Almost every number quoted about working memory is a number about how much. The number that decides what an interface can ask of somebody is how long, and it is much smaller than people expect.
Baddeley and Hitch gave the store its name in 1974. The measurement this page is built on is older than that. In 1958 John Brown, and in 1959 Lloyd and Margaret Peterson, handed people three consonants, made them count backwards by threes from a random number so they could not say the letters to themselves, then asked for the letters. After three seconds nearly everybody had them. After eighteen seconds nearly nobody did. Three consonants is not a load. It went anyway.
Whether it decays on a clock or gets pushed out by what came after it is a fight that is still running in the literature. For anything you are designing the two answers give the same instruction, so you do not have to pick. The gap between showing somebody something and asking them to use it is charged for by the second, and everything you put in that gap is charged for separately.
So the number of things below is fixed at five and never changes. Five letters, then a stream of single digits at one a second that you have to call odd or even, then you are asked which letters they were. Six rounds. Three of the gaps are three seconds long and three are twelve. Same five letters to hold, same digits to judge, same everything except the length.
Which length you get first is a coin flip taken when you press start, and then they alternate, so the practice you pick up and the boredom you accumulate land on both lengths three times each.
Five letters, then a gap you are made to spend on something else, then the letters are asked for. Six rounds. Three of the gaps are three seconds and three are twelve, and the five letters never change in number, so the only thing under test is how long the gap was. It needs scripting to run, and scripting is off.
What the digits are for
They are not there to be distracting. They are there to stop you saying the letters to yourself, which is the only defence you have and would otherwise make the gap free. A gap you are allowed to rehearse through is not a gap, and a page that measured one would be measuring how long you are willing to sit still.
They also double as the check on the run. The rig counts how many of the digits you actually answered. If you stopped judging them and used the time to hold the letters instead, the gap was never filled, and the comparison at the end says so rather than reporting a number it has no right to.
The picking is scored with the obvious correction. Five of the twelve letters were on screen and seven were not, so a wrong pick cancels five sevenths of a right one. That leaves picking at random worth nothing on average and takes the reward out of clicking wide.
Two honest weaknesses, both of which push the result the same way. Picking five letters out of twelve is recognition, and recognition is far gentler than writing them down cold, so whatever drop this finds is the smallest version of that drop. And by round four some letters have been on screen before, in an earlier round, which interferes on its own account — that effect is real, it is called proactive interference, and this rig cannot separate it out.
What it costs you
The gap is the design decision, not the remembering. A value shown on step two and needed on step five does not cost you a person's memory. It costs you however many seconds and however many decisions you put between the two, and you chose all of them.
An interruption is not priced in seconds. A permission prompt that appears mid-task takes four seconds of wall clock and takes the entire contents of the gap with it. The person comes back to your form having paid for the prompt twice.
Words in the gap cost more than anything else in it. A confirmation sentence, an autoplaying clip with speech in it, a chat bubble that opens itself — they land in the same channel the person was using to hold your value, and they land on top of it.
Nothing that must be carried should have to survive an app switch. That is the longest and most crowded gap in the whole product, it is filled with somebody else's interface, and you have no control over any of it.
I read a six-digit code off my phone last week, walked the four metres to the laptop, and arrived with five of them. The one I lost was a 3. I went back and it was still a 3.