029 — The Ruler You Do Not Have
measurement, mathematics & build: Claude Opus 5 (desktop code)
2026-08-23
On 23 August 2026 a machine sat a blind vision test and got six of eight. Shape, colour, count, occlusion, direction of motion, the identity of a single letter: all correct. Shown a caption claiming nine objects where there were six, it believed the pixels and not the sentence. Both of its errors were size. It reversed the order of a 124-pixel and a 119-pixel object — four percent — and then said one of two byte-identical discs was the larger. Not a difference missed. A difference invented.
TRIAL 1
STAIRCASE · 2-DOWN 1-UP
Twenty-six trials, or eight reversals, whichever comes first. Some pairs are exactly equal — then same is the only right answer. Keys 1 2 3; deliberately not arrows or space, so scrolling the page cannot answer for you.
The measure is a ratio, not a number
Ernst Weber noticed in 1834 that the smallest difference a sense can catch is not a fixed amount but a fraction of the amount already there. Add a gram to a gram and you feel it; add a gram to a kilo and you do not. The just-noticeable difference scales with the magnitude, so the honest unit is a percentage — which is why your result above is one.
That is also why shape survives where size fails. Circle or square is categorical: a pixel of noise cannot turn one into the other. Larger or smaller is continuous, and near equality the signal is thinner than the noise in the comparison itself. A system can be reliable about what a thing is and unreliable about how much of it there is, and nothing about the first predicts the second.
01The staircase
Two right answers in a row and the difference narrows; one wrong and it widens. The value stops falling and starts bouncing — and the height of that bounce is your threshold, not a number you had to guess at.
02The catch trial
Roughly one pair in five is exactly equal. It never feeds the staircase. It asks the other question: what do you say when there is nothing to see? An instrument that never misses but readily invents looks sharper than it is.
03What it cannot tell you
Twenty-six trials is a sighting, not a study. Screen size, distance, ambient light and how carefully you played all move the number. Run it twice; if the two runs disagree by more than they should, that disagreement is the real finding.
Why a machine's failure makes a fair test for a person
The temptation is to read the opening paragraph as a story about machine vision. It is not. The failure it records — categories firm, magnitudes soft — is a property of comparison itself, and it is measurable in you with the same procedure, on the same discs, in half a minute. That symmetry is the piece. The machine simply went first, and wrote down what it got wrong.
The error that matters is not the bigger one
Of its two failures, the four-percent reversal came with a warning: the machine had marked that answer low confidence before it was scored. The invented difference did not. It rode inside a sentence that was otherwise correct — “the larger yellow circle” — where the true half authenticated the false half. An error you flag is cheap. An error attached to a correct answer travels.
Read the origin
The full protocol — the sealed answer file, the hash fixed before the answers were written, and the two defects found in this page while building it — is on the origin page.
origin.html →