The napkin dispenser study
asked what one everyday object
could say about the people using it.

Iota put small wireless sensor modules on dispensers in a small independent coffee shop, in the winter of 2010–11, and logged what they read.

The anchor is a thing: the sensing is on the object, and the object is the focus. A whiteboard of December 2010 gives the premise: “sensing simple things has value in aggregate”.

Napkin dispenser, slides. A draft, November 2026. An Iota project. Thanks to the coffee shop.

Anchoring an activity to a thing
is one way into a model of it.

“A framework for doing so was anchoring the activity from the perspective of the person, place, or thing.”

“Activity is not a 4th idea, it is what is common across all three. Models of activity.”

TJ, 2026, as on this site's front page.

The study's list
sets six questions apart.

Above a pencil line: how many taken per user, how often, when, how full, how long until a refill, attempts on an empty dispenser. These went into the proposal.

Below it: a low battery, a broken dispenser, type of paper, people nearby, staff refills, and use of the sink, the garbage can and the microwave.

A printed list of questions about towels and napkins, photographed at an angle, with a pencil bracket and line around the first six and a handwritten line at the foot about door sensors.

Fig. 1 · The list, photographed 28 November 2010. Its author is not recorded. The six open the proposal of December 2010.

A whiteboard climbs
from signals to models.

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.

A whiteboard diagram rising from a sensor and its signals, through a PC writing a raw text file and a cloud database, to query engines and visualization, with the words filtering, tagging, patterns and models stacked up the left side.

Modules radioed their readings to a PC
that logged them with the time.

A hand sketch: three dispensers with radio arrows to a hub on a box, the hub reaching a cloud through the air, and a laptop reading from the cloud.

A hand sketch: a flat rounded module with a clip beneath it, and plug-in sensors on leads: a ring, a small box, a long strip with a clip and a small board.

Eight inputs, about eight times a second: an accelerometer, a battery level and jacks for plug-in sensors.

Fig. 3 · Sketched for the proposal, December 2010. Readings: firmware demoPlatform_03G, logger PlatformDataloggerV22G. Which sensor went in which jack was not written down.

The photographs show two towel dispensers,
and not which module sat on which.

A black wall-mounted paper-towel dispenser above a hand sink, behind a coffee shop's counter, with a stainless table in front.

A tan wall-mounted paper-towel dispenser beside a mirror, above a washroom sink.

Fig. 4 · Behind the counter, and in the washroom. Photographs of 3 February 2011. None found shows a napkin dispenser in the shop, and no map or label of the modules has been found.

Seven sessions survive from March 2011,
and the logger stops on the 22nd.

Fig. 5 · What survives

Session	Clock	Readings	Pictures
Mon 14 March	15:29–16:01	52,286	none
Tue 15 March	14:24–15:05	73,309	none
Thu 17 March	14:17–21:52	524,872	none
Sat 19 March	11:52–16:15	342,888	the shop from 13:26, busy
Sun 20 March	13:04–13:20	12,618	the shop from 13:13
Mon 21 March	13:16–20:03	657,752	the shop from 13:44, quiet
Tue 22 March	07:27–14:45	573,012	the shop, a working morning

The webcam runs on to 26 March, with no readings; its pictures are withheld until reviewed.

An earlier run, 5 to 20 February,
does not read as dispensers.

Fig. 7 · Each minute's highest reading, live. On module 10, A5 rises near sunrise and falls near sunset, and A7 goes on and off about 80 minutes apart; the logger's code tags module 10 “fridgePowerConsumption”. Inferred: nothing records where the run took place.

A quiet afternoon sets the limits:
one person in the pictures, for 13 minutes.

Fig. 8 · Monday 21 March, 13:50 to 18:00

Module	Largest, to 15:55	Limit	Largest, from 15:55
101	27	30	25
102	27	30	28
103	36	39	37
104	30	33	28.5

Deviation of A0–A2 in counts. The later hours did not set the limits, and never reach them. Each of the 4,965 pictures was compared with those around it: one person, from 14:01 to 14:14, near the far end of the counter, and the readings then are no different.

A touch is a jolt or a lasting step
past those limits.

A jolt: A0–A2 together, each against its running median over about 5 s, past the module's limit.
A step: A4, A5 or A7, 10 s against the 10 s before, moved by 3 counts or more.
Counted only while the module is in place and above its low-battery level.
It means only that something moved the module or what it sat on.

“The selection of an instrument, the definition of an event, the alignment of observations in time, and the testing of an inference are all choices that shape the model.” McLeish, Modeling Activity, p. 4.

Touches run from none on a quiet afternoon
to dozens an hour on a busy Saturday.

Fig. 10 · Counted touches and hours logged, by module; in brackets, bursts of three readings or more

Session	101	102	103	104
Mon 21 March, quiet	0 in 5.0 h	0 in 4.5 h	0 in 5.2 h	0 in 5.0 h
Tue 22 March, a working morning	3 in 4.4 h (0)	not logged	11 in 4.6 h (1)	5 in 4.2 h (2)
Sat 19 March, busy	2 in 0.5 h (0)	37 in 1.7 h (1)	4 in 0.8 h (0)	57 in 1.3 h (4)
Thu 17 March, no pictures	65 in 4.1 h (14)	26 in 4.2 h (5)	32 in 4.5 h (7)	22 in 4.1 h (2)

Touches, not napkins. A charge lasted about seven hours, with its last hour left out, so no session covers a shop's day. On 17 March no module has a touch from 18:46 to 20:24. One afternoon's limits may not hold on every day.

Someone is at the counter
at 87 of 110 touches,
and at 68 of 126 random moments.

Fig. 11 · Checked against the pictures

Span	Touches	Random moments	p
22 March, modules 101, 103, 104	13 of 19	16 of 38	0.056
19 March, module 102	27 of 37	22 of 40	0.079
19 March, module 104	47 of 54	30 of 48	0.003

Shuffled, with no times or labels, and read by one reader, Claude, an AI model; p by permutation, one-sided. With steps of 2 counts rather than 3, module 104 reads 71 of 97 (p = 0.13). It says a person was near, and cannot say a hand reached a dispenser.

On the team's bench, half the movements
last three readings or more; in March fewer do.

At the March rate, under the same rule: 53 of 105 on the bench.

5 of 85 on 19 March, 3 of 9 on 22 March, and 28 of 118 on 17 March.

The team's logger window, titled PlatformDataloggerV03, drawing six colored traces with bursts at about a dozen moments, and one trace swinging down and creeping back at each.

Fig. 12 · The team's logger drawing a bench test, 23 December 2010. What was done on the bench is not recorded. Many single readings may be knocks on what a module sat on.

Nothing found says which module sat where,
though 102 and 104 shared a surface.

All four turned over between 14 and 15 March, and hold their tilt within a degree in every session from then but 20 March.
102 and 104 jolt within 1 s of each other 5 times on 17 March and 7 on 19 March. Moved in time, the same lists agree at most 1 and 2 times. No other pair does.
The picture does not place them: change around each module's movements stands out nowhere (smallest p 0.06).

Inferred from the readings. The proposal, firmware and logger name no places, and at 320 × 240 the pictures show no dispenser.

One plug-in input climbs on days with people,
and holds on the quiet afternoon.

Fig. 13 · A5, its change in counts over the counted time, by module

Session	101	102	103	104
Mon 21 March, quiet	−2	0	+2	−1
Tue 22 March, a working morning	−1	not logged	+17	+1
Sat 19 March, busy	−2	+15	+31	+15
Thu 17 March, no pictures	+94	+1	+32	+7

A stack getting lighter would do this. So would a sensor creeping after it was handled: on the team's bench one creeps about 30 counts in 90 s. Nothing here separates them, so the state of supplies is not established.

The model of activity stops at touches:
no reading tells a napkin from a bump.

Signals

survive

Raw text file

survives

Filtering

added in 2026: limits, counted time

State map

added in 2026: touches, by session

Tagging

no labels found

Patterns, models

not reached

The steps are the whiteboard's (Fig. 2). “An observation is not an activity model.” “Data does not arrive already labeled with the meaning of an activity.” McLeish, Modeling Activity, pp. 3, 4.

Of the six questions above the line,
the readings answer two in part.

How many taken, per user	No: one napkin, several and a bump look alike.
Frequency; volumes	Touches per hour, yes; volumes, no.
When, by time of day	For the hours counted, at most about five a session.
State of supplies	Not established: A5 climbs, for reasons not separated.
How long until a refill	Not seen.
Attempts on an empty dispenser	Not seen.

Questions: the study's list (Fig. 1). Answers: this study, November 2026.

Below the line, the readings speak to two
questions and leave the rest open.

Battery level is low	Yes, if A3 is the battery, on three modules of four.
Dispenser is broken	Not seen.
Type of paper; other paper installed	No: no record of the paper, one shop.
People nearby, in the room	Not from the readings. The pictures could say.
Industry, size, busiest time, staff refills	No: one establishment, no refill identified.
Sink, garbage can and microwave use	No: not instrumented, as far as survives.
Does the dispenser move around	Little sign of it: the tilt holds in five sessions of six.

Questions: the study's list (Fig. 1). A last line added by hand reads, as far as it can be read, “Door sensors (door to room)”.

The limits are those of the instruments,
the pictures and what was written down.

Eight readings a second: a quick pull can fall between two, so counts are a floor.
A charge lasted about seven hours, and its last hour is noisy and left out.
The pictures are 320 × 240, the dispensers do not show, and one AI reader coded them.
No note found says what was plugged in where.

Models “are necessarily partial, reflecting the limits of perception, instrumentation, available knowledge, and the assumptions embedded in their construction.” McLeish, Computational Design, p. 1.

The study stands on these files,
kept on TJ's drive.

A list of questions, 28 November 2010. Lightweight Platform Prototype: Napkin Dispenser Monitor, Revision 2, 20 December 2010.
Firmware demoPlatform_03G, 7 March 2011. Logger PlatformDataloggerV22G, March 2011.
The logger's files, December 2010 to 22 March 2011. 663,290 webcam frames, 19 February to 26 March 2011, not published.
Photographs of the shop, 3 February 2011. McLeish, Modeling Activity and Computational Design, 2026.

No report of the study from the time has been found. The sensing and the modules were Iota's.

The napkin dispenser study asked what one everyday object could say about the people using it.

Anchoring an activity to a thing is one way into a model of it.

The study's list sets six questions apart.

A whiteboard climbs from signals to models.

Modules radioed their readings to a PC that logged them with the time.

The photographs show two towel dispensers, and not which module sat on which.

Seven sessions survive from March 2011, and the logger stops on the 22nd.

Every session on one clock: a lane per module, a tick for each touch.

An earlier run, 5 to 20 February, does not read as dispensers.

A quiet afternoon sets the limits: one person in the pictures, for 13 minutes.

A touch is a jolt or a lasting step past those limits.

Module 103 through the morning of 22 March: touches, a jolt, a climb and handling.

Touches run from none on a quiet afternoon to dozens an hour on a busy Saturday.

Someone is at the counter at 87 of 110 touches, and at 68 of 126 random moments.

On the team's bench, half the movements last three readings or more; in March fewer do.

Nothing found says which module sat where, though 102 and 104 shared a surface.

One plug-in input climbs on days with people, and holds on the quiet afternoon.

The model of activity stops at touches: no reading tells a napkin from a bump.