What Is Tolerable Uncertainty?

The objective of consequential decision-making is not to eliminate uncertainty.

It is to determine whether what remains is tolerable.

DECISIONABLE ⇔ Uᴿ ≤ τ(O)

Every consequential decision contains uncertainty. The mistake is assuming the objective is to eliminate it.

It isn’t.

The objective is to understand what uncertainty matters, how it propagates, what can be reduced, and whether what remains is acceptable against the outcome we care about.

The Fantasy of Certainty

We have spent decades building increasingly sophisticated systems for forecasting, simulation, optimization and now AI. With each generation comes an implicit promise: perhaps with enough data, compute and modeling, we can finally know. But physical systems do not become deterministic because our models improve.

Power prices move. Permits slip. Equipment fails. Demand shifts. Capital costs change. Variables that appeared independent turn out to be coupled.

So the question cannot simply be:

How do we eliminate uncertainty?

A better question is:

How much uncertainty can this decision tolerate?

The Largest Uncertainty Is Not Necessarily the Most Important

Imagine a decision involving power, capacity, capital, time, performance and risk.

Each contains uncertainty.

One variable may have a 20% uncertainty range. Another only 3%.

It is tempting to assume the 20% uncertainty deserves the most attention.

Not necessarily.

If the outcome barely moves when the first variable changes, that uncertainty may be immaterial.

If a 3% change in another variable pushes the project across a financing threshold or destroys the economics, that smaller uncertainty matters far more.

Conceptually:

Uᴰ = Σ wᵢuᵢsᵢ

Where Uᴰ is decision-material uncertainty, uᵢ is uncertainty in variable i, sᵢ is outcome sensitivity, and wᵢ reflects consequence or economic weighting.

The point is simple:

Uncertainty has to be evaluated against consequence.

The uncertainty that matters is the uncertainty capable of changing the decision.

But Variables Don’t Move Alone

Real systems are coupled. Power affects capacity. Capacity affects revenue. Permitting affects time. Time affects capital. Performance affects operating cost.

So evaluating each variable independently is insufficient.

A more realistic representation is:

Uᴰ = Σ wᵢuᵢsᵢ + Σ γᵢⱼ(uᵢ,uⱼ)

The second term captures interaction effects. A permitting delay alone may be manageable. Power uncertainty alone may be manageable. Higher capital costs alone may be manageable. But when those uncertainties interact, the decision can change completely.

The risk may not reside in any individual variable.

It may reside in the relationship between them.

What Remains

Some uncertainty can be reduced through better information, modeling, measurement, experimentation or additional analysis. Some uncertainty that appears irreducible may simply reflect relationships we have not yet modeled.

Eventually, however, we arrive at residual uncertainty:

Uᴿ = Uᴰ − ΔU

This is not meant as literal scalar arithmetic in every system. It is a conceptual progression.

We identify what matters. We reduce what can reasonably be reduced.

Then we examine what remains. And that brings us to the actual question:

Can we tolerate what we still don’t know?

Tolerance Depends on the Outcome

There is no universal acceptable uncertainty threshold. The same uncertainty may be tolerable under one objective and unacceptable under another.

τ(capital preservation) ≠ τ(time-to-power) ≠ τ(revenue) ≠ τ(operational continuity)

This is why uncertainty without an outcome is incomplete. We should not merely ask:

How uncertain is this?

We should ask:

How does this uncertainty affect the outcome we are trying to protect?

That is the difference between uncertainty as a statistical property and uncertainty as a decision variable.

From Uncertainty to Decisionability

This brings us back to:

DECISIONABLE ⇔ Uᴿ ≤ τ(O)

A decision becomes decisionable when the remaining decision-material uncertainty falls within the tolerance established for the relevant outcome. That does not mean uncertainty has disappeared. It does not mean the model knows the future. It does not mean there is one correct answer. It means the uncertainty that remains is understood well enough for someone accountable for the consequence to decide whether they are willing to carry it.

The Human Has to Decide

A model can quantify uncertainty. A simulation can expose possible futures. AI can identify relationships humans may miss. Well-built decision infrastructure can trace how uncertainty propagates and show which uncertainties are material against an economic or operational outcome. But none of those systems should determine how much residual uncertainty an accountable operator is willing to carry.

That is judgment.

That is governance.

That is risk appetite.

That is accountability.

The human stays in the loop because tolerance is ultimately a human and institutional decision.

The goal is not certainty.

It is to understand the decision well enough to know what risk you are actually taking.

And then someone has to answer the question no model can answer for them:

Can we live with what remains?

That is tolerable uncertainty.

And that is what makes a consequential decision defensible.

---

Definition: