, this is control theory. Like this is how controls work.”

This is the part most AI teams skip. They focus on making individual model calls more accurate rather than designing the system-level feedback architecture that makes accuracy a convergent property. The model doesn’t need to be perfect on any single call — it needs to be inside a system that corrects over time.

Why this matters for anyone building agent systems

The takeaway isn’t “hallucination is fine.” It’s that how you frame the problem determines what solutions are available. If hallucination is an inherent model flaw, your options are limited to better prompting, fine-tuning, or retrieval augmentation. If hallucination is a selection error in a system with insufficient output constraints, your options expand to the entire toolkit of control theory — feedback loops, bounded state spaces, competing heuristics, and convergence proofs.

Jones argues this isn’t novel engineering. It’s forgotten engineering. The math exists. The frameworks exist. Software engineers building agent systems just haven’t looked in the right textbooks yet.

FAQ

How can you reduce AI hallucination in production systems?

Constrain outputs to discrete, finite options instead of allowing open-ended generation. When the model picks from a bounded set, wrong answers become measurable selection errors with clear feedback loops — not mysterious hallucinations. This approach borrows from control theory’s principle of creating bounded outputs from unbounded inputs.

What is Doss and what does it do?

Doss builds an AI-native operations cloud called ARP (Adaptive Resource Platform) for mid-market physical operations companies with $20M-$250M revenue. It replaces fragmented ERP stacks — procurement, inventory, orders, finance — with a single composable system that implements in 3-4 months versus 12-24 months for SAP or Oracle.

How does Doss handle errors in its AI-native operations platform?

Doss uses control theory principles to create bounded outputs at every decision point. The system sets competing heuristics — accuracy targets, resolution time limits — and converges toward governed optimal behavior through feedback loops. When the system touches high-risk paths like order processing, it flags changes for human review automatically.

What industries does Doss serve?

Physical operations companies — retailers, food and beverage brands, consumer goods manufacturers — that manage inventory, procurement, orders, and multiple sales channels. Customers include Verve Coffee, Eight Sleep, and Mezcla. The sweet spot is mid-market companies too complex for spreadsheets but growing too fast for 12-24 month SAP implementations.

How long does a Doss implementation take compared to traditional ERP?

Three to four months versus 12-24 months for enterprise ERP systems like SAP or Oracle. The technical configuration is fast — tailored demos take about two hours. The remaining time is human coordination: getting organizations to decide how they want to operate and aligning stakeholders on business processes.

What is the sandwich theory of determinism in AI?

Specific concrete inputs on one end, specific concrete outputs on the other, and freedom for the AI system to work however it needs in the middle. You optimize the middle for speed and cost, not correctness. Correctness is enforced at the boundaries through output constraints and feedback loops. Wiley Jones uses this to build deterministic behavior from probabilistic models.

How does control theory apply to building AI agents?

Control theory creates stable, bounded outputs from unbounded inputs using feedback loops — a discipline over 200 years old. Applied to AI agents, it means designing systems where errors are fed back as corrections, outputs are constrained to finite sets, and competing heuristics converge toward optimal behavior. The math for multi-agent coordination already exists.

What is a selection error in AI?

When a model’s output is constrained to a finite set of discrete options, a wrong answer is a selection error — the model picked the wrong item from a known list. Unlike open-ended hallucination, selection errors are measurable, have clear feedback loops, and can be corrected probabilistically. This reframe makes AI reliability an engineering problem, not a research problem.

Can you build deterministic AI systems from probabilistic models?

Yes, by reducing the output dimensionality to truly discrete options. Wiley Jones argues hallucinations “only feel like a hallucination because the possible set of outputs is so large that it feels like it’s just making things up.” Reduce that dimensional scale, and wrong outputs become basic probabilistic errors with well-understood correction mechanisms.