Measuring the Ruler

If you measure a table with a ruler you do not trust, you learn as much about the ruler as about the table. Nassim Taleb called this Wittgenstein’s ruler, drawing on Wittgenstein’s late remarks on doubt and certainty. A result flows in two directions. It informs you about the object and about the tool.

The direction depends on prior confidence. A good steel rule says the table is 1.2 metres long: the table learns. A stranger’s tape says the same table is four metres long: the tape learns. The reading is the same kind of evidence in both cases. What differs is which belief was cheaper to give up.

Bayes makes this exact. An observation updates every hypothesis it bears on, in proportion to how surprising it is under each. A surprising reading does not tell you which belief to revise. Your priors do.

In 2011 the OPERA experiment reported neutrinos that seemed to travel faster than light. Most physicists read the result as a measurement of the apparatus. That turned out to be right; a badly connected fibre-optic cable was among the causes. The ruler was a detector. The table was special relativity, which had far more evidence behind it.

The risk runs the other way too. If you always read anomalies as instrument error, you never revise the theory. Mercury’s orbit had an unexplained drift for over half a century, and the first instinct was to look for an unseen planet. In 1915 the drift became evidence for general relativity. At some point the table has to be allowed to measure the ruler and win.

Metrology turns the idea into a procedure. Calibration is the act of letting a trusted object measure the instrument, deliberately and in advance. Gauge blocks check micrometers. Reference weights check balances.

The chain has to end somewhere. Until 2019 the kilogram was defined by a single metal cylinder kept near Paris. Comparisons with its official copies showed small drifts over the decades, and there was no way to say which side had moved. The 2019 redefinition ended the chain in a constant of nature, Planck’s constant, instead of an object. The last ruler is now one that nothing can measure.

The same logic covers testimony. When someone tells you something wild, you learn about the world and about the person. Benchmarks and model outputs behave the same way. A model that reports an extraordinary result tells you first about the model.

My working rule is to ask, before reading any result: which of my beliefs is weakest here? That is the one the result will probably end up measuring.