Load-Bearing Assumption
A judgement that determines a reported result while presenting as a fact, so the result inherits an uncertainty nobody sees.
Reviewed July 16, 2026Stable
When not to use it
- Where the assumption has been stress-tested and the result is insensitive, which is the case the concept exists to distinguish.
- As an accusation. An assumption doing heavy work is usually documented, reviewed and defensible where it was made.
- Where genuine measurement exists and is simply being ignored, which is a different failure.
Reach for something else instead
- Sensitivity reporting — publish the result under the plausible alternatives rather than only the chosen one.
- Stated counterfactual — name in advance the assumption change that would reverse the conclusion, which is the falsification discipline applied to estimates.
- -
Read more on the blog
- Every AI chip passes through one company's machinesThe concentration risk in AI hardware is usually discussed as a country. It is more precisely a single firm, and that firm has not priced like a monopolist.
- A $3.5bn guarantee book and a $250bn commitmentThe financing structure behind the AI buildout is disclosed, legal and defensible. It also routes several distinct-looking exposures to the same underlying variable.
- Same company, same year, revenue figures 63% apartThe most quoted numbers in AI are run-rates from private companies, reported by outlets triangulating from leaks, and sources covering the same period differ by multiples.
- Inference prices fell 9x a year. Also 900x.The most cited number in AI economics is a single rate. The study behind it reports a range spanning two orders of magnitude, and a caveat that undercuts its own fastest figures.
Further reading
- Raji et al. (2021), AI and the Everything in the Whole Wide World Benchmark — a measurement's operationalisation doing the work of the construct it stands for.
- Bouthillier et al. (2021), Accounting for Variance in Machine Learning Benchmarks — how analytic choices within a defensible range move reported conclusions.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading a footnoted estimate as a measured quantity because it appears in an audited or peer-reviewed document.
- Treating disagreement between competent parties as evidence that one is wrong, rather than as evidence that the range is wide.
- Accepting a disclosed policy range so wide it permits any assumption as though it were disclosure.
At a glance
FieldFoundations
Three shapesdistributional, temporal, contextual
Best signalcompetent parties disagreeing on the same evidence
Remedysensitivity or stress testing
Obstaclea point estimate is more persuasive than a range
DifficultyIntermediate
Where this sits
A starting point. Nothing needs to come before it.
0Levelstarting point
0Needs firstnothing
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works