Instrument methodology

Every number this station shows you comes from a sensor in your own device. Not one of them is invented, smoothed toward something more exciting, or seeded with a random figure to keep the display moving.

This is a build rule before it is a principle. There is no call to a random number generator anywhere in the path between a sensor and a reading. Where an instrument has nothing to report, it reports nothing, and the panel sits there looking unhelpful. We think that is the correct behaviour for an instrument, and it is certainly the more unnerving one.

What is invented is the frame: the Survey, the station, the case files, the numbering on the panels. That is fiction, we enjoy it, and this page is where it stops.

Less than people assume, and in stranger shapes. A modern browser will give an instrument a compass heading, the accelerometer and gyroscope, the camera, the microphone, and your position. It will not give it a raw magnetometer, a thermometer, a barometer you can rely on, or anything at all resembling a radiation sensor — whatever the app stores suggest.

It also asks your permission for most of it, and on some devices it will only ask if you touched something first. That is why instruments here have a switch you press rather than starting on their own. It is not showmanship; the platform genuinely requires it.

Desktop

A laptop has a camera and a microphone and no motion sensors whatsoever. Instruments that need movement will tell you so and offer a way to use them anyway. Nothing on this station shows you a dead screen and leaves you to work out why.

The most common question about this station is whether the detector is a real EMF meter. It is not, and the reason is worth knowing.

A real EMF meter reads a magnetometer directly and reports field strength in microtesla. Your phone has that hardware. Your browser is not allowed to touch it — the raw sensor was closed off to web pages years ago on both major platforms, for reasons of fingerprinting rather than anything to do with us. Any website claiming to show you microtesla is showing you a number it made up.

What a browser can read is the compass heading: the direction the device believes is north. A compass heading is a magnetometer reading that has already been interpreted, and crucially it carries the interference with it. When the field around the phone is disturbed, the heading stops being able to settle.

What the needle is made of

Instrument 01 samples heading many times a second and computes its circular dispersion — how much the readings scatter around their own mean. Not the direction: the disagreement.

Circular is the important word, and it is worth a figure rather than a sentence. Headings wrap around: 359° and 1° are two degrees apart, not three hundred and fifty-eight. Ordinary statistics get this catastrophically wrong, and the way they get it wrong is very specific — they report an enormous reading at exactly one bearing, which is the one the operator is facing when the readings straddle north.

That is not a hypothetical. It is the single most common defect in software of this kind, it makes an instrument appear to detect something whenever its user turns to face a particular direction, and it is invisible to anybody testing indoors facing a wall. Dispersion here is computed from unit vectors instead — each heading becomes a point on a circle, the points are averaged as vectors, and how tightly they cluster is the reading. A circle has no seam, so north is not special to any of it. The arithmetic is in the field note on compass work, with the formulae written out.

The result is a genuine physical measurement. Put your phone flat on a wooden table and it will read low. Hold it against a speaker, a laptop hinge, a fridge door, a wireless charger, or a wall with wiring behind it, and it will climb, because those things really do bend the field the compass is reading. You can verify the whole instrument in about four seconds, which is more than can be said for most of this genre.

What it does not mean

A high reading means the magnetic environment around your phone is disturbed. That is all it means. It is not evidence of a presence, and the Survey has never said otherwise in fifty-two years of trying not to.

Instrument 04 does three things, and its panel keeps them apart because they are not the same kind of thing.

It measures the frame. The picture is divided into 192 tiles and each is scored on how dark it is, how much it changes between frames relative to its own brightness, and how little structure it contains. Those three properties are why pose models fail into false positives, so the score is an honest answer to a real question: where in this picture can the array not tell what it is looking at? That figure is the resolution floor, and it is a measurement of your camera.

It tracks people, with a pose-estimation model running on your device, and marks anything shaped enough like a person to satisfy it. The model is looking for people. In an empty room it will still find candidates, because a chair back and a doorway and a coat on a hook are, to a pose model, extremely promising. That is the method, not a malfunction, and it is exactly how the original hardware behaved. The panel counts these as tracked returns.

And it draws figures. Where the frame cannot be resolved, the instrument will eventually put a figure there. Those are reconstructions: they are drawn by the software and they are not detections of anything. The panel counts them as a separate number, the register names them as reconstructions when they appear and again when they go, and every plate you capture states how many of each are in it. Nothing on this site adds the two counts together.

The one rule worth knowing is that a reconstruction can only come out of a region the array has measured as unresolvable, and the rate depends on that measurement. In a well-lit room the rate is zero and no figure will ever appear, however long you hold it there.

Instrument 12 — the thermal camera

Instrument 12 has no thermal sensor to read, because your phone does not have one and no software can give it one. Thermal imaging happens at around eight to fourteen micrometres and needs a microbolometer; a phone camera is silicon, which stops responding at about 1.1 micrometres, behind an infrared filter that removes most of the rest.

So it maps visible brightness onto the false-colour ramps thermal imagers use. A bright surface comes out looking hot because that is what the palette does, not because anything was measured about its temperature. The instrument never prints a temperature in any unit anywhere— not on the panel, not on the scale, not on a plate, not in your record. Every figure it produces is relative luminance on a scale of nought to a hundred, and the scale bar on the picture is labelled REL — not temperature in the place a real imager prints its degrees.

What it genuinely measures is the span — the window of brightness actually present in the frame, taken from its histogram, which is the level-and-span control on every real imager — and, if you hold a reference frame, the change since that moment. That second one is not an imitation of anything: differential work is what thermography is actually used for, and the visible-light version of it is a real measurement a phone can really make. Hold a reference and everything that has not moved goes black.

Both run entirely on your device. No frame is uploaded, none is stored, and no server is involved at any point.

Both of these listen to your room, and neither of them records it unless you press record. The microphone goes into an analyser and what comes back out is a handful of summary numbers — level, the room’s own noise floor, spectral shape. Nothing is uploaded from either instrument and there is no endpoint that could receive it.

Instrument 07 — the spirit box

It sweeps sixty-four synthesised channels faster than a word can complete. What you hear assembling itself into speech is being assembled by you — auditory pareidolia, extremely well documented, and stronger the more you want it.

There is no radio in it and there are no recordings in it. The static, the stations and every word are generated in your browser out of oscillators and filters. The words come from a model of a vocal tract: a buzzing source through four resonant filters driven to the formants of each speech sound in turn, which is how a voice physically works. We did not use your phone’s speech engine because its output cannot be routed into the audio graph on any browser, so it could not be filtered or swept and would arrive over the top of the instrument as a clean assistant voice.

What the word is, is synthesised. When it happens, is a measurement. The instrument tracks your room’s noise floor and only speaks in three defined conditions: answering into the quiet after a measured transient, slowly in a settled room, or unprompted when there is no microphone at all — and it says which every time. Above about eleven decibels over your room’s floor the rate is exactly zero: shout at it continuously and it will not produce a single word. That gate is the part you can check in fifteen seconds, and it is why the responsiveness is worth believing.

There is no random number generator in it. The word is chosen by a digest of the room measurement at the instant it fired, so the same room in the same state tends to produce the same word twice, and the transcript prints the seed. The band is seeded the same way, from your room’s measured floor, so a different room is a different band.

Instrument 09 — the EVP recorder

It records the room uncompressed and lets you go through the result slowly with the gain up. The capture never touches a codec, because a lossy codec is a perceptual model that deletes the quiet detail it decides you cannot hear — and that detail is the entire subject of the instrument.

The review chain does one thing: it makes the noise floor audible. Heavy compression with a very low threshold, a great deal of make-up gain, and a band-pass across the speech range. What rises is whatever was below the threshold, which is the room, and a room brought up thirty decibels is full of structure.

It marks regions, and a mark is defined precisely: band energy holding above the recording’s own tracked floor for the length of a syllable, with the level moving while it does. The third condition exists to reject the fridge. A mark is not a detection of speech — the instrument has no idea what speech is, and that sentence is printed next to the count everywhere the count appears.

Recordings are held in the page and are not written to your device’s storage at all. Close the tab and the recording is gone; the download button exists because of that. What goes to your case file is numbers and a small envelope, never audio.

Neither instrument starts making noise on its own. Nothing on this station does.

The Zener series and the intention protocol are the two things here that are not instruments at all. They are the actual experimental designs used at Duke in the 1930s and at Princeton between 1979 and 2007, run honestly, scored against chance, with the deck committed before you call.

They are the most defensible content on this site and, in the Survey’s opinion, the most interesting, because the honest result is genuinely stranger to sit with than a rigged one. Everybody gets a hot run eventually. Watching yourself decide that this one means something is the experiment.

Nowhere. Camera frames, microphone audio, motion, heading and location are read in your browser, used, and discarded. None of it is uploaded, none of it is sent to an analytics service, and there is no server-side component to any instrument on this station.

Your operator designation and your logged readings live in your browser’s local storage. Clear your browsing data and they are gone permanently — we cannot restore them, because we never had them. The full statement is in the privacy note.

These are the ones we are actually asked. Anything we would have to hedge to answer is not on the list, and the honest answer to several of them is less flattering than the marketing answer would have been.

It means the measurements are real and the conclusions are yours. Nothing on this site detects ghosts, spirits, presences, energies, or anything else outside the ordinary physics of a room with a phone in it.

It does not mean the numbers are fake. Those are two different claims, and most of this genre relies on you not separating them. We would rather you did.