Instrumented People, Places and Things: office socialization, 2012

Most of the study's readings have not been found. Of the eight-a-second records of 24 instruments, what survives is about 27 hours of one device's weight, kept inside the team's own tool, a little over a day of another device's sound moments, and one sound training set. The weight is about 1 part in 1,100 of what the study recorded. Those fragments are drawn from their readings. Everything else here is the team's plots, traced, and drawn gray.

Can a container of candy say something about an office? In the autumn of 2012, for a confectionery client, the Iota team set out 24 open containers of candy across a Chicago company's offices, from busy walkways to small kitchenettes. Each stood on an instrument that stayed put while the office moved around it: a kitchen scale's load cell, a temperature and air pressure sensor, and a microphone that kept seven band levels and never speech. The team counted the study from 24 September to 14 November.

The study's anchor is a place, and its model of activity is socialization, with taking candy as its trace: who stops, when, alone or together. In TJ's words:

“So the place is the office, the activity is socialization, I think.”

A proof of concept. Could instruments this cheap see taking at all, tell one kind from another, and say anything about the talk around it? So read this page for the kind of picture the approach gives: cases and patterns. Any number here describes these few devices, and none measures offices in general. The work's scope, as TJ has described it, “did not include generating findings or interpreting the data”.

A draft, November 2026.

Moments

1. 11:12 amThree takings in 15 seconds. Device 1315, Monday, in the team's tool. Five searchers ask whether each was a single piece, a handful, a dig or a refill. A solid dot is the answer it keeps.
2. 3:05 pmOne moment, close up. Where the team's own recording of the tool stops to explain itself. Columns compare the moment's duration, weight change and energy with each fingerprint.
3. 10:05 amSeven takings in two minutes. One person lingering, or five going by? The weight has no person in it.
4. 12:37 pmA refill it did not expect. The count falls by about 260 in two drops. Refills in training were far smaller, so the Refills row throws both out, and the tool calls the first a handful. A fingerprint knows only its training set.
5. All MondayA day at one container. As the tool opens it. The count rises as candy leaves, so the tool draws the line falling: 197 at 7:07 am, 451 by noon, 189 after the refill, 349 by evening, and almost no change overnight.
6. ShinglesThe team's other way to count. Blocks of time, with no idea of what happened in them. The block's length decides how many events there were.
7. Eight weeksThe work week shows. Every charted device, traced. The weekends are empty, batteries die, and some containers go missing.
8. A dayA rhythm by the minute. All the study's single pieces and handfuls folded into one 24 hours, with the team's names for what it saw. It asks of a spike near 11 at night: “Facilities?” The office's cleaners may be in this picture too.
9. 18 OctThe team's own footsteps. Between 3:30 and 4:30 a line of bumps crosses the floors: the team changing batteries. An instrument can't tell who it hears.
10. SoundWhat the microphone kept. Seven levels, eight times a second, in short windows through the workday. Never speech.
11. PlacesWhere they stood, and the talk around them. 24 places by kind, and the team's map of each device's taking against the voices near it.
12. NotesThe team reading its first plots. “first day novelty?”, “weird detail”. One plot carries two different notes.
13. InstrumentA kitchen scale, pulled apart. What it held, and what it wrote down.

How to read what you see

Sherlock

The team's tool, redrawn from its source and running here on device 1315's own readings. Five rows, one searcher each. Gray bands are stretches where the weight moved. A column's height is its similarity to that row's fingerprint: duration, weight change, energy. A solid dot is a stretch the tool counts; a hollow one is thrown out; a dark bar lost to a closer match on another row. Scroll to zoom, drag to pan. The t, w and e bars on the left set how much each measure matters: try them.

Shingles

The team's unsupervised track, written here from its description and run on the same points. The shapes below are stretches of the Handfuls searcher, smoothed, scaled and grouped by dynamic time warping, as the team did. Added here to show the method. In the team's words of its own clusters: “Although it looks convincing, we don't know if it worked.”

Gray

“The team's plot, traced.” The team's charts survive and their data doesn't, so their marks are read from the drawings. Read them for shape. The eight weeks are good to about half an hour, and the week to about a minute.

Places

Floors are lettered, and products and messages left out. The plan is a schematic: one floor's two wings, straightened, with ten places read by eye from the team's plan.

Clock

The instruments' own, in local time.

What survives, and what does not

From readings: device 1315, Monday 24 September at 7:07 am to Tuesday at 9:46 am, as the moments when its weight changed: two exports the team bundled with its tool as a demo, which agree where they overlap. Device 1414, from when it was switched on at 12:30 pm on Thursday 20 September until 1:36 pm on Friday, and on the afternoon of Friday 19 October, as moments its sound tool classed as voice-like, bump or other: 241,070 samples, all of them classes and only 42 of them levels. A 12-minute training set of 15 November, the day after the study ended, with the band levels of 1,066 of its 5,944 samples. Where the training set was recorded isn't known. Where device 1414 stood that first Thursday isn't recorded either: its firmware was still in a test mode.

Also survive: the team's tool in four versions and its source, the instrument's firmware, the sound tools, the team's plots of three devices, and its charts.

Not found: the eight-a-second records of the 24 devices, the database made from them, the team's analysis scripts for shingling and clustering, and any sound of the rooms, which was never kept. The drive's index was searched by device number, date and file name. It stops at July 2020.

What the model could and could not tell

Viable, as far as these fragments go. A kitchen scale could see a hand go into a container. The surviving Monday holds a day's shape, a refill at lunchtime and quiet overnight. Across the study the work week shows, a day has a rhythm, and some places were loud and some nearly silent.

What it couldn't. Who took candy, or whether two people stood together: the weight has no person in it. How much was taken: the team names “drift of the scale measurement (not being able to tare)”. The microphone heard about 2¼ hours of a workday, so most taking has no sound beside it. And nothing checked an event against what happened: the team also names “acquiring data to validate events, capturing video (privacy concerns), and incomplete reporting on device locations.”

“What we really wanted to know”, the team wrote, “was how conversation and taking events relate.” Its map on the Places view is its answer, and it is a reading of clouds, with no statistic behind it that has survived.

TJ writes that such a record “could be evidence relevant to a social hypothesis without itself being a measure of sociability. The model's value depended on whether the relationship … could be examined critically, not on whether a dispenser could count events accurately.” (Modeling Activity, p. 4)

And: “An observation is not an activity model.” (Modeling Activity, p. 3)

One of the studies in Instrumented People, Places and Things.

The study, the instruments and the tools were the Iota team's, for a confectionery client not named here, in a company's offices not named here. This page was made in 2026 at TJ's direction, with Claude Code.