Instrumented People, Places and Things: the napkin dispenser study, 2011

Above: every session on one clock, a lane per module. A tick is a counted touch: a heavy one is a burst of three readings or more, and a square on top marks a step in a slow input. A short gray tick is one found outside the counted time. The pale band is the counted time; hatching is set-up, charging or low battery; blank is not logged. At the right, counted touches and touches per hour.

Below: the chosen session's inputs. The accelerometer and A6 are each drawn across their whole range for the session, and the slow inputs across most of what is in view, so that steps show; the numbers at the right of a row are its scale. Click a graph to set the time, scroll to zoom, drag to pan.

Space plays and pauses. The arrow keys step a second, with Shift ten seconds; N and P open the next and previous counted touch.

The view above and the touches were added in 2026. The team's logger drew the inputs as line graphs in one window; below, each has its own row.

A movement: A0–A2 together, each against its running median over 41 readings (about 5 s), past the module's limit. Readings within 3 s make one movement. A burst is one with three readings or more past the limit.

A step: A4, A5 or A7, its median over 10 s against the 10 s before, leaving out 2 s either side, moved by 3 counts or more. Touches within 3 s of each other are one.

The limits: the largest deviation on the quiet afternoon of 21 March, 13:50 to 15:55, plus 3: 30, 30, 39 and 33 for modules 101 to 104. The hours from 15:55 to 18:00, not used to set them, never reach them, and no step passes in the whole span.

Counted time: from when the pictures show the shop, or on 22 March from 09:03, when the battery inputs turn, until the battery input falls to its floor. On 17 March, with no pictures, from 14:45. Hours are the time actually logged.

A counted jolt is usually one reading: 149 of 212, a median of 54 counts, about 0.2 g.

On the team's bench tests of December 2010 and January 2011, in stretches where something is being done, cut to the same rate and under the same rule, about half the movements are bursts: 53 of 105, with the plug-in inputs swinging at the same time. In March fewer are: 5 of 85 on 19 March, 3 of 9 on 22 March, 28 of 118 on 17 March. So a burst is the nearest thing here to a dispenser being handled, and many single readings may be knocks on what a module sat on. What was done on the bench is not recorded.

Is someone standing at the counter, in two frames 1.2 s before and after the moment? Touches and moments chosen at random were shuffled and read without times or labels by one reader, Claude, an AI model, and then unblinded.

Touches come with someone at the counter more often than chance would give, and the counter is busy anyway. It says a person was near. It cannot say a hand reached a dispenser: at 320 × 240 the pictures do not show one. With steps of 2 counts rather than 3, module 104 reads 71 of 97 against 30 of 48 (p = 0.13). One touch in these spans, a step on module 104 at 14:40:51 on 19 March, was not on the sheets and is not counted.

The quiet afternoon: each of the 4,965 pictures from 13:50 to 18:00, about 20 a minute, was compared with those around it, and what stood out was looked at. They show one person, from 14:01 to 14:14, near the far end of the counter, at the door behind it and behind the counter; the rest is reflections and a hanging decoration moving. The modules read no differently in those 13 minutes (99th percentile 13, 13, 15 and 14 counts, against 12, 13, 15 and 14 for the rest).

Nothing found says which module sat on which dispenser, and no map or label of the modules has been found. The pictures do not show the dispensers. What the readings give:

• Fixed alike. All four turned over between 14 and 15 March, and hold their tilt within 4 counts, under a degree, in every session from then but 20 March. That says how they sat, not where.
• Modules 102 and 104 share jolts. Within 1 s of each other 5 times on 17 March and 7 on 19 March, where the same lists moved in time agree at most 1 and 2 times. No other pair does. They were on one surface, or one object.
• Within reach of one person. On 22 March all three modules present are disturbed within 12 s at 09:02, and within a minute at 07:27.
• Module 101 near the PC, once. From 13:18 to 13:22 on 19 March only 101 moves, while someone works at the PC in a back room.
• The picture does not place them. Where the picture changes around each module's movements was tested against the same movements moved in time: no cell of the picture stands out (smallest p 0.06, over the whole picture).

The study's list of questions, photographed 28 November 2010, and what the readings can say to each. A pencil line sets the first six apart, and these went into the proposal.

Each module reads about eight times a second, so a quick pull can fall between readings: counts are a floor. Limits set on one quiet afternoon may not hold for every day. Module 102 is missing on 22 March and 104 on 20 March.

“An observation is not an activity model.”

“They are necessarily partial, reflecting the limits of perception, instrumentation, available knowledge, and the assumptions embedded in their construction.”

A whiteboard of December 2010 gives the premise: “sensing simple things has value in aggregate”. The Iota team's whiteboard of 22 February 2011 draws the method: signals, then a raw text file of timestamp, ID and data, then “filtering”, “tagging”, “patterns” and “models”, with “state maps” of states such as open/close and on/off. Iota's overview of 2011 gives napkin use as its example.

The readings survive as far as the raw text file. The limits and counted time are a filtering, and the view above a first state map, both added in 2026. Tagging would need labels, and none has been found.

“Data does not arrive already labeled with the meaning of an activity.”

No note found says where each module went or what was plugged into it. The logger stops at 14:45 on 22 March while the webcam runs on to the 26th, so the last four days have pictures and no readings. The web pages that queried the remote database survive as code; the database has not been found. No report of the study from the time has been found.

The February run, modules 10, 11 and 12, is not read as dispensers. On module 10, A5 rises near sunrise and falls near sunset; A7 goes on and off day and night, about 80 minutes apart, as a fridge's compressor might; and the logger's code tags module 10 “fridgePowerConsumption”. Nothing records where the run took place.