What the line actually does, measured from the cameras already installed — with no reference to the standard work sheet.
Work content varies 2.6× across the twelve stations measured — from 2.1 to 5.5 person-minutes
per board, and each station's own load swings two- to eight-fold from board to board. Seven stations show
operators chasing boards downstream because the work did not fit; two habitually work ahead. The imbalance does
not travel down the line, which means it can be fixed one station at a time.
Work content per board
person-minutes of hands-on work · every board measured, not planned
B2C14
5.53
B2C17
4.25
B2C15
4.23
B2C10
4.17
B2C18
4.15
B2C8
4.12
B2C13
3.76
B2C6
3.61
B2C7
3.23
B2C5
3.20
B2C11
2.78
B2C12
2.10
02.04.06.08.010.05.97 — available now7.03 at takt
▭ middle 80% of boards (p10–p90)│ median board ● average — the single number reported before├─┤ lightest to heaviest board■ work does not fit — operators chase boards downstream■ spare capacity
Every station is shown as a distribution, not a number. At the line's
current speed — a board every 179 seconds — a two-person station has 5.97 person-minutes to
finish each board; at takt (211 s) it would have 7.03. Every average sits below both lines, and that
is the trap. B2C14 averages 5.53 but runs to 7.96 on its heaviest board — it is over the line on a
large minority of boards. B2C5 averages 3.20 and reaches 9.70. The work fits on average at every station and
fails to fit regularly at several. Board-by-board detail:
every station, every board.
Where the work does not fit
chase index against work content · twelve stations · dashed lines are the medians
The line tells you this itself. When an operator cannot finish inside the pitch, they follow
the board downstream rather than let it go incomplete. We counted how often that happens at each station —
no standard times, no plan, just what the operators did.
The two measurements are independent, and that is the finding. Chase
index is the share of a station's own operator-appearances that end with the operator leaving downstream — a
comparison between stations, not a count of boards (note 3). It is uncorrelated with work content
(r = −0.16), and equally uncorrelated with everything the distribution above shows (note 4). Stations
land in all four quadrants: B2C14 carries the most work on the line and copes, while
B2C7 carries below-average work and cannot. So a station in trouble is not identified by how
much work it holds — adding minutes to a coping station is safe, and taking them off a struggling one may not
help if volume was never its problem. Dashed lines are the medians; hover a point for its raw counts.
Two moves to make first
adjacent stations · every measurement agrees
Move work from B2C7 into B2C8
the strongest signal on the line
B2C7 has the highest chase index on the line at 17.3 — its operators followed the board out 67 times
against 3 the other way. B2C8, immediately downstream, has the lowest at 2.2 and habitually works ahead.
They are adjacent, so the transfer is a work-element move rather than a re-layout. This move rests on
behaviour, not on volume — per operator present the two stations are within 1% of each other — which is
why it holds whatever the crew numbers turn out to be. Size it carefully all the same: B2C8 is already
over the 5.97 line on about one board in ten.
B2C7 · 3.23 p-min
chase 17.3 — cannot finish
→
B2C8 · 4.12 p-min
chase 2.2 — works ahead today
Move work from B2C13 into B2C12
the widest actionable gap — pending crew numbers
B2C12 carries the least work on the line at 2.10 person-minutes per board and spends its spare time
working ahead; B2C13 next door carries 3.76 and chases boards at 12.1. That 1.66 p-min gap is the second
widest between adjacent stations — B2C13 ↔ B2C14 is wider at 1.77, but B2C14 already carries the
most work on the line, so it is not a candidate to receive any.
Confirm the manning before acting on this one. The comparison above
treats both stations as equally crewed. They are not: measured presence is 0.89 operators at B2C12
against 2.07 at B2C13. Per operator present that reverses the move — B2C13 carries 1.82
person-minutes per operator, B2C12 carries 2.36 — so on a headcount basis B2C12 is the busier of the two
and this transfer would widen the gap rather than close it. Five of the nine adjacent pairs we measured
flip direction the same way. One number from Lear — operators assigned per station — settles
it, and until it arrives this move is a hypothesis and the one below is not.
B2C13 · 3.76 p-min
chase 12.1
→
B2C12 · 2.10 p-min
chase 3.9 — 35% loaded
No cascading-down effect
an overload at one station does not push work onto the next
A reasonable worry with this kind of measurement is that one
overloaded station pushes work onto the next, which pushes it onto the next. If that were happening you
could not act on any single station without tracing the whole line first.
It is not happening here. Pressure builds over two
or three stations and is then absorbed by a station with spare capacity. B2C8 and B2C12 both sit
immediately after a run of chasing and act as relief valves.
We also checked whether an operator chasing a board out of one station
makes the next station chase too, sweeping every delay from 0 to 300 seconds in 15-second steps — fine
enough to catch a whiplash inside a single board cycle. It does not: the transit window sits 4% above
chance (note 5), and the effect is the same size in the reverse direction, which a real hand-off
could not be. Each station's number is its own, which is why the moves above can be made
independently and in either order.
The other ten stations. Twelve of twenty-two were measurable. Four cameras showed as
offline on Lear's own layout; six more were not in the export. The gaps at B2C9 and B2C16 also break the
line into segments, which limits how far any flow conclusion can travel.
More than one product mix. This is two hours of one evening. Work content per station
moves with the harness variant, and a balance conclusion drawn on one mix can invert on another.
One camera needs repositioning; a second has been cleared. B2C11 views a walkway
rather than its work area, and its numbers are flagged throughout. B2C12 has since been checked and
passes — its camera does view a real work area with operators engaged on the board. Its region does
cut through a populated strip on one side, but no more than seven of the other eleven cameras do, so its
2.10 figure is a sound measurement of what that camera sees.
An unexplained shift at two stations. Comparing like with like across the eleven
stations filmed in both hours, engaged operators fell 13.8% in the second hour — but the
drop is not spread across the line. B2C5 (−68%) and B2C8 (−38%) account for 89% of it,
four stations rose, and the remaining nine are flat to within 2%. The question for Lear is therefore not what
changed on the line at 7 PM but what happened at those two stations — most likely two
operators leaving. Note that B2C14 has no second-hour footage at all; including it makes the fall look
roughly twice as large as it is.
Notes on method
What was measured. Operators were detected by a pose model in a region drawn on each
station's work area. An operator counts as working when their wrists are low relative to their own shoulders
and hips — reaching into the board. Validated against 36 frames scored by hand across 9 stations: mean
absolute error 0.28 operators for engagement, 0.39 for presence, within one operator on 97% of frames.
Work content per board = engaged person-minutes per hour ÷ 20.1 boards per hour. Pitch
of 179 s was measured three independent ways — board QR tracking, and occupancy rhythm at both 1-second and
15-second sampling — agreeing within 4 seconds. Takt of 211 s was supplied by Lear.
Chase index is the percentage of a station's own operator-appearances that end at the
downstream edge. It is deliberately not expressed per board: the tracker fragments a single person into
several tracks when they are occluded, which inflates absolute counts by an unknown factor. Dividing by
appearances — inflated the same way — largely cancels this. Ranking by raw hourly rate instead gives a
materially different order (rank correlation 0.62), and we use the fragmentation-robust version.
Work content and chase index are independent — across the twelve stations
r = −0.16 (Spearman −0.21). They measure different things: how much work is done, versus how much
did not fit in the pitch. The chase index is equally uncorrelated with everything the distribution shows —
against the share of boards over the 5.97 line r = −0.29, against a station's spread r = +0.27. Knowing
how much work a station carries tells you almost nothing about whether its operators can finish it, which is
why both measurements are on this page.
Cascade test. Two passes. At 5-minute resolution, correlation between downstream-exit
counts across nine adjacent pairs was +0.14 in the same block and +0.02 one block later. Because station
transit is 120 s against a 300 s block, that pass could not rule out a faster whiplash inside a single block,
so it was repeated on timestamped tracks at 15-second resolution, sweeping lags 0–300 s.
The transit window (90–150 s) came in +4.0% over chance across all nine pairs (p = 0.34);
no lag near 120 s rose above noise. Two controls confirm it: the reversed direction shows the same
elevation as the forward one (+11% against +12% on the seven pairs excluding B2C11) — and a cascade is
directional, so a symmetric effect is not one — while non-adjacent stations show +2.7%. Both nulls preserve
the line's own 179 s rhythm, to which every station is phase-locked. This excludes a cascade
accounting for more than about 12% of downstream chasing events; it does not exclude a smaller one.
Direction is behaviour, not camera geometry. Correlation between where operators
typically stand within a region and which edge they exit by is −0.01. Stations with high chase indices also
drift downstream while working (+0.04 to +0.10 of region width); B2C12 drifts upstream.
Cameras do not overlap, so no operator is counted twice. Confirmed three ways: feature
matching between adjacent views, visual inspection, and board-ID coincidence — no board appeared in two
cameras at the same instant across 14,000 tracked appearances.