The boundary between the two duration populations sits just under 60 seconds, and 60 seconds is the poll interval, not a property of the radio [9]. The line in the data was drawn by the instrument, and the tape then reported that line as a fact about the link.
The reason this survived two weeks of quotation is arithmetic. Sixteen episodes at a mean of 55.3 seconds is about 885 seconds of real dark time inside a 7,200-second window [21]. The eight blips the poller never saw account for roughly 114 of those seconds, about 13% of the total [20]. Lose half the events, lose an eighth of the downtime: recorded dark time comes out at 9.58% against a real 12.28%, or 78% of the truth [12][13]. Anyone checking the poller against an availability figure would have passed it. Anyone checking it against a rate would have been out by 2.3x, and the median duration by 1.99x [5][11].
The over-statement is a separate defect with a separate cause. Each row's duration is down-checks times the tick, written honestly as a lower bound, ">=60s dark" [14]. Five of the seven rows land near the truth. Two claim roughly double, and those two are exactly the rows where the remediation ladder climbed, four down-checks and six against one everywhere else [15]. The ladder's rungs are reassociate and bounce, so the monitor tears the link down itself and the seconds it spends doing that land inside the episode it is recording [16]. For those rows the instrument is a participant in the event it is measuring.
Note also that the long mode had eight members and the tape has seven rows, so the poller missed one episode above its own tick as well [22]. And one piece of arithmetic the writeup does not do: if a 14.22-second blip arrives at uniformly random phase against a 60-second check, it covers a poll about 24% of the time, so eight of them should have produced roughly two rows rather than none, and a clean sweep of eight misses has about a one-in-nine likelihood [23]. That is not evidence of a mechanism, but it is a reason to stop assuming the blips and the tick are independent.
The general shape holds well beyond one wifi dongle. Undersampling a duration distribution leaves the integral roughly intact while destroying the distribution, because the mass you drop is by construction the low-mass end [17]. Sums over time keep looking healthy; counts, means, medians and percentiles are wrong by an amount the record contains no trace of [18]. That is the trap in the dev.to writeup worth carrying away: one poller feeding a dashboard that shows uptime while engineers reason from it about how often and how long is answering three questions and is only correct for the first [24].
