Work content,
board by board

The bar chart on the work-content page shows each station's average and says the problem is the spread. This is the spread. Same measurement, same twelve stations — cut into the 40 board cycles the two hours actually contain, instead of averaged flat.

Every station, every board

person-minutes of hands-on work · one cell = one board cycle (179 s)

18:00 UTC — first hour

19:00 UTC — second hour

less work more work cycle exceeds 5.97 p-min ▨ no data

What the eye should catch first. B2C5 runs hot through the first hour — including the single heaviest cycle measured anywhere on the line, 9.70 p-min — and then goes cold for the whole second hour. B2C8 does the same, less sharply. Those two stations account for 89% of the line's entire second-hour drop; the other nine are flat. That is a staffing event at two stations, not the line slowing down.

The spread, not the average

box = p10–p90 of cycles · line = full range · dot = station average · 5.97 = a crew of two at the 179 s pitch

This is the drop-in change. The existing chart plots the black dot alone. Adding the box and the range says the thing the page already asserts in prose: B2C14's average of 5.53 sits just under the 5.97 line, but its cycles run to 7.96 — it is over capacity on a large minority of boards. B2C5 swings 8.3× between its quietest and busiest cycle.

As it happens

each station's own trace · horizontal rule = 5.97 p-min · shaded = second hour

What this measurement can and cannot carry

stated plainly, so the numbers are used for what they support

The fluctuation is real, not sampling noise. Engagement is sampled every 15 s, so a cycle holds only 12 samples and some spread is inevitable. Shuffling each station's samples in time gives the noise floor: 56–87% of the observed cycle-to-cycle spread survives it, varying by station. The traces are signal.
The 5.97 line is confirmed. Lear has verified that every station is crewed with two operators, so the capacity line is 2 × 179 s ÷ 60 = 5.97 person-minutes per board — measured, not assumed. That also makes total work content the right basis for comparing stations.
What the cameras see varies, and it tracks the result. Because the crew is known to be two everywhere, measured presence is no longer a headcount — it is occupancy, the share of expected operator-time the camera actually saw. It ranges from 44% at B2C12 to over 100% at B2C14 and B2C17, where the region also admits people who are not that station's operators. Across the twelve stations, unseen operator-time and measured work content correlate at r = −0.85. Some of that is real — a station whose operators leave to chase boards is genuinely doing less work in front of its camera. But it means a station reading light cannot yet be separated from a station being watched poorly, and B2C12, the lightest on this chart, is also the least observed. The section below is what closes that gap.
Cycle boundaries are approximate. Cells are fixed 180 s windows from the clip start, against a true pitch of 179 s — they are not phase-locked to real board arrivals, and drift about one cycle-width across the two hours. Fine for "how much did this station swing", wrong for "which board was heavy".
Two stations are short. B2C14 has no second-hour clip (18 of 40 cycles) and B2C7's first-hour clip ends early (29 of 40). Missing cycles are hatched, never zero-filled.
B2C11's camera views a walkway rather than its work area, so its numbers are flagged throughout the study. It is shown here for completeness, not for use.

How to improve the measurement

four changes, in the order they would pay back

Point the cameras over the operators' shoulders, with the whole board in view. Today they look across the line, so an operator's own body hides the part of the board being worked on and the measurement has to infer hands-down posture rather than see the hands. Framing from behind, with the full board in frame, would let it see what the hands are on — and would remove the largest single source of missed work, which is an operator standing between the camera and the board. This is the change that would move the r = −0.85 above closer to zero, and it is worth doing first.
Map each station as its own shape, not as a rectangle. A station is a curved run of conveyor seen through a wide-angle lens; a rectangle is neither. Every region here is a box with its bottom extended to the frame edge, and it shows: on some cameras 10–33% of the engaged work in frame falls outside the region, while others admit the neighbouring station's operators. A polygon drawn on the work area — or better, a mapping from each camera onto the line's own floor plan — would let a region follow the station instead of approximating it, and would let the same operator be attributed correctly wherever they stand.
Close the gaps between stations. A board is visible inside one camera for 30–60 s against a 120 s station-to-station transit, so the fields of view do not meet. An operator working in the space between two cameras is counted by neither — which is exactly where a chasing operator goes. Two stations, B2C9 and B2C16, have no camera at all, breaking the line into three segments and limiting how far any flow conclusion can travel. Overlapping the views slightly would cost nothing and close both holes.
Run one resolution across the line. Four of the twelve cameras record at 640×480 and seven at 2592×1944. The board placards decode reliably on the high-resolution cameras and not at all on the low ones, so board identity and true per-board timing are available on part of the line only — which is why the cycles on this page are fixed 180-second windows rather than real board arrivals. Matching the resolution would make per-board measurement possible everywhere.

None of these need line time. Three are camera placement and one is configuration. The measurement itself does not change — the same signals get read from a view that supports them, which is what turns a study into a number you can run a line on.