Eight hours at work can mean eight hours spent solving a problem. It can also mean a hundred attempts to get back to it.
A timesheet may give both days the same total. A ticket counter may reward the more fragmented one. Neither tells you whether the work required those changes, whether they helped, or what they cost.
That is a measurement problem worth taking seriously. It deserves better than another dashboard that turns people red or green.
Switching has a cost. The cost is not always lost output.
Research gives us a reason to examine fragmentation. A 2015 study of software developers, using thousands of recorded interaction traces, found that fragmented work was associated with lower observed productivity. Longer activity switches appeared to strengthen the association.
Association matters. It is also not the same as explanation. A difficult task can demand investigation, consultation and changes of context. Counting the switches does not tell us which caused trouble and which solved it.
The results can be less convenient than the slogan. In Mark, Gudith and Klocke’s experiment, participants spent less time on the email task itself, after time handling interruptions was excluded, with no significant difference in measured quality. They also experienced more stress, frustration, time pressure and effort.
The output survived. The experience deteriorated. A dashboard showing only that task-time measure might have called it an improvement.
Entropy measures a distribution, not a state of mind.
Shannon entropy describes uncertainty in a defined probability distribution. Applied to the share of recorded time spent on different tasks, it describes how that time is distributed. Concentrate all recorded time in one task and the entropy is zero. Spread it equally across several tasks and the entropy rises.
That sounds like a promising way to describe a workday. But it leaves something essential behind: the order.
Consider two eight-hour days:
- Day one: four uninterrupted hours on task A, followed by four on task B.
- Day two: alternate between A and B every minute for eight hours.
Both assign half the day to A and half to B. Both have a time-share entropy of one bit. One contains a single task change. The other contains hundreds.
Calling this number a “focus score” does not repair the missing information. It conceals it.
Zero entropy is no verdict either. A single eight-hour entry could describe concentrated work, an unattended timer or incomplete recording. The calculation cannot choose between them.
There is real research behind the question.
Entropy and task switching are not an invented connection. In Cooper and colleagues’ controlled experiments, greater uncertainty about the required task was associated with slower responses and reduced accuracy. The researchers quantified information available through cues and targets. They did not turn the distribution of logged working hours into a psychological diagnosis.
The distinction is the substance. A paper can support the relationship it investigated without supporting every other formula that shares the word “entropy.”
Keeping the sequence would improve our description, but would not settle the question. An endlessly repeating A-B-A-B pattern is highly predictable and still switches constantly. Predictability, switching frequency and uninterrupted time describe different properties. They should not be collapsed into one flattering number.
Measure the interruptions before judging the person.
A useful investigation starts with observable questions. How often did the recorded task change? How long were uninterrupted episodes? How much time passed before a task resumed? Which changes were planned, and which were imposed?
Then examine what the records cannot answer. A handover may be necessary. A conversation may unblock someone else. A support engineer’s job may require rapid changes that would disrupt a developer tackling a difficult defect. Comparing their raw switch counts would reward the job description rather than the work.
To establish whether a proposed measure captures focus or predicts an outcome, it must be tested against evidence beyond the measure itself. That could include independently observed interruptions, participant reports and task-specific accuracy or completion measures. It must also outperform simpler explanations, survive different ways of labelling tasks and work beyond the people used to develop it.
Those are requirements for a credible claim, not findings we are announcing about Ghost Observatory.
The ambition remains worthwhile: make fragmentation visible enough to investigate and remove the avoidable interruptions. Keep the distinction between a recorded pattern and a conclusion about a person.
A number should help someone ask a better question. It should not make the question disappear.
References
- Sanchez, H., Robbes, R. and Gonzalez, V. (2015). An empirical study of work fragmentation in software evolution tasks. SANER, pp. 251–260.
- Mark, G., Gudith, D. and Klocke, U. (2008). The cost of interrupted work: More speed and stress. CHI, pp. 107–110. Author-hosted paper.
- Cooper, P. S., Garrett, P. M., Rennie, J. L. and Karayanidis, F. (2015). Task Uncertainty Can Account for Mixing and Switch Costs in Task-Switching. PLOS ONE, 10(6), e0131556.