What Counts as a Signal

A signal is not an unusual observation. It is an observation that survives a specified transformation, control, or comparison. Until you have named the transformation it must survive, you do not have a signal. You have a surprise, which is a different and much weaker thing.

The Yard treats signal as operational. Every candidate signal is required to declare its source, its transformation path, its noise model, its expected failure surface, and the condition under which the observer is obligated to stop calling it a signal. That last requirement is the one most often skipped and the one that matters most. A claim with no stopping condition is not a measurement; it is a commitment.

The problem sharpens when the observation process may itself alter what is observed. In disputed and under-measured corpora, the analyst, the protocol, and the choice of which trials to keep are all part of the instrument, whether or not anyone admits it. A result that appears only after trials are discarded is not a result about the world. It is a result about the discarding.

This is why the harder work is usually upstream of the data. Before asking whether an effect is real, the Yard asks what observation would remain after the easiest explanations are removed: instrument drift, leakage between conditions, post-hoc selection, the analyst's own expectation. Whatever survives that subtraction is the only thing worth arguing about. It is frequently smaller than the original claim. Smaller is not a defeat. Smaller can be instrumented.

The public aperture shows selected instruments and records. The deeper protocols stay controlled where disclosure would compromise the work, the participants, or the invention strategy. What does not stay controlled is the standard itself, which is stated plainly so that anyone can hold the Yard to it: name the transformation, or you do not have a signal.

The surprise is free. The signal has to be earned against a control.