The instruments were already in your pocket.
A microphone and a compass. The Observatory is what happens when a register takes them seriously.
A phone is a poor instrument for a watch until it is asked the right question in the right way. The bench and the compass in the app ask two: how is this movement running, and has it met a magnet. Neither needs anything you do not already own, and neither asks a model for its reading — the reading is made on the phone and filed against the watch, in your own register. A model sees it only when you put a question to the Registrar.
It listens for the escapement.
Rest a mechanical watch on the microphone in a quiet room and the bench listens for the tick of the escapement — a click a few thousandths of a second long, in the band a phone hears well. It finds the beat rate among the ones watchmakers use, from 18,000 to 36,000 beats an hour, holds the rate in seconds a day, and reads the beat error in milliseconds: how unevenly the balance swings one way and the other. A quartz watch is refused; there is no escapement to hear.
Ten seconds give a first reading. The bench goes on listening, and the longer it does the steadier the figure; every reading carries its own spread, so a settled rate and a nervous one are never written the same way. Amplitude is not guessed at. It needs a cleanly separated impulse and a known lift angle, which a phone cannot promise, and the bench says so rather than invent a number.
It feels the steel.
Magnetism is the commonest reason a mechanical movement gains overnight, and nothing looks wrong. The compass check uses the magnetometer the phone already carries, and begins by knowing where it is. Apple moves it from one iPhone to the next — to the foot, beside the charging port, on most 16s and 17s; to the top-left corner on the 14s and most 15s — so the house carries a map of the phones it knows, and on one it does not, the first check finds the spot itself, the gyroscope confirming the phone lies still enough to trust.
Then the watch away for two seconds, for a baseline. Then slowly to the compass, dial up, and again dial down: from just beyond the phone’s edge, level with the glass, where the compass sits at one, and never over the speaker beside it. What swings the other way when the watch is turned over is a magnet; what stays is the steel of the movement, which bends the Earth’s field on its own and is named as steel. Never the back of the phone, whose own magnets would drown the reading.
Under six, nothing the phone can feel. Past twenty, magnetised: a watchmaker’s demagnetiser takes seconds, and it is the likeliest reason for a rate that jumped.
It looks, and is not believed.
A photograph is a poor way to identify a watch. Two references from one maker can differ by a single digit, one step of a bezel, or the metal itself — and to a lens, steel, titanium, white gold and platinum are all simply silver. The looking is done by a frontier vision model, asked to read the photograph the way a collector would: the printed line beneath the maker’s name, the shape of the hands, the reference engraved between the lugs. It is remarkably good at this. It is also, like anyone, occasionally confident about something that is not there — which is why it is never given the last word.
The model proposes. The register decides.
Judge the frame
Blur, darkness and glare are measured before anything is sent. A poor photograph is refused on the spot.
See, then conclude
What the camera can actually see is recorded separately from what the model concludes, so the evidence can be used against the conclusion.
Check the shelf
Every candidate is checked against the register’s own records, and any the photograph contradicts is struck out, with the reason in plain English.
Tell the twins apart
Where two references are near-identical, the register works out what separates that pair and puts that single question back to the model.
Or say so
When the evidence supports no single answer, it declines rather than guesses, and shows you the shortlist it could not choose between.
The instrument measures. The register keeps the measurement with its doubt attached. And every figure is one watch, on one day, on one phone — on the figures we publish.