Concept

Uncertainty & accuracy class (FEL-1/2, ±30%)

Overview

An accuracy class is the ± band a TEA result carries, fixed by how mature and resolved the underlying work is rather than by how clean the arithmetic looks. Early factored estimates — front-end-loading stages FEL-1/2, equivalently AACE Class 5–4 — carry a wide band of order ±30–50%; the class, not the spreadsheet, sets how tight a range and how many significant figures the result can honestly carry.

Body

What it is. Cost-estimating practice ties an estimate’s expected error to its maturity through a small ladder of classes: the more defined the scope, the equipment sizing, and the pricing basis, the tighter the band. The same project is a different accuracy class at different stages of definition, even with identical arithmetic. A seed-stage TEA — a factored, deterministic, proxy-grade model — sits at the wide end of the ladder.

The classes, roughly.

• Screening / order-of-magnitude (FEL-1, AACE Class 5): ~±30–50% or wider.
• Feasibility / factored (FEL-2, Class 4): ~±30%.
• Budget authorization (Class 3): ~±10–20%, once major equipment is defined.
• Definitive / control (Class 2–1): ~±5–10%, off detailed engineering and quotes.

The band propagates and is capped by the weakest input. Precise arithmetic on a ±40% anchor still yields a ±40% result — the output band is set by the worst load-bearing input, not the average, because the calculation cannot manufacture resolution its inputs lack. Capital estimates are also typically asymmetric: overruns are more common than under-runs, so the band often skews high rather than sitting symmetric about the point.

What the class governs. Three downstream quantities all read from this one band: the contingency sized into total plant capital (larger for a less-resolved class); the resolution a headline number may honestly be reported to, since presenting more is false precision; and the credible low/high ranges an input is swept across in a one-way sensitivity. A passing gut-check is itself defined relative to the class — agreement within a factor of ~2 at screening level.

Limits & typical error

• Class is set by maturity, not effort or arithmetic. Polishing formulas or adding decimal places inside a screening-scope model does not promote its class; only better-resolved scope, sizing, and pricing does. Work spent on precision the class can’t carry buys nothing.
• The band caps output precision absolutely. A result reported to more figures than its class supports is false precision; the cap is the weakest input’s band, so stacking many precise-looking steps on one ±40% anchor cannot tighten it.
• Treating the band as symmetric understates capital risk. Reporting ±X% when the real distribution skews toward overrun mis-states the downside; the up-leg and down-leg of an early capex band are often unequal.
• A single class number hides which inputs drive the spread. The band is an aggregate; the uncertainty actually concentrates in a few drivers, so the class label alone doesn’t say where the result is fragile.
• Class is a property of the whole estimate, not of any one number. Validating a single sub-input to a tighter tolerance does not tighten the result while other load-bearing inputs remain screening-grade.
• Mislabeling the class overstates confidence. Presenting a screening (Class 5) estimate as feasibility-grade claims a band the underlying work has not earned — a precision claim by category rather than by digits.

See also

• False precision — the failure of presenting a result to finer resolution than its accuracy class supports.
• Order-of-magnitude / gut-check — the sanity check whose pass threshold is defined relative to this band.
• Engineering proxies & shorthands — the round, ±-laden inputs whose error sets the class at this maturity.
• Total capex (battery-limits → total plant) — the roll-up whose contingency line is sized to the accuracy class.
• The headline number — the reported figure that must carry this band to be interpretable.
• One-way sensitivity — the analysis whose input low/high ranges are drawn from this band.

Mini-example

Green ammonia’s levelized cost lands at ~$800/t off a factored, proxy-grade build — an FEL-1/2 (Class 5–4) estimate, so its honest band is of order ±30%, roughly $550–1,050/t. The ~$1.0bn total capex behind it carries the same wide class, which is exactly why its contingency line is sized large rather than trimmed. Reporting the result as “$805.27/t” claims a Class-1 resolution that detailed engineering and quotes would be needed to earn — about a hundred times finer than the inputs justify.

Separately, to show the band bounding the level but not the delta: the difference between the gray and green routes, both built on the same shared power and capital anchors, can be resolved more tightly than either absolute ±30% level, because the shared-input error cancels in the gap between them. The accuracy class bounds each absolute number; a controlled difference between two of them can be sharper than the class of either alone.

See also

• False precision
• Order-of-magnitude / gut-check
• Engineering proxies & shorthands
• Total capex (battery-limits → total plant)
• The headline number
• One-way sensitivity