The Difference Between Knowing More and Deciding Better
Why Prediction, Simulation, and Uncertainty Quantification Must Ultimately Answer To Consequence.
Beyond the Tools We Use
Over the three years building Chateauz, I've become increasingly interested in how we categorize emerging technologies. We often define them by the tools they employ rather than the problems they are designed to solve.
When a system analyzes data, we call it predictive analytics. When it represents a physical environment, we call it a digital twin. When it explores scenarios, we call it simulation.
These classifications are useful, but they can also obscure an important distinction: a technology's capabilities are not necessarily its purpose.
That distinction has shaped my thinking about the relationship between prediction, simulation, uncertainty, and decision-making in critical physical systems.
We have unprecedented computational power and access to data. We can forecast demand, simulate infrastructure performance, and evaluate thousands of potential futures. Yet organizations continue to make consequential capital commitments under profound uncertainty.
There is an assumption embedded in much of our enthusiasm for computational intelligence: that more information, more sophisticated models, and more simulations must necessarily lead to better decisions.
But information has value only in relation to what it enables us to understand, evaluate, or decide. Greater analytical precision does not automatically translate into better judgment, particularly when the consequences of a decision extend far beyond the boundaries of a model.
Perhaps we've confused the ability to model a system with the ability to make a defensible decision about it.
1. Prediction Is Not a Decision
Predictive analytics asks what is likely to happen. Through statistical inference, historical observations, and mathematical modeling, we can estimate future conditions with increasing sophistication.
For an energy infrastructure project, we might forecast electricity demand, generation capacity, or long-term operating costs. These are valuable inputs, but an accurate forecast does not independently establish whether a multibillion-dollar investment is justified.
What if interconnection is delayed, equipment costs escalate, or several individually manageable risks materialize together? What if a less probable outcome carries disproportionately greater consequences?
The most likely future is not necessarily the future against which an investment must be defended. A decision must account not only for expected performance, but also for the consequences of being wrong.
Prediction estimates what may happen. A defensible decision must account for what happens when the prediction fails.
2. Simulation Is Not Confidence
Simulation allows us to investigate how physical systems behave under different conditions. Digital twins can extend these capabilities by connecting physical assets with evolving digital representations, supporting monitoring, analysis, and operational decisions.
These are essential engineering disciplines. Their contribution to scientific discovery and infrastructure development is substantial. Yet the sophistication of a simulation does not, by itself, establish the reliability of the decision it informs.
A model can produce precise results while remaining sensitive to poorly understood assumptions. Thousands of scenarios can reveal countless possible outcomes without identifying which uncertainties materially affect a capital commitment.
Consider two scenarios. In the first, significant changes in modeled conditions produce different performance estimates, yet the preferred investment remains economically viable. In the second, a relatively small change in one assumption makes the investment unacceptable.
The second scenario may be far more consequential, despite involving less variability. The distinction lies not in the sophistication of the simulation but in the relevance of its findings to the decision.
The most valuable result of a simulation may not be the outcome it predicts, but the assumption it reveals we cannot afford to get wrong.
3. Not All Uncertainty Is Created Equal
In computational science, uncertainty is commonly distinguished between two categories: aleatoric and epistemic uncertainty, as recognized in established engineering and modeling standards, including NASA's Handbook for Models and Simulations (Ref 1).
Aleatoric uncertainty reflects inherent variability within a system, while epistemic uncertainty arises from incomplete knowledge, including uncertain parameters, insufficient observations, and limitations in our models.
This distinction matters because different uncertainties call for different responses. Some require resilience and adaptive design; others may be reduced through additional evidence or improved understanding.
Research by Chateauz Scientific Advisor Dr. Rylan Spence and his collaborators contributes to the broader scientific literature on inference and uncertainty in complex dynamical systems (Refs 2–3).
For critical physical infrastructure, the more consequential question is how uncertainty affects decisions involving capital, operational reliability, and long-term performance.
The magnitude of uncertainty and the consequence of uncertainty are not the same thing.
Even when additional information can reduce uncertainty, it does not automatically follow that acquiring it is worthwhile.
Before accepting uncertainty as unavoidable, have we established whether it is genuinely irreducible—or simply unresolved?
Even when additional information can reduce uncertainty, however, it does not automatically follow that acquiring it is worthwhile.
4. Not All Uncertainty Deserves Investigation
Imagine an infrastructure developer evaluating a new energy campus. The project faces uncertainty across grid interconnection, equipment performance, construction schedules, regulatory approvals, and future demand.
Should the developer devote equal resources to investigating every variable? Should the objective be to reduce uncertainty wherever possible, regardless of cost?
Decision theory has long examined the value of information: whether acquiring additional evidence is worth its cost, given its potential to improve a decision.
Suppose one investigation costs $500,000 and significantly improves forecast precision but is unlikely to alter the investment decision. Another costs $50,000 and could reveal whether a critical constraint makes the entire project economically unviable.
The second investigation may have greater decision value, even if it produces far less information. What matters is not simply how much uncertainty can be reduced, but whether reducing it could change the preferred course of action and justify the cost of acquiring that knowledge.
For critical physical infrastructure, the more consequential question is how uncertainty affects decisions involving capital, operational reliability, and long-term performance.
The magnitude of uncertainty and the consequence of uncertainty are not the same thing. Before accepting uncertainty as unavoidable, have we established whether it is genuinely irreducible—or simply unresolved?
Even when additional information can reduce uncertainty, however, it does not automatically follow that acquiring it is worthwhile.
5. From Modeling Systems to Defending Decisions
Predictive analytics, simulation, and uncertainty quantification are complementary capabilities that contribute to a larger process of evaluating choices.
Established fields such as decision analysis, robust optimization, and decision-making under deep uncertainty already recognize the importance of evaluating alternatives across uncertain futures.
The opportunity we see is to make that decision orientation foundational to the architecture of computational systems themselves.
At Chateauz, our work on outcome-conditioned simulation begins with the decision being considered and the outcomes that matter, using those requirements to guide what evidence, modeling, and uncertainty analysis should be prioritized.
It is a shift in organizing principle: from generating analytical outputs and subsequently interpreting their relevance, toward designing the analytical inquiry around the decision and its consequences.
We believe this distinction becomes especially important in critical physical systems, where the cost of a decision cannot be measured solely by predictive error or simulation accuracy.
The question is not whether a system can be modeled with sufficient detail. It is whether the evidence can establish which decisions remain defensible when assumptions, constraints, and conditions change.
This is the broader purpose of decision infrastructure: connecting computational evidence, uncertainty, and system behavior to consequential human decisions while preserving accountability for the outcomes.
Before Decisions Become Reality
The physical world is unforgiving of assumptions that remain unexamined. Capital cannot always be recovered, infrastructure cannot always be reconfigured, and the consequences of decisions can persist for decades.
We cannot eliminate every uncertainty or guarantee that reality will behave as predicted. But we can become more deliberate about which questions deserve investigation, which evidence is worth acquiring, and which assumptions are too consequential to leave unchallenged.
For decision-critical physical systems, greater analytical capability should not be an end in itself. Its value lies in the quality of the decisions it helps inform.
We do not need another prediction of the most likely future. We need to know which decisions still hold when the future refuses to cooperate.
References
NASA (2026). NASA-HDBK-7009B: NASA Handbook for Models and Simulations—An Implementation Guide for NASA-STD-7009B. NASA Office of the Chief Engineer. https://standards.nasa.gov/standard/NASA/NASA-HDBK-7009
del-Castillo-Negrete, C., Spence, R., Butler, T., & Dawson, C. (2025). Sequential Maximal Updated Density Parameter Estimation for Dynamical Systems With Parameter Drift. International Journal for Numerical Methods in Engineering, 126(3), e7618. https://doi.org/10.1002/nme.7618
Spence, R., Butler, T., & Dawson, C. (2026). Variational Data-Consistent Assimilation. Computer Methods in Applied Mechanics and Engineering, 453, 118804. https://doi.org/10.1016/j.cma.2026.118804
