Night Colonies · Look Under the Hood

Concept

A watch has been recording me while I sleep since 2021, and I have never once looked at what it wrote. I gave the whole export to a machine and asked it to keep a laboratory diary about me. It calls me Subject 01.

Every other experiment here is me working with a machine. This one is the machine working on me. It reads the nights, decides what is worth noticing, writes the note, grows the plate. I am the material.

That inversion cost me the thing I usually keep. I am not the author of this one and I am not quoted inside it. Not one sentence of mine is in it.

“It is trying to answer whether the body reports what the person cannot. It cannot see my dreams. It only measures the body that had them.”

From my own description of the experiment · it had never appeared on the site until this page

Why a dish

Because the container never changes. Every night gets the same glass circle, the same ground, the same light, so whatever is different between two nights has to come out in how the thing grew. A chart cannot do that. Its frame moves with its data, so the frame is doing half the talking.

So each reading got a colony shape instead of a bar. Deep sleep is a rounded body half filled, like a dish that has settled. REM is a scatter inside a dotted edge. Variability is concentric rings. Movement is a ruffled star. The shape says which signal it is, the size says what the number was.

The lab clothes are a register, not a claim. This began as a research question about health and I dropped that framing early. It is an art project with a subject, and the subject is enough.

One petri dish on a pale ground, six colonies inside it and nothing else: a smooth cream dome, a blue-grey fur, a ridged tan spiral, a small white ruffle, a green segmented disc, and a flat amber pool. — One night's plate: 16 August 2026, the longest of the fourteen at 8h 22m · six colonies, one for each reading, grown by an image model from that night's own numbers · Fig. 01

What it runs

Nobody wrote the diary. A script reads the six readings of a night, compares them against the fortnight, and builds the note out of a fixed vocabulary. Its own header says it: rules only, no sentence composed by a person.

The plates are not photographs. Each night's six numbers become a written description of a dish, and an image model grows the picture from those words. A plate is a picture of a description of a night.

Everything runs inside one document that never navigates. The room tone cannot survive a page load, so the page does not load.

A single night on the log. On the left a photograph-like petri dish holding six colonies, with a method line under it. On the right the six readings in a list, the machine's note, and a vertical bar of the night's sleep stages. — One night on the log: the plate, its six readings, and what the machine noticed · screenshot, 30 August 2026 · Fig. 02

The study

Fourteen nights, 15 to 29 August 2026, six readings a night: deep sleep, REM, variability, skin temperature, body response, movement. Eighty three colonies out of a possible eighty four.

Both gaps are printed. Variability did not report on the last night, so that plate grew five. And 24 August returned nothing at all, so the study ran fifteen days to reach fourteen nights. The gap is the reading here, not a missing number.

Six line drawings in a row, each in its own colour: a rounded half filled body, a dotted scatter, concentric rings, a blob holding a smaller solid shape, a cluster of circles, and a ruffled star. Named deep sleep, REM, variability, skin temp, body response, movement. — The six shapes, one for each reading · the piece's own drawings, shown together at size for the first time · Fig. 03

What it found

No night was typical. The shortest was 3h 13m, the longest 8h 22m, the middle night 6h 02m. That is the first thing the plates showed me that asking had not.

After my busiest day, 17,561 steps, REM fell to 9 minutes against a middle of 66.5, and deep sleep barely moved. A hard day cost me a particular kind of sleep, not an amount of it.

And not only that one pair. Of the twelve pairs of nights that ran back to back, ten moved opposite ways: the busier the day, the less REM the night after. It is the strongest thing here and it is still one fortnight of one person.

The lab report's second section: fourteen petri dishes in two rows of seven, each captioned with its date and how long I slept, the method line below them and a key of the six colony shapes at the foot. — All fourteen plates and the key, from the lab report · screenshot, 30 August 2026 · Fig. 04

What was sent back

Most of this build is rejections. The versions I threw away are more useful than the one that survived.

The ageing went first. A plate was meant to thicken across a month. It could not be built honestly: ask an image model twice and it returns a different dish, not an older one. There is no version of that which is not a new picture pretending to have grown. I dropped it rather than fake it.

Nine ways into a plate were built and I threw away eight. The one left moves nothing: the ground crosses and the plate is already where it belongs when the black arrives. What I sent back, again and again, was the dish flying across the screen.

A dark screen with the six colonies of one night laid out around a faint circle, each labelled with its signal name, and the night's readings listed on the left. — The same night broken apart, one colony to a reading · screenshot, 30 August 2026 · Fig. 05

“I still see the sad PowerPoint animation.”

Me, on the seventh version of the same move · 27 August 2026

Where it fails

One subject, fourteen nights, nothing held constant. Every sentence here describes this fortnight and nothing else. The pair of nights the report leans on is the closest thing to an experiment, and it is still two nights that fell next to each other. It shows a difference. It does not show a cause.

The plates are drawn, not photographed, and the piece did not say so until the day it went live. The line under every dish was written on 29 August and built on the 30th, hours before this page existed. It showed generated images, called each one the plate for its night, and never named the machine that made it.

And I am nowhere in it. That was my call and it took me longest to agree to. It is a cost, not a feature: a site called The Robot and Me has one experiment on it with no me in it at all.

A woman curled asleep on her side inside a petri dish, drawn in black halftone dots on white, the dish ringed with a heavy outline against an ultramarine ground. Dark colonies bloom around her hair and her legs. — The subject, in the dish · made with Gemini, 30 August 2026 · Fig. 06